You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
7626 lines
273 KiB
7626 lines
273 KiB
# Copyright 2020 Google LLC
|
|
#
|
|
# This source code is licensed under the BSD-style license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
#
|
|
# Description:
|
|
# XNNPACK - optimized floating-point neural network operators library
|
|
|
|
load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
|
|
|
|
licenses(["notice"])
|
|
|
|
exports_files(["LICENSE"])
|
|
|
|
OPERATOR_BENCHMARK_DEPS = [
|
|
":XNNPACK",
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
]
|
|
|
|
MICROKERNEL_BENCHMARK_DEPS = [
|
|
":ukernels",
|
|
":bench_utils",
|
|
":enable_assembly",
|
|
"@cpuinfo",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
]
|
|
|
|
ACCURACY_EVAL_DEPS = [
|
|
":XNNPACK",
|
|
":ukernels",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
]
|
|
|
|
MICROKERNEL_TEST_DEPS = [
|
|
":ukernels_test_mode",
|
|
":enable_assembly",
|
|
"@cpuinfo",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
]
|
|
|
|
OPERATOR_TEST_DEPS = [
|
|
":XNNPACK_test_mode",
|
|
"@pthreadpool",
|
|
"@FP16",
|
|
]
|
|
|
|
OPERATOR_SRCS = [
|
|
"src/operators/argmax-pooling-nhwc.c",
|
|
"src/operators/average-pooling-nhwc.c",
|
|
"src/operators/binary-elementwise-nd.c",
|
|
"src/operators/channel-shuffle-nc.c",
|
|
"src/operators/constant-pad-nd.c",
|
|
"src/operators/convolution-nchw.c",
|
|
"src/operators/convolution-nhwc.c",
|
|
"src/operators/deconvolution-nhwc.c",
|
|
"src/operators/depth-to-space-nchw2nhwc.c",
|
|
"src/operators/depth-to-space-nhwc.c",
|
|
"src/operators/fully-connected-nc.c",
|
|
"src/operators/global-average-pooling-ncw.c",
|
|
"src/operators/global-average-pooling-nwc.c",
|
|
"src/operators/leaky-relu-nc.c",
|
|
"src/operators/max-pooling-nhwc.c",
|
|
"src/operators/prelu-nc.c",
|
|
"src/operators/resize-bilinear-nchw.c",
|
|
"src/operators/resize-bilinear-nhwc.c",
|
|
"src/operators/sigmoid-nc.c",
|
|
"src/operators/softmax-nc.c",
|
|
"src/operators/unary-elementwise-nc.c",
|
|
"src/operators/unpooling-nhwc.c",
|
|
]
|
|
|
|
SUBGRAPH_SRCS = [
|
|
"src/subgraph/abs.c",
|
|
"src/subgraph/add2.c",
|
|
"src/subgraph/argmax-pooling-2d.c",
|
|
"src/subgraph/average-pooling-2d.c",
|
|
"src/subgraph/bankers-rounding.c",
|
|
"src/subgraph/ceiling.c",
|
|
"src/subgraph/clamp.c",
|
|
"src/subgraph/convolution-2d.c",
|
|
"src/subgraph/deconvolution-2d.c",
|
|
"src/subgraph/depth-to-space.c",
|
|
"src/subgraph/depthwise-convolution-2d.c",
|
|
"src/subgraph/divide.c",
|
|
"src/subgraph/elu.c",
|
|
"src/subgraph/floor.c",
|
|
"src/subgraph/fully-connected.c",
|
|
"src/subgraph/global-average-pooling-2d.c",
|
|
"src/subgraph/hardswish.c",
|
|
"src/subgraph/leaky-relu.c",
|
|
"src/subgraph/max-pooling-2d.c",
|
|
"src/subgraph/maximum2.c",
|
|
"src/subgraph/minimum2.c",
|
|
"src/subgraph/multiply2.c",
|
|
"src/subgraph/negate.c",
|
|
"src/subgraph/prelu.c",
|
|
"src/subgraph/sigmoid.c",
|
|
"src/subgraph/softmax.c",
|
|
"src/subgraph/square-root.c",
|
|
"src/subgraph/square.c",
|
|
"src/subgraph/squared-difference.c",
|
|
"src/subgraph/static-constant-pad.c",
|
|
"src/subgraph/static-reshape.c",
|
|
"src/subgraph/static-resize-bilinear-2d.c",
|
|
"src/subgraph/subtract.c",
|
|
"src/subgraph/unpooling-2d.c",
|
|
]
|
|
|
|
TABLE_SRCS = [
|
|
"src/tables/exp2-k-over-64.c",
|
|
"src/tables/exp2-k-over-2048.c",
|
|
"src/tables/exp2minus-k-over-4.c",
|
|
"src/tables/exp2minus-k-over-8.c",
|
|
"src/tables/exp2minus-k-over-16.c",
|
|
"src/tables/exp2minus-k-over-64.c",
|
|
"src/tables/exp2minus-k-over-2048.c",
|
|
]
|
|
|
|
SCALAR_UKERNELS = [
|
|
"src/f32-argmaxpool/4x-scalar-c1.c",
|
|
"src/f32-argmaxpool/9p8x-scalar-c1.c",
|
|
"src/f32-argmaxpool/9x-scalar-c1.c",
|
|
"src/f32-avgpool/9p8x-minmax-scalar-c1.c",
|
|
"src/f32-avgpool/9x-minmax-scalar-c1.c",
|
|
"src/f32-clamp/gen/scalar-x1.c",
|
|
"src/f32-clamp/gen/scalar-x2.c",
|
|
"src/f32-clamp/gen/scalar-x4.c",
|
|
"src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
|
|
"src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
|
|
"src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x4-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up1x4-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x4-scalar.c",
|
|
"src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x9-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up1x9-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x9-scalar.c",
|
|
"src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x25-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up1x25-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up1x25-scalar.c",
|
|
"src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x4-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up2x4-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x4-scalar.c",
|
|
"src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x9-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up2x9-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x9-scalar.c",
|
|
"src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x25-minmax-scalar.c",
|
|
"src/f32-dwconv/gen/up2x25-scalar-acc2.c",
|
|
"src/f32-dwconv/gen/up2x25-scalar.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
|
|
"src/f32-gavgpool-cw/scalar-x1.c",
|
|
"src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
|
|
"src/f32-gavgpool/7x-minmax-scalar-c1.c",
|
|
"src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
|
|
"src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
|
|
"src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
|
|
"src/f32-gemm/gen/1x4-minmax-scalar.c",
|
|
"src/f32-gemm/gen/1x4-relu-scalar.c",
|
|
"src/f32-gemm/gen/1x4-scalar.c",
|
|
"src/f32-gemm/gen/2x4-minmax-scalar.c",
|
|
"src/f32-gemm/gen/2x4-relu-scalar.c",
|
|
"src/f32-gemm/gen/2x4-scalar.c",
|
|
"src/f32-gemm/gen/4x2-minmax-scalar.c",
|
|
"src/f32-gemm/gen/4x2-relu-scalar.c",
|
|
"src/f32-gemm/gen/4x2-scalar.c",
|
|
"src/f32-gemm/gen/4x4-minmax-scalar.c",
|
|
"src/f32-gemm/gen/4x4-relu-scalar.c",
|
|
"src/f32-gemm/gen/4x4-scalar.c",
|
|
"src/f32-hswish/gen/hswish-scalar-x1.c",
|
|
"src/f32-hswish/gen/hswish-scalar-x2.c",
|
|
"src/f32-hswish/gen/hswish-scalar-x4.c",
|
|
"src/f32-ibilinear-chw/gen/scalar-p1.c",
|
|
"src/f32-ibilinear-chw/gen/scalar-p2.c",
|
|
"src/f32-ibilinear-chw/gen/scalar-p4.c",
|
|
"src/f32-ibilinear/gen/scalar-c1.c",
|
|
"src/f32-ibilinear/gen/scalar-c2.c",
|
|
"src/f32-ibilinear/gen/scalar-c4.c",
|
|
"src/f32-igemm/gen/1x4-minmax-scalar.c",
|
|
"src/f32-igemm/gen/1x4-relu-scalar.c",
|
|
"src/f32-igemm/gen/1x4-scalar.c",
|
|
"src/f32-igemm/gen/2x4-minmax-scalar.c",
|
|
"src/f32-igemm/gen/2x4-relu-scalar.c",
|
|
"src/f32-igemm/gen/2x4-scalar.c",
|
|
"src/f32-igemm/gen/4x2-minmax-scalar.c",
|
|
"src/f32-igemm/gen/4x2-relu-scalar.c",
|
|
"src/f32-igemm/gen/4x2-scalar.c",
|
|
"src/f32-igemm/gen/4x4-minmax-scalar.c",
|
|
"src/f32-igemm/gen/4x4-relu-scalar.c",
|
|
"src/f32-igemm/gen/4x4-scalar.c",
|
|
"src/f32-maxpool/9p8x-minmax-scalar-c1.c",
|
|
"src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
|
|
"src/f32-pavgpool/9x-minmax-scalar-c1.c",
|
|
"src/f32-ppmm/gen/2x4-minmax-scalar.c",
|
|
"src/f32-ppmm/gen/3x3-minmax-scalar.c",
|
|
"src/f32-ppmm/gen/4x2-minmax-scalar.c",
|
|
"src/f32-ppmm/gen/4x4-minmax-scalar.c",
|
|
"src/f32-prelu/gen/scalar-2x1.c",
|
|
"src/f32-prelu/gen/scalar-2x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
|
|
"src/f32-relu/gen/scalar-x1.c",
|
|
"src/f32-relu/gen/scalar-x2.c",
|
|
"src/f32-relu/gen/scalar-x4.c",
|
|
"src/f32-relu/gen/scalar-x8.c",
|
|
"src/f32-rmax/scalar.c",
|
|
"src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
|
|
"src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
|
|
"src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
|
|
"src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
|
|
"src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
|
|
"src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
|
|
"src/f32-sigmoid/gen/scalar-p5-div-x1.c",
|
|
"src/f32-sigmoid/gen/scalar-p5-div-x2.c",
|
|
"src/f32-sigmoid/gen/scalar-p5-div-x4.c",
|
|
"src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
|
|
"src/f32-spmm/gen/1x1-minmax-scalar.c",
|
|
"src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
|
|
"src/f32-spmm/gen/2x1-minmax-scalar.c",
|
|
"src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
|
|
"src/f32-spmm/gen/4x1-minmax-scalar.c",
|
|
"src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
|
|
"src/f32-spmm/gen/8x1-minmax-scalar.c",
|
|
"src/f32-spmm/gen/8x2-minmax-scalar.c",
|
|
"src/f32-spmm/gen/8x4-minmax-scalar.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vadd-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vadd-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vadd-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vadd-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vaddc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vaddc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdiv-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdiv-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vdivc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vdivc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmaxc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmaxc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmin-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmin-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmin-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmin-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vminc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vminc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vminc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vminc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmul-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmul-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmul-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmul-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vmulc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vmulc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrdivc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrdivc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vrsubc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vrsubc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsub-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsub-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsub-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsub-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-scalar-x1.c",
|
|
"src/f32-vbinary/gen/vsubc-scalar-x2.c",
|
|
"src/f32-vbinary/gen/vsubc-scalar-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-scalar-x8.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
|
|
"src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
|
|
"src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
|
|
"src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
|
|
"src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
|
|
"src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
|
|
"src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
|
|
"src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
|
|
"src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
|
|
"src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
|
|
"src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
|
|
"src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
|
|
"src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
|
|
"src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
|
|
"src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
|
|
"src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
|
|
"src/f32-vsqrt/gen/scalar-sqrt-x1.c",
|
|
"src/f32-vsqrt/gen/scalar-sqrt-x2.c",
|
|
"src/f32-vsqrt/gen/scalar-sqrt-x4.c",
|
|
"src/f32-vunary/gen/vabs-scalar-x1.c",
|
|
"src/f32-vunary/gen/vabs-scalar-x2.c",
|
|
"src/f32-vunary/gen/vabs-scalar-x4.c",
|
|
"src/f32-vunary/gen/vneg-scalar-x1.c",
|
|
"src/f32-vunary/gen/vneg-scalar-x2.c",
|
|
"src/f32-vunary/gen/vneg-scalar-x4.c",
|
|
"src/f32-vunary/gen/vsqr-scalar-x1.c",
|
|
"src/f32-vunary/gen/vsqr-scalar-x2.c",
|
|
"src/f32-vunary/gen/vsqr-scalar-x4.c",
|
|
"src/math/expm1minus-scalar-rr2-lut4-p4.c",
|
|
"src/math/expm1minus-scalar-rr2-lut8-p3.c",
|
|
"src/math/expm1minus-scalar-rr2-lut8-p4.c",
|
|
"src/math/expm1minus-scalar-rr2-lut16-p3.c",
|
|
"src/math/expm1minus-scalar-rr2-lut16-p4.c",
|
|
"src/math/expm1minus-scalar-rr2-p5.c",
|
|
"src/math/expm1minus-scalar-rr2-p6.c",
|
|
"src/math/expminus-scalar-rr2-lut64-p2.c",
|
|
"src/math/expminus-scalar-rr2-lut2048-p1.c",
|
|
"src/math/expminus-scalar-rr2-p5.c",
|
|
"src/math/roundd-scalar-addsub.c",
|
|
"src/math/roundd-scalar-cvt.c",
|
|
"src/math/roundd-scalar-floor.c",
|
|
"src/math/roundne-scalar-addsub.c",
|
|
"src/math/roundne-scalar-nearbyint.c",
|
|
"src/math/roundne-scalar-rint.c",
|
|
"src/math/roundu-scalar-addsub.c",
|
|
"src/math/roundu-scalar-ceil.c",
|
|
"src/math/roundu-scalar-cvt.c",
|
|
"src/math/roundz-scalar-addsub.c",
|
|
"src/math/roundz-scalar-cvt.c",
|
|
"src/math/roundz-scalar-trunc.c",
|
|
"src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
|
|
"src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
|
|
"src/math/sigmoid-scalar-rr2-p5-div.c",
|
|
"src/qs8-requantization/fp32-scalar-lrintf.c",
|
|
"src/qs8-requantization/fp32-scalar-magic.c",
|
|
"src/qs8-requantization/precise-scalar-signed64.c",
|
|
"src/qs8-requantization/precise-scalar-unsigned32.c",
|
|
"src/qs8-requantization/precise-scalar-unsigned64.c",
|
|
"src/qs8-requantization/q31-scalar.c",
|
|
"src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
|
|
"src/qu8-avgpool/9x-minmax-scalar-c1.c",
|
|
"src/qu8-dwconv/up1x9-minmax-scalar.c",
|
|
"src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
|
|
"src/qu8-gavgpool/7x-minmax-scalar-c1.c",
|
|
"src/qu8-gemm/2x2-minmax-scalar.c",
|
|
"src/qu8-igemm/2x2-minmax-scalar.c",
|
|
"src/qu8-requantization/fp32-scalar-lrintf.c",
|
|
"src/qu8-requantization/fp32-scalar-magic.c",
|
|
"src/qu8-requantization/precise-scalar-signed64.c",
|
|
"src/qu8-requantization/precise-scalar-unsigned32.c",
|
|
"src/qu8-requantization/precise-scalar-unsigned64.c",
|
|
"src/qu8-requantization/q31-scalar.c",
|
|
"src/qu8-vadd/minmax-scalar.c",
|
|
"src/u8-clamp/scalar-x4.c",
|
|
"src/u8-lut32norm/scalar.c",
|
|
"src/u8-maxpool/9p8x-minmax-scalar-c1.c",
|
|
"src/u8-rmax/scalar.c",
|
|
"src/x8-lut/scalar.c",
|
|
"src/x8-zip/x2-scalar.c",
|
|
"src/x8-zip/x3-scalar.c",
|
|
"src/x8-zip/x4-scalar.c",
|
|
"src/x8-zip/xm-scalar.c",
|
|
"src/x32-depthtospace2d-chw2hwc/scalar.c",
|
|
"src/x32-fill/scalar-float.c",
|
|
"src/x32-fill/scalar-int.c",
|
|
"src/x32-packx/x2-scalar.c",
|
|
"src/x32-packx/x3-scalar.c",
|
|
"src/x32-packx/x4-scalar.c",
|
|
"src/x32-pad/scalar-float.c",
|
|
"src/x32-pad/scalar-int.c",
|
|
"src/x32-unpool/scalar.c",
|
|
"src/x32-zip/x2-scalar.c",
|
|
"src/x32-zip/x3-scalar.c",
|
|
"src/x32-zip/x4-scalar.c",
|
|
"src/x32-zip/xm-scalar.c",
|
|
"src/xx-copy/memcpy.c",
|
|
]
|
|
|
|
WASM_UKERNELS = [
|
|
"src/f32-avgpool/9p8x-minmax-wasm-c1.c",
|
|
"src/f32-avgpool/9x-minmax-wasm-c1.c",
|
|
"src/f32-clamp/gen/wasm-x1.c",
|
|
"src/f32-clamp/gen/wasm-x2.c",
|
|
"src/f32-clamp/gen/wasm-x4.c",
|
|
"src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x4-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up1x4-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x4-wasm.c",
|
|
"src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x9-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up1x9-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x9-wasm.c",
|
|
"src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x25-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up1x25-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up1x25-wasm.c",
|
|
"src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x4-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up2x4-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x4-wasm.c",
|
|
"src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x9-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up2x9-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x9-wasm.c",
|
|
"src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x25-minmax-wasm.c",
|
|
"src/f32-dwconv/gen/up2x25-wasm-acc2.c",
|
|
"src/f32-dwconv/gen/up2x25-wasm.c",
|
|
"src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
|
|
"src/f32-gavgpool/7x-minmax-wasm-c1.c",
|
|
"src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
|
|
"src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
|
|
"src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
|
|
"src/f32-gemm/gen/1x4-minmax-wasm.c",
|
|
"src/f32-gemm/gen/1x4-relu-wasm.c",
|
|
"src/f32-gemm/gen/1x4-wasm.c",
|
|
"src/f32-gemm/gen/2x4-minmax-wasm.c",
|
|
"src/f32-gemm/gen/2x4-relu-wasm.c",
|
|
"src/f32-gemm/gen/2x4-wasm.c",
|
|
"src/f32-gemm/gen/4x2-minmax-wasm.c",
|
|
"src/f32-gemm/gen/4x2-relu-wasm.c",
|
|
"src/f32-gemm/gen/4x2-wasm.c",
|
|
"src/f32-gemm/gen/4x4-minmax-wasm.c",
|
|
"src/f32-gemm/gen/4x4-relu-wasm.c",
|
|
"src/f32-gemm/gen/4x4-wasm.c",
|
|
"src/f32-hswish/gen/hswish-wasm-x1.c",
|
|
"src/f32-hswish/gen/hswish-wasm-x2.c",
|
|
"src/f32-hswish/gen/hswish-wasm-x4.c",
|
|
"src/f32-igemm/gen/1x4-minmax-wasm.c",
|
|
"src/f32-igemm/gen/1x4-relu-wasm.c",
|
|
"src/f32-igemm/gen/1x4-wasm.c",
|
|
"src/f32-igemm/gen/2x4-minmax-wasm.c",
|
|
"src/f32-igemm/gen/2x4-relu-wasm.c",
|
|
"src/f32-igemm/gen/2x4-wasm.c",
|
|
"src/f32-igemm/gen/4x2-minmax-wasm.c",
|
|
"src/f32-igemm/gen/4x2-relu-wasm.c",
|
|
"src/f32-igemm/gen/4x2-wasm.c",
|
|
"src/f32-igemm/gen/4x4-minmax-wasm.c",
|
|
"src/f32-igemm/gen/4x4-relu-wasm.c",
|
|
"src/f32-igemm/gen/4x4-wasm.c",
|
|
"src/f32-maxpool/9p8x-minmax-wasm-c1.c",
|
|
"src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
|
|
"src/f32-pavgpool/9x-minmax-wasm-c1.c",
|
|
"src/f32-prelu/gen/wasm-2x1.c",
|
|
"src/f32-prelu/gen/wasm-2x4.c",
|
|
"src/f32-relu/gen/wasm-x1.c",
|
|
"src/f32-relu/gen/wasm-x2.c",
|
|
"src/f32-relu/gen/wasm-x4.c",
|
|
"src/f32-relu/gen/wasm-x8.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmin-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmin-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmin-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmin-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vminc-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vminc-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vminc-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vminc-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
|
|
"src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
|
|
"src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
|
|
"src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
|
|
]
|
|
|
|
WASMSIMD_UKERNELS = [
|
|
"src/f32-argmaxpool/4x-wasmsimd-c4.c",
|
|
"src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
|
|
"src/f32-argmaxpool/9x-wasmsimd-c4.c",
|
|
"src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-clamp/gen/wasmsimd-arm-x4.c",
|
|
"src/f32-clamp/gen/wasmsimd-arm-x8.c",
|
|
"src/f32-clamp/gen/wasmsimd-x86-x4.c",
|
|
"src/f32-clamp/gen/wasmsimd-x86-x8.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up4x4-wasmsimd.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up4x9-wasmsimd.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up4x25-wasmsimd.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up8x4-wasmsimd.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up8x9-wasmsimd.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
|
|
"src/f32-dwconv/gen/up8x25-wasmsimd.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-splat-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-splat-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-splat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-splat-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c",
|
|
"src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
|
|
"src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
|
|
"src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/1x8-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
|
|
"src/f32-gemm/gen/4x2c4-wasmsimd.c",
|
|
"src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/4x8-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/5x8-wasmsimd-splat.c",
|
|
"src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-gemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-gemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-hswish/gen/hswish-wasmsimd-x4.c",
|
|
"src/f32-hswish/gen/hswish-wasmsimd-x8.c",
|
|
"src/f32-hswish/gen/hswish-wasmsimd-x16.c",
|
|
"src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
|
|
"src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
|
|
"src/f32-ibilinear/gen/wasmsimd-c4.c",
|
|
"src/f32-ibilinear/gen/wasmsimd-c8.c",
|
|
"src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-igemm/gen/1x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-igemm/gen/1x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/1x8-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-igemm/gen/3x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-igemm/gen/3x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
|
|
"src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
|
|
"src/f32-igemm/gen/4x2c4-wasmsimd.c",
|
|
"src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-igemm/gen/4x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-igemm/gen/4x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/4x8-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-igemm/gen/5x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-igemm/gen/5x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/5x8-wasmsimd-splat.c",
|
|
"src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-loadsplat.c",
|
|
"src/f32-igemm/gen/6x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-loadsplat.c",
|
|
"src/f32-igemm/gen/6x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
|
|
"src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
|
|
"src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
|
|
"src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
|
|
"src/f32-ppmm/gen/4x8-minmax-wasmsimd-arm-splat.c",
|
|
"src/f32-ppmm/gen/4x8-minmax-wasmsimd-x86-splat.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
|
|
"src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
|
|
"src/f32-relu/gen/wasmsimd-x4.c",
|
|
"src/f32-relu/gen/wasmsimd-x8.c",
|
|
"src/f32-relu/gen/wasmsimd-x16.c",
|
|
"src/f32-rmax/wasmsimd-arm.c",
|
|
"src/f32-rmax/wasmsimd-x86.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
|
|
"src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
|
|
"src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
|
|
"src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
|
|
"src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
|
|
"src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
|
|
"src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
|
|
"src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
|
|
"src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
|
|
"src/f32-vunary/gen/vabs-wasmsimd-x4.c",
|
|
"src/f32-vunary/gen/vabs-wasmsimd-x8.c",
|
|
"src/f32-vunary/gen/vneg-wasmsimd-x4.c",
|
|
"src/f32-vunary/gen/vneg-wasmsimd-x8.c",
|
|
"src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
|
|
"src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
|
|
"src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
|
|
"src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
|
|
"src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
|
|
"src/math/expm1minus-wasmsimd-rr2-p6-max.c",
|
|
"src/math/roundd-wasmsimd-addsub.c",
|
|
"src/math/roundd-wasmsimd-cvt.c",
|
|
"src/math/roundne-wasmsimd-addsub.c",
|
|
"src/math/roundu-wasmsimd-addsub.c",
|
|
"src/math/roundu-wasmsimd-cvt.c",
|
|
"src/math/roundz-wasmsimd-addsub.c",
|
|
"src/math/roundz-wasmsimd-cvt.c",
|
|
"src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
|
|
"src/math/sigmoid-wasmsimd-rr2-p5-div.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
|
|
"src/qs8-requantization/fp32-wasmsimd.c",
|
|
"src/qs8-requantization/q31-wasmsimd.c",
|
|
"src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
|
|
"src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
|
|
"src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
|
|
"src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
|
|
"src/qu8-requantization/fp32-wasmsimd.c",
|
|
"src/qu8-requantization/q31-wasmsimd.c",
|
|
"src/x32-fill/wasmsimd.c",
|
|
"src/x32-packx/x4-wasmsimd.c",
|
|
"src/x32-pad/wasmsimd.c",
|
|
"src/x32-unpool/wasmsimd.c",
|
|
"src/x32-zip/x2-wasmsimd.c",
|
|
"src/x32-zip/x3-wasmsimd.c",
|
|
"src/x32-zip/x4-wasmsimd.c",
|
|
"src/x32-zip/xm-wasmsimd.c",
|
|
]
|
|
|
|
# ISA-specific micro-kernels
|
|
NEON_UKERNELS = [
|
|
"src/f32-argmaxpool/4x-neon-c4.c",
|
|
"src/f32-argmaxpool/9p8x-neon-c4.c",
|
|
"src/f32-argmaxpool/9x-neon-c4.c",
|
|
"src/f32-avgpool/9p8x-minmax-neon-c4.c",
|
|
"src/f32-avgpool/9x-minmax-neon-c4.c",
|
|
"src/f32-clamp/gen/neon-x4.c",
|
|
"src/f32-clamp/gen/neon-x8.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-neon.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-neon.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-neon.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-neon.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-neon.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-neon.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
|
|
"src/f32-gavgpool-cw/neon-x4.c",
|
|
"src/f32-gavgpool/7p7x-minmax-neon-c4.c",
|
|
"src/f32-gavgpool/7x-minmax-neon-c4.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
|
|
"src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
|
|
"src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
|
|
"src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
|
|
"src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen/1x8s4-minmax-neon.c",
|
|
"src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
|
|
"src/f32-gemm/gen/4x8s4-minmax-neon.c",
|
|
"src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
|
|
"src/f32-gemm/gen/6x8s4-minmax-neon.c",
|
|
"src/f32-gemm/gen/8x8s4-minmax-neon.c",
|
|
"src/f32-hswish/gen/hswish-neon-x4.c",
|
|
"src/f32-hswish/gen/hswish-neon-x8.c",
|
|
"src/f32-hswish/gen/hswish-neon-x16.c",
|
|
"src/f32-ibilinear-chw/gen/neon-p4.c",
|
|
"src/f32-ibilinear-chw/gen/neon-p8.c",
|
|
"src/f32-ibilinear/gen/neon-c4.c",
|
|
"src/f32-ibilinear/gen/neon-c8.c",
|
|
"src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-igemm/gen/1x8s4-minmax-neon.c",
|
|
"src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
|
|
"src/f32-igemm/gen/4x8s4-minmax-neon.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
|
|
"src/f32-igemm/gen/6x8s4-minmax-neon.c",
|
|
"src/f32-igemm/gen/8x8s4-minmax-neon.c",
|
|
"src/f32-maxpool/9p8x-minmax-neon-c4.c",
|
|
"src/f32-pavgpool/9p8x-minmax-neon-c4.c",
|
|
"src/f32-pavgpool/9x-minmax-neon-c4.c",
|
|
"src/f32-ppmm/gen/4x8-minmax-neon.c",
|
|
"src/f32-ppmm/gen/8x8-minmax-neon.c",
|
|
"src/f32-prelu/gen/neon-1x4.c",
|
|
"src/f32-prelu/gen/neon-1x8.c",
|
|
"src/f32-prelu/gen/neon-1x16.c",
|
|
"src/f32-prelu/gen/neon-2x4.c",
|
|
"src/f32-prelu/gen/neon-2x8.c",
|
|
"src/f32-prelu/gen/neon-2x16.c",
|
|
"src/f32-prelu/gen/neon-4x4.c",
|
|
"src/f32-prelu/gen/neon-4x8.c",
|
|
"src/f32-prelu/gen/neon-4x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
|
|
"src/f32-relu/gen/neon-x4.c",
|
|
"src/f32-relu/gen/neon-x8.c",
|
|
"src/f32-rmax/neon.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neon-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neon.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neon-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neon.c",
|
|
"src/f32-spmm/gen/12x1-minmax-neon.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neon-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neon.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neon-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neon.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-neon-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-neon-x8.c",
|
|
"src/f32-vbinary/gen/vmin-neon-x4.c",
|
|
"src/f32-vbinary/gen/vmin-neon-x8.c",
|
|
"src/f32-vbinary/gen/vminc-neon-x4.c",
|
|
"src/f32-vbinary/gen/vminc-neon-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
|
|
"src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
|
|
"src/f32-vlrelu/gen/vlrelu-neon-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-neon-x8.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
|
|
"src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
|
|
"src/f32-vrnd/gen/vrndd-neon-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-neon-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-neon-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-neon-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-neon-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-neon-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-neon-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-neon-x8.c",
|
|
"src/f32-vunary/gen/vabs-neon-x4.c",
|
|
"src/f32-vunary/gen/vabs-neon-x8.c",
|
|
"src/f32-vunary/gen/vneg-neon-x4.c",
|
|
"src/f32-vunary/gen/vneg-neon-x8.c",
|
|
"src/f32-vunary/gen/vsqr-neon-x4.c",
|
|
"src/f32-vunary/gen/vsqr-neon-x8.c",
|
|
"src/math/expm1minus-neon-rr2-lut16-p3.c",
|
|
"src/math/expm1minus-neon-rr2-p6.c",
|
|
"src/math/roundd-neon-addsub.c",
|
|
"src/math/roundd-neon-cvt.c",
|
|
"src/math/roundne-neon-addsub.c",
|
|
"src/math/roundu-neon-addsub.c",
|
|
"src/math/roundu-neon-cvt.c",
|
|
"src/math/roundz-neon-addsub.c",
|
|
"src/math/roundz-neon-cvt.c",
|
|
"src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
|
|
"src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
|
|
"src/math/sigmoid-neon-rr2-p5-nr2recps.c",
|
|
"src/math/sqrt-neon-nr1rsqrts.c",
|
|
"src/math/sqrt-neon-nr2rsqrts.c",
|
|
"src/math/sqrt-neon-nr3rsqrts.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
|
|
"src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
|
|
"src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
|
|
"src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
|
|
"src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
|
|
"src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
|
|
"src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
|
|
"src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
|
|
"src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
|
|
"src/qs8-requantization/fp32-neon.c",
|
|
"src/qs8-requantization/precise-neon.c",
|
|
"src/qs8-requantization/q31-neon.c",
|
|
"src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
|
|
"src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
|
|
"src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
|
|
"src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
|
|
"src/qu8-avgpool/9p8x-minmax-neon-c8.c",
|
|
"src/qu8-avgpool/9x-minmax-neon-c8.c",
|
|
"src/qu8-dwconv/up8x9-minmax-neon.c",
|
|
"src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
|
|
"src/qu8-gavgpool/7x-minmax-neon-c8.c",
|
|
"src/qu8-gemm/4x8-minmax-neon.c",
|
|
"src/qu8-gemm/8x8-minmax-neon.c",
|
|
"src/qu8-igemm/4x8-minmax-neon.c",
|
|
"src/qu8-igemm/8x8-minmax-neon.c",
|
|
"src/qu8-requantization/fp32-neon.c",
|
|
"src/qu8-requantization/precise-neon.c",
|
|
"src/qu8-requantization/q31-neon.c",
|
|
"src/qu8-vadd/minmax-neon.c",
|
|
"src/u8-clamp/neon-x64.c",
|
|
"src/u8-maxpool/9p8x-minmax-neon-c16.c",
|
|
"src/u8-rmax/neon.c",
|
|
"src/x8-zip/x2-neon.c",
|
|
"src/x8-zip/x3-neon.c",
|
|
"src/x8-zip/x4-neon.c",
|
|
"src/x8-zip/xm-neon.c",
|
|
"src/x32-fill/neon.c",
|
|
"src/x32-packx/x4-neon-st4.c",
|
|
"src/x32-pad/neon.c",
|
|
"src/x32-unpool/neon.c",
|
|
"src/x32-zip/x2-neon.c",
|
|
"src/x32-zip/x3-neon.c",
|
|
"src/x32-zip/x4-neon.c",
|
|
"src/x32-zip/xm-neon.c",
|
|
]
|
|
|
|
NEONFMA_UKERNELS = [
|
|
"src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
|
|
"src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
|
|
"src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
|
|
"src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
|
|
"src/f32-ibilinear-chw/gen/neonfma-p4.c",
|
|
"src/f32-ibilinear-chw/gen/neonfma-p8.c",
|
|
"src/f32-ibilinear/gen/neonfma-c4.c",
|
|
"src/f32-ibilinear/gen/neonfma-c8.c",
|
|
"src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
|
|
"src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
|
|
"src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
|
|
"src/f32-ppmm/gen/4x8-minmax-neonfma.c",
|
|
"src/f32-ppmm/gen/8x8-minmax-neonfma.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
|
|
"src/f32-spmm/gen/4x1-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
|
|
"src/f32-spmm/gen/8x1-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/12x1-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
|
|
"src/f32-spmm/gen/16x1-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
|
|
"src/f32-spmm/gen/32x1-minmax-neonfma.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
|
|
"src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
|
|
"src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
|
|
"src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
|
|
"src/math/exp-neonfma-rr2-lut64-p2.c",
|
|
"src/math/exp-neonfma-rr2-p5.c",
|
|
"src/math/expm1minus-neonfma-rr1-lut16-p3.c",
|
|
"src/math/expm1minus-neonfma-rr1-p6.c",
|
|
"src/math/expminus-neonfma-rr2-lut64-p2.c",
|
|
"src/math/expminus-neonfma-rr2-lut2048-p1.c",
|
|
"src/math/expminus-neonfma-rr2-p5.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
|
|
"src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
|
|
"src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
|
|
"src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
|
|
"src/math/sqrt-neonfma-nr1fma.c",
|
|
"src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
|
|
"src/math/sqrt-neonfma-nr2fma.c",
|
|
"src/math/sqrt-neonfma-nr2fma1adj.c",
|
|
"src/math/sqrt-neonfma-nr3fma.c",
|
|
]
|
|
|
|
AARCH64_NEONFMA_UKERNELS = [
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
|
|
"src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
|
|
"src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
|
|
"src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
|
|
"src/f32-spmm/gen/4x2-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/4x4-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/8x2-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/8x4-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/12x2-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/12x4-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/16x2-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/16x4-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/32x2-minmax-neonfma.c",
|
|
"src/f32-spmm/gen/32x4-minmax-neonfma.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
|
|
"src/f32-vsqrt/gen/neon-sqrt-x4.c",
|
|
"src/f32-vsqrt/gen/neon-sqrt-x8.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
|
|
"src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
|
|
"src/math/sigmoid-neonfma-rr1-p5-div.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
|
|
"src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
|
|
"src/math/sigmoid-neonfma-rr2-p5-div.c",
|
|
]
|
|
|
|
NEONV8_UKERNELS = [
|
|
"src/f32-vrnd/gen/vrndd-neonv8-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-neonv8-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-neonv8-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-neonv8-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-neonv8-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-neonv8-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-neonv8-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-neonv8-x8.c",
|
|
"src/math/roundd-neonv8.c",
|
|
"src/math/roundne-neonv8.c",
|
|
"src/math/roundu-neonv8.c",
|
|
"src/math/roundz-neonv8.c",
|
|
]
|
|
|
|
AARCH64_NEONFP16ARITH_UKERNELS = [
|
|
"src/f16-clamp/gen/neonfp16arith-x8.c",
|
|
"src/f16-clamp/gen/neonfp16arith-x16.c",
|
|
"src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
|
|
"src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
|
|
"src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
|
|
"src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
|
|
"src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
|
|
"src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
|
|
"src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
|
|
"src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
|
|
"src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
|
|
"src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
|
|
"src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
|
|
"src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
|
|
"src/f16-prelu/gen/neonfp16arith-2x8.c",
|
|
"src/f16-prelu/gen/neonfp16arith-2x16.c",
|
|
"src/f16-relu/gen/neonfp16arith-x8.c",
|
|
"src/f16-relu/gen/neonfp16arith-x16.c",
|
|
"src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
|
|
"src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
|
|
"src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
|
|
"src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
|
|
"src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
|
|
"src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
|
|
"src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
|
|
"src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
|
|
"src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
|
|
"src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
|
|
"src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
|
|
"src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
|
|
]
|
|
|
|
NEONDOT_UKERNELS = [
|
|
"src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
|
|
"src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
|
|
"src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
|
|
]
|
|
|
|
SSE_UKERNELS = [
|
|
"src/f32-avgpool/9p8x-minmax-sse-c4.c",
|
|
"src/f32-avgpool/9x-minmax-sse-c4.c",
|
|
"src/f32-clamp/gen/sse-x4.c",
|
|
"src/f32-clamp/gen/sse-x8.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
|
|
"src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up4x4-minmax-sse.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up4x9-minmax-sse.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up4x25-minmax-sse.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-sse.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-sse.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-sse.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
|
|
"src/f32-gavgpool-cw/sse-x4.c",
|
|
"src/f32-gavgpool/7p7x-minmax-sse-c4.c",
|
|
"src/f32-gavgpool/7x-minmax-sse-c4.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
|
|
"src/f32-gemm/gen/1x8-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen/1x8-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen/1x8s4-minmax-sse.c",
|
|
"src/f32-gemm/gen/3x8-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen/3x8-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen/3x8s4-minmax-sse.c",
|
|
"src/f32-gemm/gen/4x2c4-minmax-sse.c",
|
|
"src/f32-gemm/gen/4x8-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen/4x8-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen/4x8s4-minmax-sse.c",
|
|
"src/f32-gemm/gen/5x8-minmax-sse-dup.c",
|
|
"src/f32-gemm/gen/5x8-minmax-sse-load1.c",
|
|
"src/f32-gemm/gen/5x8s4-minmax-sse.c",
|
|
"src/f32-hswish/gen/hswish-sse-x4.c",
|
|
"src/f32-hswish/gen/hswish-sse-x8.c",
|
|
"src/f32-ibilinear/gen/sse-c4.c",
|
|
"src/f32-ibilinear/gen/sse-c8.c",
|
|
"src/f32-igemm/gen/1x8-minmax-sse-dup.c",
|
|
"src/f32-igemm/gen/1x8-minmax-sse-load1.c",
|
|
"src/f32-igemm/gen/1x8s4-minmax-sse.c",
|
|
"src/f32-igemm/gen/3x8-minmax-sse-dup.c",
|
|
"src/f32-igemm/gen/3x8-minmax-sse-load1.c",
|
|
"src/f32-igemm/gen/3x8s4-minmax-sse.c",
|
|
"src/f32-igemm/gen/4x2c4-minmax-sse.c",
|
|
"src/f32-igemm/gen/4x8-minmax-sse-dup.c",
|
|
"src/f32-igemm/gen/4x8-minmax-sse-load1.c",
|
|
"src/f32-igemm/gen/4x8s4-minmax-sse.c",
|
|
"src/f32-igemm/gen/5x8-minmax-sse-dup.c",
|
|
"src/f32-igemm/gen/5x8-minmax-sse-load1.c",
|
|
"src/f32-igemm/gen/5x8s4-minmax-sse.c",
|
|
"src/f32-maxpool/9p8x-minmax-sse-c4.c",
|
|
"src/f32-pavgpool/9p8x-minmax-sse-c4.c",
|
|
"src/f32-pavgpool/9x-minmax-sse-c4.c",
|
|
"src/f32-ppmm/gen/4x8-minmax-sse.c",
|
|
"src/f32-prelu/gen/sse-2x4.c",
|
|
"src/f32-prelu/gen/sse-2x8.c",
|
|
"src/f32-relu/gen/sse-x4.c",
|
|
"src/f32-relu/gen/sse-x8.c",
|
|
"src/f32-rmax/sse.c",
|
|
"src/f32-spmm/gen/4x1-minmax-sse.c",
|
|
"src/f32-spmm/gen/8x1-minmax-sse.c",
|
|
"src/f32-spmm/gen/16x1-minmax-sse.c",
|
|
"src/f32-spmm/gen/32x1-minmax-sse.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-sse-x4.c",
|
|
"src/f32-vbinary/gen/vmaxc-sse-x8.c",
|
|
"src/f32-vbinary/gen/vmin-sse-x4.c",
|
|
"src/f32-vbinary/gen/vmin-sse-x8.c",
|
|
"src/f32-vbinary/gen/vminc-sse-x4.c",
|
|
"src/f32-vbinary/gen/vminc-sse-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse-x8.c",
|
|
"src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
|
|
"src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
|
|
"src/f32-vsqrt/gen/sse-sqrt-x4.c",
|
|
"src/f32-vsqrt/gen/sse-sqrt-x8.c",
|
|
"src/f32-vunary/gen/vabs-sse-x4.c",
|
|
"src/f32-vunary/gen/vabs-sse-x8.c",
|
|
"src/f32-vunary/gen/vneg-sse-x4.c",
|
|
"src/f32-vunary/gen/vneg-sse-x8.c",
|
|
"src/f32-vunary/gen/vsqr-sse-x4.c",
|
|
"src/f32-vunary/gen/vsqr-sse-x8.c",
|
|
"src/math/roundd-sse-addsub.c",
|
|
"src/math/roundne-sse-addsub.c",
|
|
"src/math/roundu-sse-addsub.c",
|
|
"src/math/roundz-sse-addsub.c",
|
|
"src/math/sqrt-sse-hh1mac.c",
|
|
"src/math/sqrt-sse-nr1mac.c",
|
|
"src/math/sqrt-sse-nr2mac.c",
|
|
"src/x32-fill/sse.c",
|
|
"src/x32-packx/x4-sse.c",
|
|
"src/x32-pad/sse.c",
|
|
]
|
|
|
|
SSE2_UKERNELS = [
|
|
"src/f32-argmaxpool/4x-sse2-c4.c",
|
|
"src/f32-argmaxpool/9p8x-sse2-c4.c",
|
|
"src/f32-argmaxpool/9x-sse2-c4.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
|
|
"src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
|
|
"src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
|
|
"src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
|
|
"src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
|
|
"src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
|
|
"src/f32-prelu/gen/sse2-2x4.c",
|
|
"src/f32-prelu/gen/sse2-2x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
|
|
"src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x4.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x12.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x20.c",
|
|
"src/f32-sigmoid/gen/sse2-p5-div-x24.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
|
|
"src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
|
|
"src/f32-vrnd/gen/vrndd-sse2-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-sse2-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-sse2-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-sse2-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-sse2-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-sse2-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-sse2-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-sse2-x8.c",
|
|
"src/math/exp-sse2-rr2-lut64-p2.c",
|
|
"src/math/exp-sse2-rr2-p5.c",
|
|
"src/math/expm1minus-sse2-rr2-lut16-p3.c",
|
|
"src/math/expm1minus-sse2-rr2-p6.c",
|
|
"src/math/expminus-sse2-rr2-p5.c",
|
|
"src/math/roundd-sse2-cvt.c",
|
|
"src/math/roundne-sse2-cvt.c",
|
|
"src/math/roundu-sse2-cvt.c",
|
|
"src/math/roundz-sse2-cvt.c",
|
|
"src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
|
|
"src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
|
|
"src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
|
|
"src/math/sigmoid-sse2-rr2-p5-div.c",
|
|
"src/math/sigmoid-sse2-rr2-p5-nr1.c",
|
|
"src/math/sigmoid-sse2-rr2-p5-nr2.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
|
|
"src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
|
|
"src/qs8-requantization/fp32-sse2.c",
|
|
"src/qs8-requantization/precise-sse2.c",
|
|
"src/qs8-requantization/q31-sse2.c",
|
|
"src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
|
|
"src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
|
|
"src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
|
|
"src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
|
|
"src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
|
|
"src/qu8-avgpool/9x-minmax-sse2-c8.c",
|
|
"src/qu8-dwconv/up8x9-minmax-sse2.c",
|
|
"src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
|
|
"src/qu8-gavgpool/7x-minmax-sse2-c8.c",
|
|
"src/qu8-gemm/2x4c8-minmax-sse2.c",
|
|
"src/qu8-gemm/4x4c2-minmax-sse2.c",
|
|
"src/qu8-igemm/4x4c2-minmax-sse2.c",
|
|
"src/qu8-requantization/fp32-sse2.c",
|
|
"src/qu8-requantization/precise-sse2.c",
|
|
"src/qu8-requantization/q31-sse2.c",
|
|
"src/qu8-vadd/minmax-sse2.c",
|
|
"src/u8-clamp/sse2-x64.c",
|
|
"src/u8-maxpool/9p8x-minmax-sse2-c16.c",
|
|
"src/u8-rmax/sse2.c",
|
|
"src/x8-zip/x2-sse2.c",
|
|
"src/x8-zip/x3-sse2.c",
|
|
"src/x8-zip/x4-sse2.c",
|
|
"src/x8-zip/xm-sse2.c",
|
|
"src/x32-unpool/sse2.c",
|
|
"src/x32-zip/x2-sse2.c",
|
|
"src/x32-zip/x3-sse2.c",
|
|
"src/x32-zip/x4-sse2.c",
|
|
"src/x32-zip/xm-sse2.c",
|
|
]
|
|
|
|
SSSE3_UKERNELS = [
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
|
|
"src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
|
|
"src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
|
|
"src/qs8-requantization/precise-ssse3.c",
|
|
"src/qs8-requantization/q31-ssse3.c",
|
|
"src/qu8-requantization/precise-ssse3.c",
|
|
"src/qu8-requantization/q31-ssse3.c",
|
|
]
|
|
|
|
SSE41_UKERNELS = [
|
|
"src/f32-prelu/gen/sse41-2x4.c",
|
|
"src/f32-prelu/gen/sse41-2x8.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
|
|
"src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x4.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x12.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x20.c",
|
|
"src/f32-sigmoid/gen/sse41-p5-div-x24.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
|
|
"src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
|
|
"src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
|
|
"src/f32-vrnd/gen/vrndd-sse41-x4.c",
|
|
"src/f32-vrnd/gen/vrndd-sse41-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-sse41-x4.c",
|
|
"src/f32-vrnd/gen/vrndne-sse41-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-sse41-x4.c",
|
|
"src/f32-vrnd/gen/vrndu-sse41-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-sse41-x4.c",
|
|
"src/f32-vrnd/gen/vrndz-sse41-x8.c",
|
|
"src/math/roundd-sse41.c",
|
|
"src/math/roundne-sse41.c",
|
|
"src/math/roundu-sse41.c",
|
|
"src/math/roundz-sse41.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
|
|
"src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
|
|
"src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
|
|
"src/qs8-requantization/fp32-sse4.c",
|
|
"src/qs8-requantization/precise-sse4.c",
|
|
"src/qs8-requantization/q31-sse4.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
|
|
"src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
|
|
"src/qu8-requantization/precise-sse4.c",
|
|
"src/qu8-requantization/q31-sse4.c",
|
|
]
|
|
|
|
AVX_UKERNELS = [
|
|
"src/f32-clamp/gen/avx-x8.c",
|
|
"src/f32-clamp/gen/avx-x16.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-avx.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-avx.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-avx.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-avx.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-avx.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-avx.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
|
|
"src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
|
|
"src/f32-hswish/gen/hswish-avx-x8.c",
|
|
"src/f32-hswish/gen/hswish-avx-x16.c",
|
|
"src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
|
|
"src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
|
|
"src/f32-prelu/gen/avx-2x8.c",
|
|
"src/f32-prelu/gen/avx-2x16.c",
|
|
"src/f32-relu/gen/avx-x8.c",
|
|
"src/f32-relu/gen/avx-x16.c",
|
|
"src/f32-rmax/avx.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
|
|
"src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vmaxc-avx-x8.c",
|
|
"src/f32-vbinary/gen/vmaxc-avx-x16.c",
|
|
"src/f32-vbinary/gen/vmin-avx-x8.c",
|
|
"src/f32-vbinary/gen/vmin-avx-x16.c",
|
|
"src/f32-vbinary/gen/vminc-avx-x8.c",
|
|
"src/f32-vbinary/gen/vminc-avx-x16.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
|
|
"src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
|
|
"src/f32-vlrelu/gen/vlrelu-avx-x8.c",
|
|
"src/f32-vlrelu/gen/vlrelu-avx-x16.c",
|
|
"src/f32-vrnd/gen/vrndd-avx-x8.c",
|
|
"src/f32-vrnd/gen/vrndd-avx-x16.c",
|
|
"src/f32-vrnd/gen/vrndne-avx-x8.c",
|
|
"src/f32-vrnd/gen/vrndne-avx-x16.c",
|
|
"src/f32-vrnd/gen/vrndu-avx-x8.c",
|
|
"src/f32-vrnd/gen/vrndu-avx-x16.c",
|
|
"src/f32-vrnd/gen/vrndz-avx-x8.c",
|
|
"src/f32-vrnd/gen/vrndz-avx-x16.c",
|
|
"src/f32-vscale/avx-x32.c",
|
|
"src/f32-vsqrt/gen/avx-sqrt-x8.c",
|
|
"src/f32-vsqrt/gen/avx-sqrt-x16.c",
|
|
"src/f32-vunary/gen/vabs-avx-x8.c",
|
|
"src/f32-vunary/gen/vabs-avx-x16.c",
|
|
"src/f32-vunary/gen/vneg-avx-x8.c",
|
|
"src/f32-vunary/gen/vneg-avx-x16.c",
|
|
"src/f32-vunary/gen/vsqr-avx-x8.c",
|
|
"src/f32-vunary/gen/vsqr-avx-x16.c",
|
|
"src/math/exp-avx-rr2-p5.c",
|
|
"src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
|
|
"src/math/expm1minus-avx-rr2-lut16-p3.c",
|
|
"src/math/expm1minus-avx-rr2-p6.c",
|
|
"src/math/sigmoid-avx-rr2-lut64-p2-div.c",
|
|
"src/math/sigmoid-avx-rr2-p5-div.c",
|
|
"src/math/sigmoid-avx-rr2-p5-nr1.c",
|
|
"src/math/sigmoid-avx-rr2-p5-nr2.c",
|
|
]
|
|
|
|
XOP_UKERNELS = [
|
|
"src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
|
|
"src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
|
|
"src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
|
|
"src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
|
|
"src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
|
|
"src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
|
|
"src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
|
|
"src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
|
|
"src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
|
|
]
|
|
|
|
FMA3_UKERNELS = [
|
|
"src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up8x4-minmax-fma3.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up8x9-minmax-fma3.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up8x25-minmax-fma3.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-fma3.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-fma3.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-fma3.c",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
|
|
"src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
|
|
"src/f32-hswish/gen/hswish-fma3-x8.c",
|
|
"src/f32-hswish/gen/hswish-fma3-x16.c",
|
|
"src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
|
|
"src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
|
|
"src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
|
|
"src/math/sqrt-fma3-nr1fma.c",
|
|
"src/math/sqrt-fma3-nr1fma1adj.c",
|
|
"src/math/sqrt-fma3-nr2fma.c",
|
|
]
|
|
|
|
AVX2_UKERNELS = [
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
|
|
"src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x64.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x72.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x80.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
|
|
"src/f32-raddextexp/gen/avx2-p5-x96.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
|
|
"src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
|
|
"src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x8.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x16.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x24.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x32.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x40.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x48.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x56.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x64.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x72.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x80.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x88.c",
|
|
"src/f32-vscaleextexp/gen/avx2-p5-x96.c",
|
|
"src/math/exp-avx2-rr2-lut8-p3-perm.c",
|
|
"src/math/exp-avx2-rr2-lut8-p4-perm.c",
|
|
"src/math/exp-avx2-rr2-p5.c",
|
|
"src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
|
|
"src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
|
|
"src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
|
|
"src/math/expm1minus-avx2-rr1-p6.c",
|
|
"src/math/expminus-avx2-rr2-p5.c",
|
|
"src/math/extexp-avx2-p5.c",
|
|
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
|
|
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
|
|
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
|
|
"src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
|
|
"src/math/sigmoid-avx2-rr1-p5-div.c",
|
|
"src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
|
|
"src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
|
|
"src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
|
|
"src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
|
|
"src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
|
|
"src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
|
|
"src/math/sigmoid-avx2-rr2-p5-div.c",
|
|
"src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
|
|
"src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
|
|
"src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
|
|
"src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
|
|
"src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
|
|
"src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
|
|
"src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
|
|
"src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
|
|
"src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
|
|
"src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
|
|
"src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
|
|
"src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
|
|
"src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
|
|
"src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
|
|
"src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
|
|
"src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
|
|
"src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
|
|
"src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
|
|
"src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
|
|
"src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
|
|
"src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
|
|
"src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
|
|
"src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
|
|
"src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
|
|
]
|
|
|
|
AVX512F_UKERNELS = [
|
|
"src/f32-clamp/gen/avx512f-x16.c",
|
|
"src/f32-clamp/gen/avx512f-x32.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
|
|
"src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
|
|
"src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
|
|
"src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
|
|
"src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
|
|
"src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-hswish/gen/hswish-avx512f-x16.c",
|
|
"src/f32-hswish/gen/hswish-avx512f-x32.c",
|
|
"src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
|
|
"src/f32-prelu/gen/avx512f-2x16.c",
|
|
"src/f32-prelu/gen/avx512f-2x32.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
|
|
"src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
|
|
"src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
|
|
"src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
|
|
"src/f32-relu/gen/avx512f-x16.c",
|
|
"src/f32-relu/gen/avx512f-x32.c",
|
|
"src/f32-rmax/avx512f.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
|
|
"src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vmin-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vmin-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vminc-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vminc-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
|
|
"src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
|
|
"src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
|
|
"src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
|
|
"src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
|
|
"src/f32-vrnd/gen/vrndd-avx512f-x16.c",
|
|
"src/f32-vrnd/gen/vrndd-avx512f-x32.c",
|
|
"src/f32-vrnd/gen/vrndne-avx512f-x16.c",
|
|
"src/f32-vrnd/gen/vrndne-avx512f-x32.c",
|
|
"src/f32-vrnd/gen/vrndu-avx512f-x16.c",
|
|
"src/f32-vrnd/gen/vrndu-avx512f-x32.c",
|
|
"src/f32-vrnd/gen/vrndz-avx512f-x16.c",
|
|
"src/f32-vrnd/gen/vrndz-avx512f-x32.c",
|
|
"src/f32-vscale/avx512f-x64.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
|
|
"src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
|
|
"src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
|
|
"src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
|
|
"src/f32-vunary/gen/vabs-avx512f-x16.c",
|
|
"src/f32-vunary/gen/vabs-avx512f-x32.c",
|
|
"src/f32-vunary/gen/vneg-avx512f-x16.c",
|
|
"src/f32-vunary/gen/vneg-avx512f-x32.c",
|
|
"src/f32-vunary/gen/vsqr-avx512f-x16.c",
|
|
"src/f32-vunary/gen/vsqr-avx512f-x32.c",
|
|
"src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
|
|
"src/math/exp-avx512f-rr2-lut16-p3-perm.c",
|
|
"src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
|
|
"src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
|
|
"src/math/exp-avx512f-rr2-p5-scalef.c",
|
|
"src/math/exp-avx512f-rr2-p5.c",
|
|
"src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
|
|
"src/math/expm1minus-avx512f-rr1-p6.c",
|
|
"src/math/extexp-avx512f-p5.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
|
|
"src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
|
|
"src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
|
|
"src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
|
|
"src/math/sqrt-avx512f-nr1fma.c",
|
|
"src/math/sqrt-avx512f-nr1fma1adj.c",
|
|
"src/math/sqrt-avx512f-nr2fma.c",
|
|
]
|
|
|
|
AVX512SKX_UKERNELS = [
|
|
"src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
|
|
"src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
|
|
"src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
|
|
"src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
|
|
"src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
|
|
"src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
|
|
"src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
|
|
"src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
|
|
"src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
|
|
"src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
|
|
]
|
|
|
|
WASM32_ASM_UKERNELS = [
|
|
"src/f32-relu/wasm_shr_x1.S",
|
|
"src/f32-relu/wasm_shr_x2.S",
|
|
"src/f32-relu/wasm_shr_x4.S",
|
|
]
|
|
|
|
AARCH32_ASM_UKERNELS = [
|
|
"src/f32-gemm/4x4-aarch32-vfp-ld64.S",
|
|
"src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
|
|
"src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
|
|
"src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
|
|
"src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
|
|
"src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
|
|
]
|
|
|
|
AARCH64_ASM_UKERNELS = [
|
|
"src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
|
|
"src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
|
|
"src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
|
|
"src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
|
|
"src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
|
|
"src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
|
|
"src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
|
|
"src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
|
|
"src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
|
|
"src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
|
|
"src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
|
|
"src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
|
|
"src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
|
|
"src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
|
|
"src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
|
|
"src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
|
|
"src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
|
|
"src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
|
|
"src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
|
|
"src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
|
|
"src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
|
|
"src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
|
|
"src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
|
|
"src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
|
|
]
|
|
|
|
INTERNAL_MICROKERNEL_HDRS = [
|
|
"src/xnnpack/argmaxpool.h",
|
|
"src/xnnpack/avgpool.h",
|
|
"src/xnnpack/clamp.h",
|
|
"src/xnnpack/common.h",
|
|
"src/xnnpack/conv.h",
|
|
"src/xnnpack/depthtospace.h",
|
|
"src/xnnpack/dwconv.h",
|
|
"src/xnnpack/fill.h",
|
|
"src/xnnpack/gavgpool.h",
|
|
"src/xnnpack/gemm.h",
|
|
"src/xnnpack/hswish.h",
|
|
"src/xnnpack/ibilinear.h",
|
|
"src/xnnpack/igemm.h",
|
|
"src/xnnpack/intrinsics-polyfill.h",
|
|
"src/xnnpack/lut.h",
|
|
"src/xnnpack/math.h",
|
|
"src/xnnpack/maxpool.h",
|
|
"src/xnnpack/memory.h",
|
|
"src/xnnpack/packx.h",
|
|
"src/xnnpack/pad.h",
|
|
"src/xnnpack/params.h",
|
|
"src/xnnpack/pavgpool.h",
|
|
"src/xnnpack/ppmm.h",
|
|
"src/xnnpack/prelu.h",
|
|
"src/xnnpack/raddexpminusmax.h",
|
|
"src/xnnpack/raddextexp.h",
|
|
"src/xnnpack/raddstoreexpminusmax.h",
|
|
"src/xnnpack/rmax.h",
|
|
"src/xnnpack/scalar-utils.h",
|
|
"src/xnnpack/spmm.h",
|
|
"src/xnnpack/unpool.h",
|
|
"src/xnnpack/vadd.h",
|
|
"src/xnnpack/vbinary.h",
|
|
"src/xnnpack/vmulcaddc.h",
|
|
"src/xnnpack/vscale.h",
|
|
"src/xnnpack/vscaleexpminusmax.h",
|
|
"src/xnnpack/vscaleextexp.h",
|
|
"src/xnnpack/vunary.h",
|
|
"src/xnnpack/zip.h",
|
|
]
|
|
|
|
INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
|
|
"include/xnnpack.h",
|
|
"src/xnnpack/allocator.h",
|
|
"src/xnnpack/compute.h",
|
|
"src/xnnpack/im2col.h",
|
|
"src/xnnpack/indirection.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
"src/xnnpack/memory-planner.h",
|
|
"src/xnnpack/operator.h",
|
|
"src/xnnpack/pack.h",
|
|
"src/xnnpack/params-init.h",
|
|
"src/xnnpack/requantization-stubs.h",
|
|
"src/xnnpack/requantization.h",
|
|
"src/xnnpack/subgraph.h",
|
|
]
|
|
|
|
ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
|
|
"src/xnnpack/math-stubs.h",
|
|
]
|
|
|
|
MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
|
|
"include/xnnpack.h",
|
|
"src/xnnpack/params-init.h",
|
|
]
|
|
|
|
MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
|
|
"include/xnnpack.h",
|
|
"src/xnnpack/isa-checks.h",
|
|
"src/xnnpack/params-init.h",
|
|
"src/xnnpack/requantization.h",
|
|
]
|
|
|
|
OPERATOR_TEST_PARAMS_HDRS = [
|
|
"src/xnnpack/common.h",
|
|
"src/xnnpack/params.h",
|
|
]
|
|
|
|
WEIGHTS_PACK_HDRS = [
|
|
"src/xnnpack/compute.h",
|
|
"src/xnnpack/operator.h",
|
|
"src/xnnpack/pack.h",
|
|
]
|
|
|
|
LOGGING_COPTS = select({
|
|
# No logging in optimized mode
|
|
":optimized_build": ["-DXNN_LOG_LEVEL=0"],
|
|
# Full logging in debug mode
|
|
":debug_build": ["-DXNN_LOG_LEVEL=5"],
|
|
# Error-only logging in default (fastbuild) mode
|
|
"//conditions:default": ["-DXNN_LOG_LEVEL=2"],
|
|
})
|
|
|
|
LOGGING_SRCS = select({
|
|
# No logging in optimized mode
|
|
":optimized_build": [],
|
|
"//conditions:default": [
|
|
"src/operator-strings.c",
|
|
"src/subgraph-strings.c",
|
|
],
|
|
})
|
|
|
|
LOGGING_HDRS = [
|
|
"src/xnnpack/log.h",
|
|
]
|
|
|
|
xnnpack_cc_library(
|
|
name = "tables",
|
|
srcs = TABLE_SRCS,
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "scalar_ukernels",
|
|
srcs = SCALAR_UKERNELS,
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = ["-marm"],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "scalar_ukernels_test_mode",
|
|
srcs = SCALAR_UKERNELS,
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = ["-marm"],
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "wasm_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
wasm_srcs = WASM_UKERNELS,
|
|
wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "wasm_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
wasm_srcs = WASM_UKERNELS,
|
|
wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neon_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv7-a",
|
|
"-mfpu=neon",
|
|
],
|
|
aarch32_srcs = NEON_UKERNELS,
|
|
aarch64_srcs = NEON_UKERNELS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neon_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv7-a",
|
|
"-mfpu=neon",
|
|
],
|
|
aarch32_srcs = NEON_UKERNELS,
|
|
aarch64_srcs = NEON_UKERNELS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonfma_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv7-a",
|
|
"-mfpu=neon-vfpv4",
|
|
],
|
|
aarch32_srcs = NEONFMA_UKERNELS,
|
|
aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
|
|
apple_aarch32_copts = [
|
|
"-mcpu=swift",
|
|
"-mtune=generic",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonfma_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv7-a",
|
|
"-mfpu=neon-vfpv4",
|
|
],
|
|
aarch32_srcs = NEONFMA_UKERNELS,
|
|
aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
|
|
apple_aarch32_copts = [
|
|
"-mcpu=swift",
|
|
"-mtune=generic",
|
|
],
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonv8_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv8-a",
|
|
"-mfpu=neon-fp-armv8",
|
|
],
|
|
aarch32_srcs = NEONV8_UKERNELS,
|
|
aarch64_srcs = NEONV8_UKERNELS,
|
|
apple_aarch32_copts = [
|
|
"-mcpu=cyclone",
|
|
"-mtune=generic",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonv8_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv8-a",
|
|
"-mfpu=neon-fp-armv8",
|
|
],
|
|
aarch32_srcs = NEONV8_UKERNELS,
|
|
aarch64_srcs = NEONV8_UKERNELS,
|
|
apple_aarch32_copts = [
|
|
"-mcpu=cyclone",
|
|
"-mtune=generic",
|
|
],
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonfp16arith_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch64_copts = ["-march=armv8.2-a+fp16"],
|
|
aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neonfp16arith_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch64_copts = ["-march=armv8.2-a+fp16"],
|
|
aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neondot_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv8.2-a+dotprod",
|
|
"-mfpu=neon-fp-armv8",
|
|
],
|
|
aarch32_srcs = NEONDOT_UKERNELS,
|
|
aarch64_copts = ["-march=armv8.2-a+dotprod"],
|
|
aarch64_srcs = NEONDOT_UKERNELS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "neondot_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
aarch32_copts = [
|
|
"-marm",
|
|
"-march=armv8.2-a+dotprod",
|
|
"-mfpu=neon-fp-armv8",
|
|
],
|
|
aarch32_srcs = NEONDOT_UKERNELS,
|
|
aarch64_copts = ["-march=armv8.2-a+dotprod"],
|
|
aarch64_srcs = NEONDOT_UKERNELS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "sse2_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-msse2"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "sse2_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-msse2"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "ssse3_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mssse3"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSSE3_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "ssse3_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mssse3"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSSE3_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "sse41_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-msse4.1"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSE41_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "sse41_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-msse4.1"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:SSE2"],
|
|
x86_srcs = SSE41_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mavx"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = AVX_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mavx"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = AVX_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "xop_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mxop"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = XOP_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "xop_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mxop"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = XOP_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "fma3_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mfma"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = FMA3_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "fma3_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mfma"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX"],
|
|
msvc_x86_64_copts = ["/arch:AVX"],
|
|
x86_srcs = FMA3_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx2_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = [
|
|
"-mfma",
|
|
"-mavx2",
|
|
],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX2"],
|
|
msvc_x86_64_copts = ["/arch:AVX2"],
|
|
x86_srcs = AVX2_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx2_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = [
|
|
"-mfma",
|
|
"-mavx2",
|
|
],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX2"],
|
|
msvc_x86_64_copts = ["/arch:AVX2"],
|
|
x86_srcs = AVX2_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx512f_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mavx512f"],
|
|
mingw_copts = ["-fno-asynchronous-unwind-tables"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX512"],
|
|
msvc_x86_64_copts = ["/arch:AVX512"],
|
|
msys_copts = ["-fno-asynchronous-unwind-tables"],
|
|
x86_srcs = AVX512F_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx512f_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = ["-mavx512f"],
|
|
mingw_copts = ["-fno-asynchronous-unwind-tables"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX512"],
|
|
msvc_x86_64_copts = ["/arch:AVX512"],
|
|
msys_copts = ["-fno-asynchronous-unwind-tables"],
|
|
x86_srcs = AVX512F_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx512skx_ukernels",
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = [
|
|
"-mavx512f",
|
|
"-mavx512cd",
|
|
"-mavx512bw",
|
|
"-mavx512dq",
|
|
"-mavx512vl",
|
|
],
|
|
mingw_copts = ["-fno-asynchronous-unwind-tables"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX512"],
|
|
msvc_x86_64_copts = ["/arch:AVX512"],
|
|
msys_copts = ["-fno-asynchronous-unwind-tables"],
|
|
x86_srcs = AVX512SKX_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "avx512skx_ukernels_test_mode",
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
gcc_x86_copts = [
|
|
"-mavx512f",
|
|
"-mavx512cd",
|
|
"-mavx512bw",
|
|
"-mavx512dq",
|
|
"-mavx512vl",
|
|
],
|
|
mingw_copts = ["-fno-asynchronous-unwind-tables"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
msvc_x86_32_copts = ["/arch:AVX512"],
|
|
msvc_x86_64_copts = ["/arch:AVX512"],
|
|
msys_copts = ["-fno-asynchronous-unwind-tables"],
|
|
x86_srcs = AVX512SKX_UKERNELS,
|
|
deps = [
|
|
":tables",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "asm_ukernels",
|
|
hdrs = ["src/xnnpack/assembly.h"],
|
|
aarch32_srcs = AARCH32_ASM_UKERNELS,
|
|
aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
|
|
aarch64_srcs = AARCH64_ASM_UKERNELS,
|
|
wasm_srcs = WASM32_ASM_UKERNELS,
|
|
wasmsimd_srcs = WASM32_ASM_UKERNELS,
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "logging_utils",
|
|
srcs = LOGGING_SRCS,
|
|
hdrs = INTERNAL_HDRS + LOGGING_HDRS,
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
visibility = xnnpack_visibility(),
|
|
deps = [
|
|
"@FP16",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_aggregate_library(
|
|
name = "ukernels",
|
|
aarch32_deps = [
|
|
":neon_ukernels",
|
|
":neonfma_ukernels",
|
|
":neonv8_ukernels",
|
|
":neondot_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
aarch64_deps = [
|
|
":neon_ukernels",
|
|
":neonfma_ukernels",
|
|
":neonv8_ukernels",
|
|
":neonfp16arith_ukernels",
|
|
":neondot_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
generic_deps = [
|
|
":scalar_ukernels",
|
|
],
|
|
wasm_deps = [
|
|
":wasm_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
wasmsimd_deps = [
|
|
":wasm_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
x86_deps = [
|
|
":sse2_ukernels",
|
|
":ssse3_ukernels",
|
|
":sse41_ukernels",
|
|
":avx_ukernels",
|
|
":xop_ukernels",
|
|
":fma3_ukernels",
|
|
":avx2_ukernels",
|
|
":avx512f_ukernels",
|
|
":avx512skx_ukernels",
|
|
],
|
|
)
|
|
|
|
xnnpack_aggregate_library(
|
|
name = "ukernels_test_mode",
|
|
aarch32_deps = [
|
|
":neon_ukernels_test_mode",
|
|
":neonfma_ukernels_test_mode",
|
|
":neonv8_ukernels_test_mode",
|
|
":neondot_ukernels_test_mode",
|
|
":asm_ukernels",
|
|
],
|
|
aarch64_deps = [
|
|
":neon_ukernels_test_mode",
|
|
":neonfma_ukernels_test_mode",
|
|
":neonv8_ukernels_test_mode",
|
|
":neonfp16arith_ukernels_test_mode",
|
|
":neondot_ukernels_test_mode",
|
|
":asm_ukernels",
|
|
],
|
|
generic_deps = [
|
|
":scalar_ukernels_test_mode",
|
|
],
|
|
wasm_deps = [
|
|
":wasm_ukernels_test_mode",
|
|
":asm_ukernels",
|
|
],
|
|
wasmsimd_deps = [
|
|
":wasm_ukernels_test_mode",
|
|
":asm_ukernels",
|
|
],
|
|
x86_deps = [
|
|
":sse2_ukernels_test_mode",
|
|
":ssse3_ukernels_test_mode",
|
|
":sse41_ukernels_test_mode",
|
|
":avx_ukernels_test_mode",
|
|
":xop_ukernels_test_mode",
|
|
":fma3_ukernels_test_mode",
|
|
":avx2_ukernels_test_mode",
|
|
":avx512f_ukernels_test_mode",
|
|
":avx512skx_ukernels_test_mode",
|
|
],
|
|
)
|
|
|
|
# This target must be compatible with Android NDK r17 and compile with Clang 6.
|
|
# It is similar to :ukernels target, but excludes microkernels which require
|
|
# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
|
|
xnnpack_aggregate_library(
|
|
name = "ukernels_max_compatibility",
|
|
aarch32_deps = [
|
|
":neon_ukernels",
|
|
":neonfma_ukernels",
|
|
":neonv8_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
aarch64_deps = [
|
|
":neon_ukernels",
|
|
":neonfma_ukernels",
|
|
":neonv8_ukernels",
|
|
":asm_ukernels",
|
|
],
|
|
generic_deps = [
|
|
":scalar_ukernels",
|
|
],
|
|
wasm_deps = [
|
|
":wasm_ukernels",
|
|
],
|
|
wasmsimd_deps = [
|
|
":wasm_ukernels",
|
|
],
|
|
x86_deps = [
|
|
":sse2_ukernels",
|
|
":ssse3_ukernels",
|
|
":sse41_ukernels",
|
|
":avx_ukernels",
|
|
":xop_ukernels",
|
|
":fma3_ukernels",
|
|
":avx2_ukernels",
|
|
":avx512f_ukernels",
|
|
":avx512skx_ukernels",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "im2col",
|
|
srcs = ["src/im2col.c"],
|
|
hdrs = [
|
|
"src/xnnpack/common.h",
|
|
"src/xnnpack/im2col.h",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "indirection",
|
|
srcs = ["src/indirection.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "indirection_test_mode",
|
|
srcs = ["src/indirection.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "packing",
|
|
srcs = ["src/packing.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "packing_test_mode",
|
|
srcs = ["src/packing.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "operator_run",
|
|
srcs = ["src/operator-run.c"],
|
|
hdrs = INTERNAL_HDRS + LOGGING_HDRS,
|
|
copts = LOGGING_COPTS + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":logging_utils",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "operator_run_test_mode",
|
|
srcs = ["src/operator-run.c"],
|
|
hdrs = INTERNAL_HDRS + LOGGING_HDRS,
|
|
copts = LOGGING_COPTS + [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
] + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":logging_utils",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "memory_planner",
|
|
srcs = ["src/memory-planner.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
defines = select({
|
|
":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
|
|
":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
|
|
"//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":logging_utils",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "memory_planner_test_mode",
|
|
srcs = ["src/memory-planner.c"],
|
|
hdrs = INTERNAL_HDRS,
|
|
copts = [
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
],
|
|
defines = select({
|
|
":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
|
|
":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
|
|
"//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":logging_utils",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "enable_assembly",
|
|
defines = select({
|
|
":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
|
|
":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
|
|
"//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
|
|
}),
|
|
)
|
|
|
|
cc_library(
|
|
name = "enable_sparse",
|
|
defines = select({
|
|
":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
|
|
":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
|
|
"//conditions:default": ["XNN_ENABLE_SPARSE=1"],
|
|
}),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "operators",
|
|
srcs = OPERATOR_SRCS + [
|
|
"src/memory.c",
|
|
"src/operator-delete.c",
|
|
],
|
|
hdrs = INTERNAL_HDRS + LOGGING_HDRS,
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":indirection",
|
|
":logging_utils",
|
|
":packing",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "operators_test_mode",
|
|
srcs = OPERATOR_SRCS + [
|
|
"src/memory.c",
|
|
"src/operator-delete.c",
|
|
],
|
|
hdrs = INTERNAL_HDRS + LOGGING_HDRS,
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
deps = [
|
|
":indirection_test_mode",
|
|
":logging_utils",
|
|
":packing_test_mode",
|
|
"@FP16",
|
|
"@FXdiv",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "XNNPACK",
|
|
srcs = [
|
|
"src/init.c",
|
|
"src/runtime.c",
|
|
"src/subgraph.c",
|
|
"src/tensor.c",
|
|
] + SUBGRAPH_SRCS,
|
|
hdrs = ["include/xnnpack.h"],
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
includes = ["include"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
visibility = xnnpack_visibility(),
|
|
deps = [
|
|
":enable_assembly",
|
|
":enable_sparse",
|
|
":logging_utils",
|
|
":memory_planner",
|
|
":operator_run",
|
|
":operators",
|
|
":ukernels",
|
|
"@clog",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
] + select({
|
|
":emscripten": [],
|
|
"//conditions:default": ["@cpuinfo"],
|
|
}),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "XNNPACK_test_mode",
|
|
srcs = [
|
|
"src/init.c",
|
|
"src/runtime.c",
|
|
"src/subgraph.c",
|
|
"src/tensor.c",
|
|
] + SUBGRAPH_SRCS,
|
|
hdrs = ["include/xnnpack.h"],
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
"-UNDEBUG",
|
|
"-DXNN_TEST_MODE=1",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
includes = ["include"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
visibility = xnnpack_visibility(),
|
|
deps = [
|
|
":enable_assembly",
|
|
":enable_sparse",
|
|
":logging_utils",
|
|
":memory_planner_test_mode",
|
|
":operator_run_test_mode",
|
|
":operators_test_mode",
|
|
":ukernels_test_mode",
|
|
"@clog",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
] + select({
|
|
":emscripten": [],
|
|
"//conditions:default": ["@cpuinfo"],
|
|
}),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "xnnpack_f32",
|
|
srcs = [
|
|
"src/init.c",
|
|
"src/runtime.c",
|
|
"src/subgraph.c",
|
|
"src/tensor.c",
|
|
] + SUBGRAPH_SRCS,
|
|
hdrs = ["include/xnnpack.h"],
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
defines = [
|
|
"XNN_NO_QS8_OPERATORS",
|
|
"XNN_NO_QU8_OPERATORS",
|
|
"XNN_NO_U8_OPERATORS",
|
|
"XNN_NO_X8_OPERATORS",
|
|
"XNN_NO_F16_OPERATORS",
|
|
"XNN_NO_X16_OPERATORS",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
includes = ["include"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
visibility = xnnpack_visibility(),
|
|
deps = [
|
|
":enable_assembly",
|
|
":enable_sparse",
|
|
":logging_utils",
|
|
":memory_planner",
|
|
":operator_run",
|
|
":operators",
|
|
":ukernels_max_compatibility",
|
|
"@clog",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
] + select({
|
|
":emscripten": [],
|
|
"//conditions:default": ["@cpuinfo"],
|
|
}),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "xnnpack_operators_nhwc_f32",
|
|
srcs = [
|
|
"src/init.c",
|
|
],
|
|
hdrs = ["include/xnnpack.h"],
|
|
copts = LOGGING_COPTS + [
|
|
"-Isrc",
|
|
"-Iinclude",
|
|
] + select({
|
|
":debug_build": [],
|
|
"//conditions:default": xnnpack_min_size_copts(),
|
|
}) + select({
|
|
":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
|
|
"//conditions:default": [],
|
|
}),
|
|
defines = [
|
|
"XNN_NO_QS8_OPERATORS",
|
|
"XNN_NO_QU8_OPERATORS",
|
|
"XNN_NO_U8_OPERATORS",
|
|
"XNN_NO_X8_OPERATORS",
|
|
"XNN_NO_NCHW_OPERATORS",
|
|
],
|
|
gcc_copts = xnnpack_gcc_std_copts(),
|
|
includes = ["include"],
|
|
msvc_copts = xnnpack_msvc_std_copts(),
|
|
visibility = xnnpack_visibility(),
|
|
deps = [
|
|
":enable_assembly",
|
|
":logging_utils",
|
|
":operator_run",
|
|
":operators",
|
|
":ukernels",
|
|
"@clog",
|
|
"@pthreadpool",
|
|
] + select({
|
|
":emscripten": [],
|
|
"//conditions:default": ["@cpuinfo"],
|
|
}),
|
|
)
|
|
|
|
xnnpack_cc_library(
|
|
name = "bench_utils",
|
|
srcs = ["bench/utils.cc"],
|
|
hdrs = ["bench/utils.h"],
|
|
deps = [
|
|
"@com_google_benchmark//:benchmark",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
######################### Benchmarks for micro-kernels #########################
|
|
|
|
xnnpack_benchmark(
|
|
name = "qs8_gemm_bench",
|
|
srcs = [
|
|
"bench/gemm.h",
|
|
"bench/qs8-gemm.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "qs8_requantization_bench",
|
|
srcs = [
|
|
"bench/qs8-requantization.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/requantization-stubs.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "qu8_gemm_bench",
|
|
srcs = [
|
|
"bench/gemm.h",
|
|
"bench/qu8-gemm.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "qu8_requantization_bench",
|
|
srcs = [
|
|
"bench/qu8-requantization.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/requantization-stubs.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f16_igemm_bench",
|
|
srcs = [
|
|
"bench/f16-igemm.cc",
|
|
"bench/conv.h",
|
|
"bench/google/conv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":indirection",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f16_gemm_bench",
|
|
srcs = [
|
|
"bench/f16-gemm.cc",
|
|
"bench/gemm.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f16_spmm_bench",
|
|
srcs = [
|
|
"bench/f16-spmm.cc",
|
|
"bench/spmm.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_igemm_bench",
|
|
srcs = [
|
|
"bench/f32-igemm.cc",
|
|
"bench/conv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":indirection",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f16_relu_bench",
|
|
srcs = [
|
|
"bench/f16-relu.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_conv_hwc_bench",
|
|
srcs = [
|
|
"bench/f32-conv-hwc.cc",
|
|
"bench/dconv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_conv_hwc2chw_bench",
|
|
srcs = [
|
|
"bench/f32-conv-hwc2chw.cc",
|
|
"bench/dconv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f16_dwconv_bench",
|
|
srcs = [
|
|
"bench/f16-dwconv.cc",
|
|
"bench/dwconv.h",
|
|
"bench/google/dwconv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":indirection",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_dwconv_bench",
|
|
srcs = [
|
|
"bench/f32-dwconv.cc",
|
|
"bench/dwconv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":indirection",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_dwconv2d_chw_bench",
|
|
srcs = [
|
|
"bench/f32-dwconv2d-chw.cc",
|
|
"bench/dwconv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":indirection",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_gemm_bench",
|
|
srcs = [
|
|
"bench/f32-gemm.cc",
|
|
"bench/gemm.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
copts = xnnpack_optional_ruy_copts(),
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_hswish_bench",
|
|
srcs = [
|
|
"bench/f32-hswish.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_raddexpminusmax_bench",
|
|
srcs = [
|
|
"bench/f32-raddexpminusmax.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_raddextexp_bench",
|
|
srcs = [
|
|
"bench/f32-raddextexp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_raddstoreexpminusmax_bench",
|
|
srcs = [
|
|
"bench/f32-raddstoreexpminusmax.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_relu_bench",
|
|
srcs = [
|
|
"bench/f32-relu.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_rmax_bench",
|
|
srcs = [
|
|
"bench/f32-rmax.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_sigmoid_bench",
|
|
srcs = [
|
|
"bench/f32-sigmoid.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_spmm_bench",
|
|
srcs = [
|
|
"bench/f32-spmm.cc",
|
|
"bench/spmm.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_softmax_bench",
|
|
srcs = [
|
|
"bench/f32-softmax.cc",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
copts = xnnpack_optional_dnnl_copts(),
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_velu_bench",
|
|
srcs = [
|
|
"bench/f32-velu.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_vscaleexpminusmax_bench",
|
|
srcs = [
|
|
"bench/f32-vscaleexpminusmax.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_vscaleextexp_bench",
|
|
srcs = [
|
|
"bench/f32-vscaleextexp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_vsqrt_bench",
|
|
srcs = [
|
|
"bench/f32-vsqrt.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_im2col_gemm_bench",
|
|
srcs = [
|
|
"bench/f32-im2col-gemm.cc",
|
|
"bench/conv.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":im2col",
|
|
":packing",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "rounding_bench",
|
|
srcs = [
|
|
"bench/rounding.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS,
|
|
)
|
|
|
|
########################### Benchmarks for operators ###########################
|
|
|
|
xnnpack_benchmark(
|
|
name = "average_pooling_bench",
|
|
srcs = ["bench/average-pooling.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "bankers_rounding_bench",
|
|
srcs = ["bench/bankers-rounding.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "ceiling_bench",
|
|
srcs = ["bench/ceiling.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "channel_shuffle_bench",
|
|
srcs = ["bench/channel-shuffle.cc"],
|
|
deps = OPERATOR_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "convolution_bench",
|
|
srcs = ["bench/convolution.cc"],
|
|
copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "deconvolution_bench",
|
|
srcs = ["bench/deconvolution.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "elu_bench",
|
|
srcs = ["bench/elu.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "floor_bench",
|
|
srcs = ["bench/floor.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "global_average_pooling_bench",
|
|
srcs = ["bench/global-average-pooling.cc"],
|
|
deps = OPERATOR_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "hardswish_bench",
|
|
srcs = ["bench/hardswish.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "max_pooling_bench",
|
|
srcs = ["bench/max-pooling.cc"],
|
|
deps = OPERATOR_BENCHMARK_DEPS,
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "sigmoid_bench",
|
|
srcs = ["bench/sigmoid.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "prelu_bench",
|
|
srcs = ["bench/prelu.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "softmax_bench",
|
|
srcs = ["bench/softmax.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "square_root_bench",
|
|
srcs = ["bench/square-root.cc"],
|
|
copts = xnnpack_optional_tflite_copts(),
|
|
tags = ["nowin32"],
|
|
deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "truncation_bench",
|
|
srcs = ["bench/truncation.cc"],
|
|
deps = OPERATOR_BENCHMARK_DEPS,
|
|
)
|
|
|
|
############################# End-to-end benchmarks ############################
|
|
|
|
cc_library(
|
|
name = "fp32_mobilenet_v1",
|
|
srcs = ["models/fp32-mobilenet-v1.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_sparse_mobilenet_v1",
|
|
srcs = ["models/fp32-sparse-mobilenet-v1.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp16_mobilenet_v1",
|
|
srcs = ["models/fp16-mobilenet-v1.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "qs8_mobilenet_v1",
|
|
srcs = ["models/qs8-mobilenet-v1.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "qs8_mobilenet_v2",
|
|
srcs = ["models/qs8-mobilenet-v2.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "qu8_mobilenet_v1",
|
|
srcs = ["models/qu8-mobilenet-v1.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_mobilenet_v2",
|
|
srcs = ["models/fp32-mobilenet-v2.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_sparse_mobilenet_v2",
|
|
srcs = ["models/fp32-sparse-mobilenet-v2.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp16_mobilenet_v2",
|
|
srcs = ["models/fp16-mobilenet-v2.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_mobilenet_v3_large",
|
|
srcs = ["models/fp32-mobilenet-v3-large.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_sparse_mobilenet_v3_large",
|
|
srcs = ["models/fp32-sparse-mobilenet-v3-large.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp16_mobilenet_v3_large",
|
|
srcs = ["models/fp16-mobilenet-v3-large.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_mobilenet_v3_small",
|
|
srcs = ["models/fp32-mobilenet-v3-small.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp32_sparse_mobilenet_v3_small",
|
|
srcs = ["models/fp32-sparse-mobilenet-v3-small.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
cc_library(
|
|
name = "fp16_mobilenet_v3_small",
|
|
srcs = ["models/fp16-mobilenet-v3-small.cc"],
|
|
hdrs = ["models/models.h"],
|
|
copts = xnnpack_std_cxxopts(),
|
|
linkstatic = True,
|
|
deps = [
|
|
":XNNPACK",
|
|
"@FP16",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_dwconv_e2e_bench",
|
|
srcs = [
|
|
"bench/f32-dwconv-e2e.cc",
|
|
"bench/end2end.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":XNNPACK",
|
|
":fp32_mobilenet_v1",
|
|
":fp32_mobilenet_v2",
|
|
":fp32_mobilenet_v3_large",
|
|
":fp32_mobilenet_v3_small",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_gemm_e2e_bench",
|
|
srcs = [
|
|
"bench/f32-gemm-e2e.cc",
|
|
"bench/end2end.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":XNNPACK",
|
|
":fp32_mobilenet_v1",
|
|
":fp32_mobilenet_v2",
|
|
":fp32_mobilenet_v3_large",
|
|
":fp32_mobilenet_v3_small",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "qs8_gemm_e2e_bench",
|
|
srcs = [
|
|
"bench/qs8-gemm-e2e.cc",
|
|
"bench/end2end.h",
|
|
] + MICROKERNEL_BENCHMARK_HDRS,
|
|
deps = MICROKERNEL_BENCHMARK_DEPS + [
|
|
":XNNPACK",
|
|
":qs8_mobilenet_v1",
|
|
":qs8_mobilenet_v2",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "end2end_bench",
|
|
srcs = ["bench/end2end.cc"],
|
|
deps = [
|
|
":XNNPACK",
|
|
":bench_utils",
|
|
":fp16_mobilenet_v1",
|
|
":fp16_mobilenet_v2",
|
|
":fp16_mobilenet_v3_large",
|
|
":fp16_mobilenet_v3_small",
|
|
":fp32_mobilenet_v1",
|
|
":fp32_mobilenet_v2",
|
|
":fp32_mobilenet_v3_large",
|
|
":fp32_mobilenet_v3_small",
|
|
":fp32_sparse_mobilenet_v1",
|
|
":fp32_sparse_mobilenet_v2",
|
|
":fp32_sparse_mobilenet_v3_large",
|
|
":fp32_sparse_mobilenet_v3_small",
|
|
":qs8_mobilenet_v1",
|
|
":qs8_mobilenet_v2",
|
|
":qu8_mobilenet_v1",
|
|
"@pthreadpool",
|
|
],
|
|
)
|
|
|
|
#################### Accuracy evaluation for math functions ####################
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_exp_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-exp-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_expminus_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-expminus-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_expm1minus_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-expm1minus-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_extexp_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-extexp-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_sigmoid_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-sigmoid-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
xnnpack_benchmark(
|
|
name = "f32_sqrt_ulp_eval",
|
|
srcs = [
|
|
"eval/f32-sqrt-ulp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + ACCURACY_EVAL_HDRS,
|
|
deps = ACCURACY_EVAL_DEPS + [
|
|
":bench_utils",
|
|
"@cpuinfo",
|
|
],
|
|
)
|
|
|
|
################### Accuracy verification for math functions ##################
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_exp_eval",
|
|
srcs = [
|
|
"eval/f32-exp.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_expm1minus_eval",
|
|
srcs = [
|
|
"eval/f32-expm1minus.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_expminus_eval",
|
|
srcs = [
|
|
"eval/f32-expminus.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_roundne_eval",
|
|
srcs = [
|
|
"eval/f32-roundne.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_roundd_eval",
|
|
srcs = [
|
|
"eval/f32-roundd.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_roundu_eval",
|
|
srcs = [
|
|
"eval/f32-roundu.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_roundz_eval",
|
|
srcs = [
|
|
"eval/f32-roundz.cc",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
"src/xnnpack/math-stubs.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
automatic = False,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
######################### Unit tests for micro-kernels #########################
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_gemm_minmax_test",
|
|
srcs = [
|
|
"test/f16-gemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_spmm_minmax_test",
|
|
srcs = [
|
|
"test/f16-spmm-minmax.cc",
|
|
"test/spmm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_argmaxpool_test",
|
|
srcs = [
|
|
"test/f32-argmaxpool.cc",
|
|
"test/argmaxpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_avgpool_minmax_test",
|
|
srcs = [
|
|
"test/f32-avgpool-minmax.cc",
|
|
"test/avgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_ibilinear_test",
|
|
srcs = [
|
|
"test/f32-ibilinear.cc",
|
|
"test/ibilinear-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_ibilinear_chw_test",
|
|
srcs = [
|
|
"test/f32-ibilinear-chw.cc",
|
|
"test/ibilinear-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_clamp_test",
|
|
srcs = [
|
|
"test/f16-clamp.cc",
|
|
"test/clamp-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_clamp_test",
|
|
srcs = [
|
|
"test/f32-clamp.cc",
|
|
"test/clamp-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_igemm_test",
|
|
srcs = [
|
|
"test/f32-igemm.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_igemm_relu_test",
|
|
srcs = [
|
|
"test/f32-igemm-relu.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_igemm_minmax_test",
|
|
srcs = [
|
|
"test/f32-igemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_igemm_minmax_test",
|
|
srcs = [
|
|
"test/f16-igemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_conv_hwc_test",
|
|
srcs = [
|
|
"test/f32-conv-hwc.cc",
|
|
"test/conv-hwc-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_conv_hwc2chw_test",
|
|
srcs = [
|
|
"test/f32-conv-hwc2chw.cc",
|
|
"test/conv-hwc2chw-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_dwconv_test",
|
|
srcs = [
|
|
"test/f32-dwconv.cc",
|
|
"test/dwconv-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_dwconv_minmax_test",
|
|
srcs = [
|
|
"test/f16-dwconv-minmax.cc",
|
|
"test/dwconv-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_dwconv_minmax_test",
|
|
srcs = [
|
|
"test/f32-dwconv-minmax.cc",
|
|
"test/dwconv-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_dwconv2d_chw_test",
|
|
srcs = [
|
|
"test/f32-dwconv2d-chw.cc",
|
|
"test/dwconv2d-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_gavgpool_minmax_test",
|
|
srcs = [
|
|
"test/f16-gavgpool-minmax.cc",
|
|
"test/gavgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gavgpool_minmax_test",
|
|
srcs = [
|
|
"test/f32-gavgpool-minmax.cc",
|
|
"test/gavgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gavgpool_cw_test",
|
|
srcs = [
|
|
"test/f32-gavgpool-cw.cc",
|
|
"test/gavgpool-cw-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gemm_test",
|
|
srcs = [
|
|
"test/f32-gemm.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gemm_relu_test",
|
|
srcs = [
|
|
"test/f32-gemm-relu.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gemm_minmax_test",
|
|
srcs = [
|
|
"test/f32-gemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_gemminc_minmax_test",
|
|
srcs = [
|
|
"test/f32-gemminc-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_hswish_test",
|
|
srcs = [
|
|
"test/f16-hswish.cc",
|
|
"test/hswish-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_hswish_test",
|
|
srcs = [
|
|
"test/f32-hswish.cc",
|
|
"test/hswish-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_maxpool_minmax_test",
|
|
srcs = [
|
|
"test/f32-maxpool-minmax.cc",
|
|
"test/maxpool-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_pavgpool_minmax_test",
|
|
srcs = [
|
|
"test/f32-pavgpool-minmax.cc",
|
|
"test/avgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_ppmm_minmax_test",
|
|
srcs = [
|
|
"test/f32-ppmm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_prelu_test",
|
|
srcs = [
|
|
"test/f16-prelu.cc",
|
|
"test/prelu-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_prelu_test",
|
|
srcs = [
|
|
"test/f32-prelu.cc",
|
|
"test/prelu-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_raddexpminusmax_test",
|
|
srcs = [
|
|
"test/f32-raddexpminusmax.cc",
|
|
"test/raddexpminusmax-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_raddextexp_test",
|
|
srcs = [
|
|
"test/f32-raddextexp.cc",
|
|
"test/raddextexp-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_raddstoreexpminusmax_test",
|
|
srcs = [
|
|
"test/f32-raddstoreexpminusmax.cc",
|
|
"test/raddstoreexpminusmax-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_relu_test",
|
|
srcs = [
|
|
"test/f32-relu.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_rmax_test",
|
|
srcs = [
|
|
"test/f32-rmax.cc",
|
|
"test/rmax-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_sigmoid_test",
|
|
srcs = [
|
|
"test/f32-sigmoid.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_spmm_minmax_test",
|
|
srcs = [
|
|
"test/f32-spmm-minmax.cc",
|
|
"test/spmm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vadd_minmax_test",
|
|
srcs = [
|
|
"test/f16-vadd-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vaddc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vaddc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vdiv_minmax_test",
|
|
srcs = [
|
|
"test/f16-vdiv-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vdivc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vdivc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vrdivc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vrdivc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmax_test",
|
|
srcs = [
|
|
"test/f16-vmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmaxc_test",
|
|
srcs = [
|
|
"test/f16-vmaxc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmin_test",
|
|
srcs = [
|
|
"test/f16-vmin.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vminc_test",
|
|
srcs = [
|
|
"test/f16-vminc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmul_minmax_test",
|
|
srcs = [
|
|
"test/f16-vmul-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmulc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vmulc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vsub_minmax_test",
|
|
srcs = [
|
|
"test/f16-vsub-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vsubc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vsubc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vrsubc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vrsubc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vabs_test",
|
|
srcs = [
|
|
"test/f32-vabs.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vadd_test",
|
|
srcs = [
|
|
"test/f32-vadd.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vadd_minmax_test",
|
|
srcs = [
|
|
"test/f32-vadd-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vadd_relu_test",
|
|
srcs = [
|
|
"test/f32-vadd-relu.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vaddc_test",
|
|
srcs = [
|
|
"test/f32-vaddc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vaddc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vaddc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vaddc_relu_test",
|
|
srcs = [
|
|
"test/f32-vaddc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdiv_test",
|
|
srcs = [
|
|
"test/f32-vdiv.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdiv_minmax_test",
|
|
srcs = [
|
|
"test/f32-vdiv-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdiv_relu_test",
|
|
srcs = [
|
|
"test/f32-vdiv-relu.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdivc_test",
|
|
srcs = [
|
|
"test/f32-vdivc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdivc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vdivc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vdivc_relu_test",
|
|
srcs = [
|
|
"test/f32-vdivc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrdivc_test",
|
|
srcs = [
|
|
"test/f32-vrdivc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrdivc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vrdivc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrdivc_relu_test",
|
|
srcs = [
|
|
"test/f32-vrdivc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_velu_test",
|
|
srcs = [
|
|
"test/f32-velu.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmax_test",
|
|
srcs = [
|
|
"test/f32-vmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmaxc_test",
|
|
srcs = [
|
|
"test/f32-vmaxc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmin_test",
|
|
srcs = [
|
|
"test/f32-vmin.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vminc_test",
|
|
srcs = [
|
|
"test/f32-vminc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmul_test",
|
|
srcs = [
|
|
"test/f32-vmul.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmul_minmax_test",
|
|
srcs = [
|
|
"test/f32-vmul-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmul_relu_test",
|
|
srcs = [
|
|
"test/f32-vmul-relu.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmulc_test",
|
|
srcs = [
|
|
"test/f32-vmulc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmulc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vmulc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmulc_relu_test",
|
|
srcs = [
|
|
"test/f32-vmulc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f16_vmulcaddc_minmax_test",
|
|
srcs = [
|
|
"test/f16-vmulcaddc-minmax.cc",
|
|
"test/vmulcaddc-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vmulcaddc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vmulcaddc-minmax.cc",
|
|
"test/vmulcaddc-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vlrelu_test",
|
|
srcs = [
|
|
"test/f32-vlrelu.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vneg_test",
|
|
srcs = [
|
|
"test/f32-vneg.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrndne_test",
|
|
srcs = [
|
|
"test/f32-vrndne.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrndz_test",
|
|
srcs = [
|
|
"test/f32-vrndz.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrndu_test",
|
|
srcs = [
|
|
"test/f32-vrndu.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrndd_test",
|
|
srcs = [
|
|
"test/f32-vrndd.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vscale_test",
|
|
srcs = [
|
|
"test/f32-vscale.cc",
|
|
"test/vscale-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vscaleexpminusmax_test",
|
|
srcs = [
|
|
"test/f32-vscaleexpminusmax.cc",
|
|
"test/vscaleexpminusmax-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vscaleextexp_test",
|
|
srcs = [
|
|
"test/f32-vscaleextexp.cc",
|
|
"test/vscaleextexp-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsqr_test",
|
|
srcs = [
|
|
"test/f32-vsqr.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsqrdiff_test",
|
|
srcs = [
|
|
"test/f32-vsqrdiff.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsqrdiffc_test",
|
|
srcs = [
|
|
"test/f32-vsqrdiffc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsqrt_test",
|
|
srcs = [
|
|
"test/f32-vsqrt.cc",
|
|
"test/vunary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsub_test",
|
|
srcs = [
|
|
"test/f32-vsub.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsub_minmax_test",
|
|
srcs = [
|
|
"test/f32-vsub-minmax.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsub_relu_test",
|
|
srcs = [
|
|
"test/f32-vsub-relu.cc",
|
|
"test/vbinary-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsubc_test",
|
|
srcs = [
|
|
"test/f32-vsubc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsubc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vsubc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vsubc_relu_test",
|
|
srcs = [
|
|
"test/f32-vsubc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrsubc_test",
|
|
srcs = [
|
|
"test/f32-vrsubc.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrsubc_minmax_test",
|
|
srcs = [
|
|
"test/f32-vrsubc-minmax.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "f32_vrsubc_relu_test",
|
|
srcs = [
|
|
"test/f32-vrsubc-relu.cc",
|
|
"test/vbinaryc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_dwconv_minmax_test",
|
|
srcs = [
|
|
"test/qs8-dwconv-minmax.cc",
|
|
"test/dwconv-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_gavgpool_minmax_test",
|
|
srcs = [
|
|
"test/qs8-gavgpool-minmax.cc",
|
|
"test/gavgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_gemm_minmax_test",
|
|
srcs = [
|
|
"test/qs8-gemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_igemm_minmax_test",
|
|
srcs = [
|
|
"test/qs8-igemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_requantization_test",
|
|
srcs = [
|
|
"src/xnnpack/requantization-stubs.h",
|
|
"test/qs8-requantization.cc",
|
|
"test/requantization-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_vadd_minmax_test",
|
|
srcs = [
|
|
"test/qs8-vadd-minmax.cc",
|
|
"test/vadd-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qs8_vaddc_minmax_test",
|
|
srcs = [
|
|
"test/qs8-vaddc-minmax.cc",
|
|
"test/vaddc-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_avgpool_minmax_test",
|
|
srcs = [
|
|
"test/qu8-avgpool-minmax.cc",
|
|
"test/avgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_dwconv_minmax_test",
|
|
srcs = [
|
|
"test/qu8-dwconv-minmax.cc",
|
|
"test/dwconv-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_igemm_minmax_test",
|
|
srcs = [
|
|
"test/qu8-igemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_gavgpool_minmax_test",
|
|
srcs = [
|
|
"test/qu8-gavgpool-minmax.cc",
|
|
"test/gavgpool-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_gemm_minmax_test",
|
|
srcs = [
|
|
"test/qu8-gemm-minmax.cc",
|
|
"test/gemm-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS + [":packing"],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_requantization_test",
|
|
srcs = [
|
|
"src/xnnpack/requantization-stubs.h",
|
|
"test/qu8-requantization.cc",
|
|
"test/requantization-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "qu8_vadd_minmax_test",
|
|
srcs = [
|
|
"test/qu8-vadd-minmax.cc",
|
|
"test/vadd-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "u8_clamp_test",
|
|
srcs = [
|
|
"test/u8-clamp.cc",
|
|
"test/clamp-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "u8_lut32norm_test",
|
|
srcs = [
|
|
"test/u8-lut32norm.cc",
|
|
"test/lut-norm-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "u8_maxpool_minmax_test",
|
|
srcs = [
|
|
"test/u8-maxpool-minmax.cc",
|
|
"test/maxpool-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "u8_rmax_test",
|
|
srcs = [
|
|
"test/u8-rmax.cc",
|
|
"test/rmax-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_depthtospace2d_chw2hwc_test",
|
|
srcs = [
|
|
"test/x32-depthtospace2d-chw2hwc.cc",
|
|
"test/depthtospace-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_fill_test",
|
|
srcs = [
|
|
"test/x32-fill.cc",
|
|
"test/fill-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_packx_test",
|
|
srcs = [
|
|
"test/x32-packx.cc",
|
|
"test/pack-microkernel-tester.h",
|
|
"src/xnnpack/AlignedAllocator.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_pad_test",
|
|
srcs = [
|
|
"test/x32-pad.cc",
|
|
"test/pad-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_unpool_test",
|
|
srcs = [
|
|
"test/x32-unpool.cc",
|
|
"test/unpool-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x32_zip_test",
|
|
srcs = [
|
|
"test/x32-zip.cc",
|
|
"test/zip-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x8_lut_test",
|
|
srcs = [
|
|
"test/x8-lut.cc",
|
|
"test/lut-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "x8_zip_test",
|
|
srcs = [
|
|
"test/x8-zip.cc",
|
|
"test/zip-microkernel-tester.h",
|
|
] + MICROKERNEL_TEST_HDRS,
|
|
deps = MICROKERNEL_TEST_DEPS,
|
|
)
|
|
|
|
########################## Size tests for the library #########################
|
|
|
|
xnnpack_binary(
|
|
name = "operator_size_test",
|
|
srcs = ["test/operator-size.c"],
|
|
deps = [":xnnpack_operators_nhwc_f32"],
|
|
)
|
|
|
|
xnnpack_binary(
|
|
name = "subgraph_size_test",
|
|
srcs = ["test/subgraph-size.c"],
|
|
deps = [":XNNPACK"],
|
|
)
|
|
|
|
########################### Unit tests for operators ##########################
|
|
|
|
xnnpack_unit_test(
|
|
name = "abs_nc_test",
|
|
srcs = [
|
|
"test/abs-nc.cc",
|
|
"test/abs-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "add_nd_test",
|
|
timeout = "moderate",
|
|
srcs = [
|
|
"test/add-nd.cc",
|
|
"test/binary-elementwise-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "argmax_pooling_nhwc_test",
|
|
srcs = [
|
|
"test/argmax-pooling-nhwc.cc",
|
|
"test/argmax-pooling-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "average_pooling_nhwc_test",
|
|
srcs = [
|
|
"test/average-pooling-nhwc.cc",
|
|
"test/average-pooling-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "bankers_rounding_nc_test",
|
|
srcs = [
|
|
"test/bankers-rounding-nc.cc",
|
|
"test/bankers-rounding-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "ceiling_nc_test",
|
|
srcs = [
|
|
"test/ceiling-nc.cc",
|
|
"test/ceiling-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "channel_shuffle_nc_test",
|
|
srcs = [
|
|
"test/channel-shuffle-nc.cc",
|
|
"test/channel-shuffle-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "clamp_nc_test",
|
|
srcs = [
|
|
"test/clamp-nc.cc",
|
|
"test/clamp-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "constant_pad_nd_test",
|
|
srcs = [
|
|
"test/constant-pad-nd.cc",
|
|
"test/constant-pad-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "convolution_nhwc_test",
|
|
timeout = "moderate",
|
|
srcs = [
|
|
"test/convolution-nhwc.cc",
|
|
"test/convolution-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "convolution_nchw_test",
|
|
timeout = "moderate",
|
|
srcs = [
|
|
"test/convolution-nchw.cc",
|
|
"test/convolution-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "copy_nc_test",
|
|
srcs = [
|
|
"test/copy-nc.cc",
|
|
"test/copy-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "deconvolution_nhwc_test",
|
|
timeout = "moderate",
|
|
srcs = [
|
|
"test/deconvolution-nhwc.cc",
|
|
"test/deconvolution-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "depth_to_space_nchw2nhwc_test",
|
|
srcs = [
|
|
"test/depth-to-space-nchw2nhwc.cc",
|
|
"test/depth-to-space-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "depth_to_space_nhwc_test",
|
|
srcs = [
|
|
"test/depth-to-space-nhwc.cc",
|
|
"test/depth-to-space-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "divide_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/divide-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "elu_nc_test",
|
|
srcs = [
|
|
"test/elu-nc.cc",
|
|
"test/elu-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "fully_connected_nc_test",
|
|
srcs = [
|
|
"test/fully-connected-nc.cc",
|
|
"test/fully-connected-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "floor_nc_test",
|
|
srcs = [
|
|
"test/floor-nc.cc",
|
|
"test/floor-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "global_average_pooling_nwc_test",
|
|
srcs = [
|
|
"test/global-average-pooling-nwc.cc",
|
|
"test/global-average-pooling-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "global_average_pooling_ncw_test",
|
|
srcs = [
|
|
"test/global-average-pooling-ncw.cc",
|
|
"test/global-average-pooling-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "hardswish_nc_test",
|
|
srcs = [
|
|
"test/hardswish-nc.cc",
|
|
"test/hardswish-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "leaky_relu_nc_test",
|
|
srcs = [
|
|
"test/leaky-relu-nc.cc",
|
|
"test/leaky-relu-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "max_pooling_nhwc_test",
|
|
timeout = "moderate",
|
|
srcs = [
|
|
"test/max-pooling-nhwc.cc",
|
|
"test/max-pooling-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "maximum_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/maximum-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "minimum_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/minimum-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "multiply_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/multiply-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "negate_nc_test",
|
|
srcs = [
|
|
"test/negate-nc.cc",
|
|
"test/negate-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "prelu_nc_test",
|
|
srcs = [
|
|
"test/prelu-nc.cc",
|
|
"test/prelu-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "resize_bilinear_nhwc_test",
|
|
srcs = [
|
|
"test/resize-bilinear-nhwc.cc",
|
|
"test/resize-bilinear-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "resize_bilinear_nchw_test",
|
|
srcs = [
|
|
"test/resize-bilinear-nchw.cc",
|
|
"test/resize-bilinear-operator-tester.h",
|
|
] + OPERATOR_TEST_PARAMS_HDRS,
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "sigmoid_nc_test",
|
|
srcs = [
|
|
"test/sigmoid-nc.cc",
|
|
"test/sigmoid-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "softmax_nc_test",
|
|
srcs = [
|
|
"test/softmax-nc.cc",
|
|
"test/softmax-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "square_nc_test",
|
|
srcs = [
|
|
"test/square-nc.cc",
|
|
"test/square-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "square_root_nc_test",
|
|
srcs = [
|
|
"test/square-root-nc.cc",
|
|
"test/square-root-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "squared_difference_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/squared-difference-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "subtract_nd_test",
|
|
srcs = [
|
|
"test/binary-elementwise-operator-tester.h",
|
|
"test/subtract-nd.cc",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "truncation_nc_test",
|
|
srcs = [
|
|
"test/truncation-nc.cc",
|
|
"test/truncation-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "unpooling_nhwc_test",
|
|
srcs = [
|
|
"test/unpooling-nhwc.cc",
|
|
"test/unpooling-operator-tester.h",
|
|
],
|
|
deps = OPERATOR_TEST_DEPS,
|
|
)
|
|
|
|
############################### Misc unit tests ###############################
|
|
|
|
xnnpack_unit_test(
|
|
name = "memory_planner_test",
|
|
srcs = [
|
|
"test/memory-planner-test.cc",
|
|
],
|
|
deps = [
|
|
":XNNPACK",
|
|
":memory_planner",
|
|
],
|
|
)
|
|
|
|
xnnpack_unit_test(
|
|
name = "subgraph_nchw_test",
|
|
srcs = [
|
|
"src/xnnpack/subgraph.h",
|
|
"test/subgraph-nchw.cc",
|
|
"test/subgraph-tester.h",
|
|
],
|
|
deps = [
|
|
":XNNPACK",
|
|
],
|
|
)
|
|
|
|
############################# Build configurations #############################
|
|
|
|
# Enables usage of assembly kernels.
|
|
config_setting(
|
|
name = "xnn_enable_assembly_explicit_true",
|
|
define_values = {"xnn_enable_assembly": "true"},
|
|
)
|
|
|
|
# Disables usage of assembly kernels.
|
|
config_setting(
|
|
name = "xnn_enable_assembly_explicit_false",
|
|
define_values = {"xnn_enable_assembly": "false"},
|
|
)
|
|
|
|
# Enables usage of sparse inference.
|
|
config_setting(
|
|
name = "xnn_enable_sparse_explicit_true",
|
|
define_values = {"xnn_enable_sparse": "true"},
|
|
)
|
|
|
|
# Disables usage of sparse inference.
|
|
config_setting(
|
|
name = "xnn_enable_sparse_explicit_false",
|
|
define_values = {"xnn_enable_sparse": "false"},
|
|
)
|
|
|
|
# Disables usage of HMP-aware optimizations.
|
|
config_setting(
|
|
name = "xnn_enable_hmp_explicit_false",
|
|
define_values = {"xnn_enable_hmp": "false"},
|
|
)
|
|
|
|
# Enable usage of optimized memory allocation
|
|
config_setting(
|
|
name = "xnn_enable_memopt_explicit_true",
|
|
define_values = {"xnnpack_enable_memopt": "true"},
|
|
)
|
|
|
|
# Disable usage of optimized memory allocation
|
|
config_setting(
|
|
name = "xnn_enable_memopt_explicit_false",
|
|
define_values = {"xnnpack_enable_memopt": "false"},
|
|
)
|
|
|
|
# Builds with -c dbg
|
|
config_setting(
|
|
name = "debug_build",
|
|
values = {
|
|
"compilation_mode": "dbg",
|
|
},
|
|
)
|
|
|
|
# Builds with -c opt
|
|
config_setting(
|
|
name = "optimized_build",
|
|
values = {
|
|
"compilation_mode": "opt",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_k8",
|
|
values = {"cpu": "k8"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_arm",
|
|
values = {"cpu": "arm"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_armeabi",
|
|
values = {"cpu": "armeabi"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_armhf",
|
|
values = {"cpu": "armhf"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_armv7a",
|
|
values = {"cpu": "armv7a"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "linux_aarch64",
|
|
values = {"cpu": "aarch64"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "android",
|
|
values = {"crosstool_top": "//external:android/crosstool"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "android_armv7",
|
|
values = {
|
|
"crosstool_top": "//external:android/crosstool",
|
|
"cpu": "armeabi-v7a",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "android_arm64",
|
|
values = {
|
|
"crosstool_top": "//external:android/crosstool",
|
|
"cpu": "arm64-v8a",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "android_x86",
|
|
values = {
|
|
"crosstool_top": "//external:android/crosstool",
|
|
"cpu": "x86",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "android_x86_64",
|
|
values = {
|
|
"crosstool_top": "//external:android/crosstool",
|
|
"cpu": "x86_64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "windows_x86_64",
|
|
values = {"cpu": "x64_windows"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "windows_x86_64_clang",
|
|
values = {
|
|
"compiler": "clang-cl",
|
|
"cpu": "x64_windows",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "windows_x86_64_mingw",
|
|
values = {
|
|
"compiler": "mingw-gcc",
|
|
"cpu": "x64_windows",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "windows_x86_64_msys",
|
|
values = {
|
|
"compiler": "msys-gcc",
|
|
"cpu": "x64_windows",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "macos_x86_64",
|
|
values = {
|
|
"apple_platform_type": "macos",
|
|
"cpu": "darwin",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "macos_arm64",
|
|
values = {
|
|
"apple_platform_type": "macos",
|
|
"cpu": "darwin_arm64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "emscripten",
|
|
values = {"crosstool_top": "//emscripten_toolchain:everything"},
|
|
)
|
|
|
|
config_setting(
|
|
name = "emscripten_wasm",
|
|
values = {
|
|
"crosstool_top": "//emscripten_toolchain:everything",
|
|
"cpu": "wasm",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "emscripten_wasmsimd",
|
|
values = {
|
|
"crosstool_top": "//emscripten_toolchain:everything",
|
|
"cpu": "wasm",
|
|
"copt": "-msimd128",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "ios_armv7",
|
|
values = {
|
|
"apple_platform_type": "ios",
|
|
"cpu": "ios_armv7",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "ios_arm64",
|
|
values = {
|
|
"apple_platform_type": "ios",
|
|
"cpu": "ios_arm64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "ios_arm64e",
|
|
values = {
|
|
"apple_platform_type": "ios",
|
|
"cpu": "ios_arm64e",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "ios_x86",
|
|
values = {
|
|
"apple_platform_type": "ios",
|
|
"cpu": "ios_i386",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "ios_x86_64",
|
|
values = {
|
|
"apple_platform_type": "ios",
|
|
"cpu": "ios_x86_64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "watchos_armv7k",
|
|
values = {
|
|
"apple_platform_type": "watchos",
|
|
"cpu": "watchos_armv7k",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "watchos_arm64_32",
|
|
values = {
|
|
"apple_platform_type": "watchos",
|
|
"cpu": "watchos_arm64_32",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "watchos_x86",
|
|
values = {
|
|
"apple_platform_type": "watchos",
|
|
"cpu": "watchos_i386",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "watchos_x86_64",
|
|
values = {
|
|
"apple_platform_type": "watchos",
|
|
"cpu": "watchos_x86_64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "tvos_arm64",
|
|
values = {
|
|
"apple_platform_type": "tvos",
|
|
"cpu": "tvos_arm64",
|
|
},
|
|
)
|
|
|
|
config_setting(
|
|
name = "tvos_x86_64",
|
|
values = {
|
|
"apple_platform_type": "tvos",
|
|
"cpu": "tvos_x86_64",
|
|
},
|
|
)
|