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.

376 lines
11 KiB

//
// Copyright (c) 2020 The Khronos Group Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
#include "procs.h"
#include "subhelpers.h"
#include "harness/conversions.h"
#include "harness/typeWrappers.h"
// These need to stay in sync with the kernel source below
#define NUM_LOC 49
#define INST_LOC_MASK 0x7f
#define INST_OP_SHIFT 0
#define INST_OP_MASK 0xf
#define INST_LOC_SHIFT 4
#define INST_VAL_SHIFT 12
#define INST_VAL_MASK 0x7ffff
#define INST_END 0x0
#define INST_STORE 0x1
#define INST_WAIT 0x2
#define INST_COUNT 0x3
static const char *ifp_source =
"#define NUM_LOC 49\n"
"#define INST_LOC_MASK 0x7f\n"
"#define INST_OP_SHIFT 0\n"
"#define INST_OP_MASK 0xf\n"
"#define INST_LOC_SHIFT 4\n"
"#define INST_VAL_SHIFT 12\n"
"#define INST_VAL_MASK 0x7ffff\n"
"#define INST_END 0x0\n"
"#define INST_STORE 0x1\n"
"#define INST_WAIT 0x2\n"
"#define INST_COUNT 0x3\n"
"\n"
"__kernel void\n"
"test_ifp(const __global int *in, __global int4 *xy, __global int *out)\n"
"{\n"
" __local atomic_int loc[NUM_LOC];\n"
"\n"
" // Don't run if there is only one sub group\n"
" if (get_num_sub_groups() == 1)\n"
" return;\n"
"\n"
" // First initialize loc[]\n"
" int lid = (int)get_local_id(0);\n"
"\n"
" if (lid < NUM_LOC)\n"
" atomic_init(loc+lid, 0);\n"
"\n"
" work_group_barrier(CLK_LOCAL_MEM_FENCE);\n"
"\n"
" // Compute pointer to this sub group's \"instructions\"\n"
" const __global int *pc = in +\n"
" ((int)get_group_id(0)*(int)get_enqueued_num_sub_groups() +\n"
" (int)get_sub_group_id()) *\n"
" (NUM_LOC+1);\n"
"\n"
" // Set up to \"run\"\n"
" bool ok = (int)get_sub_group_local_id() == 0;\n"
" bool run = true;\n"
"\n"
" while (run) {\n"
" int inst = *pc++;\n"
" int iop = (inst >> INST_OP_SHIFT) & INST_OP_MASK;\n"
" int iloc = (inst >> INST_LOC_SHIFT) & INST_LOC_MASK;\n"
" int ival = (inst >> INST_VAL_SHIFT) & INST_VAL_MASK;\n"
"\n"
" switch (iop) {\n"
" case INST_STORE:\n"
" if (ok)\n"
" atomic_store(loc+iloc, ival);\n"
" break;\n"
" case INST_WAIT:\n"
" if (ok) {\n"
" while (atomic_load(loc+iloc) != ival)\n"
" ;\n"
" }\n"
" break;\n"
" case INST_COUNT:\n"
" if (ok) {\n"
" int i;\n"
" for (i=0;i<ival;++i)\n"
" atomic_fetch_add(loc+iloc, 1);\n"
" }\n"
" break;\n"
" case INST_END:\n"
" run = false;\n"
" break;\n"
" }\n"
"\n"
" sub_group_barrier(CLK_LOCAL_MEM_FENCE);\n"
" }\n"
"\n"
" work_group_barrier(CLK_LOCAL_MEM_FENCE);\n"
"\n"
" // Save this group's result\n"
" __global int *op = out + (int)get_group_id(0)*NUM_LOC;\n"
" if (lid < NUM_LOC)\n"
" op[lid] = atomic_load(loc+lid);\n"
"}\n";
// Independent forward progress stuff
// Note:
// Output needs num_groups * NUM_LOC elements
// local_size must be > NUM_LOC
// Input needs num_groups * num_sub_groups * (NUM_LOC+1) elements
static inline int inst(int op, int loc, int val)
{
return (val << INST_VAL_SHIFT) | (loc << INST_LOC_SHIFT)
| (op << INST_OP_SHIFT);
}
void gen_insts(cl_int *x, cl_int *p, int n)
{
int i, j0, j1;
int val;
int ii[NUM_LOC];
// Create a random permutation of 0...NUM_LOC-1
ii[0] = 0;
for (i = 1; i < NUM_LOC; ++i)
{
j0 = random_in_range(0, i, gMTdata);
if (j0 != i) ii[i] = ii[j0];
ii[j0] = i;
}
// Initialize "instruction pointers"
memset(p, 0, n * 4);
for (i = 0; i < NUM_LOC; ++i)
{
// Randomly choose 2 different sub groups
// One does a random amount of work, and the other waits for it
j0 = random_in_range(0, n - 1, gMTdata);
do
{
j1 = random_in_range(0, n - 1, gMTdata);
} while (j1 == j0);
// Randomly choose a wait value and assign "instructions"
val = random_in_range(100, 200 + 10 * NUM_LOC, gMTdata);
x[j0 * (NUM_LOC + 1) + p[j0]] = inst(INST_COUNT, ii[i], val);
x[j1 * (NUM_LOC + 1) + p[j1]] = inst(INST_WAIT, ii[i], val);
++p[j0];
++p[j1];
}
// Last "inst" for each sub group is END
for (i = 0; i < n; ++i) x[i * (NUM_LOC + 1) + p[i]] = inst(INST_END, 0, 0);
}
// Execute one group's "instructions"
void run_insts(cl_int *x, cl_int *p, int n)
{
int i, nend;
bool scont;
cl_int loc[NUM_LOC];
// Initialize result and "instruction pointers"
memset(loc, 0, sizeof(loc));
memset(p, 0, 4 * n);
// Repetitively loop over subgroups with each executing "instructions" until
// blocked The loop terminates when all subgroups have hit the "END
// instruction"
do
{
nend = 0;
for (i = 0; i < n; ++i)
{
do
{
cl_int inst = x[i * (NUM_LOC + 1) + p[i]];
cl_int iop = (inst >> INST_OP_SHIFT) & INST_OP_MASK;
cl_int iloc = (inst >> INST_LOC_SHIFT) & INST_LOC_MASK;
cl_int ival = (inst >> INST_VAL_SHIFT) & INST_VAL_MASK;
scont = false;
switch (iop)
{
case INST_STORE:
loc[iloc] = ival;
++p[i];
scont = true;
break;
case INST_WAIT:
if (loc[iloc] == ival)
{
++p[i];
scont = true;
}
break;
case INST_COUNT:
loc[iloc] += ival;
++p[i];
scont = true;
break;
case INST_END: ++nend; break;
}
} while (scont);
}
} while (nend < n);
// Return result, reusing "p"
memcpy(p, loc, sizeof(loc));
}
struct IFP
{
static void gen(cl_int *x, cl_int *t, cl_int *,
const WorkGroupParams &test_params)
{
int k;
int nw = test_params.local_workgroup_size;
int ns = test_params.subgroup_size;
int ng = test_params.global_workgroup_size;
int nj = (nw + ns - 1) / ns;
ng = ng / nw;
// We need at least 2 sub groups per group for this test
if (nj == 1) return;
for (k = 0; k < ng; ++k)
{
gen_insts(x, t, nj);
x += nj * (NUM_LOC + 1);
}
}
static int chk(cl_int *x, cl_int *y, cl_int *t, cl_int *, cl_int *,
const WorkGroupParams &test_params)
{
int i, k;
int nw = test_params.local_workgroup_size;
int ns = test_params.subgroup_size;
int ng = test_params.global_workgroup_size;
int nj = (nw + ns - 1) / ns;
ng = ng / nw;
// We need at least 2 sub groups per group for this tes
if (nj == 1) return 0;
log_info(" independent forward progress...\n");
for (k = 0; k < ng; ++k)
{
run_insts(x, t, nj);
for (i = 0; i < NUM_LOC; ++i)
{
if (t[i] != y[i])
{
log_error(
"ERROR: mismatch at element %d in work group %d\n", i,
k);
return -1;
}
}
x += nj * (NUM_LOC + 1);
y += NUM_LOC;
}
return 0;
}
};
int test_ifp(cl_device_id device, cl_context context, cl_command_queue queue,
int num_elements, bool useCoreSubgroups)
{
int error = TEST_PASS;
// Global/local work group sizes
// Adjust these individually below if desired/needed
constexpr size_t global_work_size = 2000;
constexpr size_t local_work_size = 200;
WorkGroupParams test_params(global_work_size, local_work_size);
test_params.use_core_subgroups = useCoreSubgroups;
test_params.dynsc = NUM_LOC + 1;
error = test<cl_int, IFP>::run(device, context, queue, num_elements,
"test_ifp", ifp_source, test_params);
return error;
}
static test_status checkIFPSupport(cl_device_id device, bool &ifpSupport)
{
cl_uint ifp_supported;
cl_uint error;
error = clGetDeviceInfo(device,
CL_DEVICE_SUB_GROUP_INDEPENDENT_FORWARD_PROGRESS,
sizeof(ifp_supported), &ifp_supported, NULL);
if (error != CL_SUCCESS)
{
print_error(
error,
"Unable to get CL_DEVICE_SUB_GROUP_INDEPENDENT_FORWARD_PROGRESS "
"capability");
return TEST_FAIL;
}
// skip testing ifp
if (ifp_supported != 1)
{
log_info("INDEPENDENT FORWARD PROGRESS not supported...\n");
ifpSupport = false;
}
else
{
log_info("INDEPENDENT FORWARD PROGRESS supported...\n");
ifpSupport = true;
}
return TEST_PASS;
}
int test_ifp_core(cl_device_id device, cl_context context,
cl_command_queue queue, int num_elements)
{
bool ifpSupport = true;
test_status error;
error = checkIFPSupport(device, ifpSupport);
if (error != TEST_PASS)
{
return error;
}
if (ifpSupport == false)
{
log_info("Independed forward progress skipped.\n");
return TEST_SKIPPED_ITSELF;
}
return test_ifp(device, context, queue, num_elements, true);
}
int test_ifp_ext(cl_device_id device, cl_context context,
cl_command_queue queue, int num_elements)
{
bool hasExtension = is_extension_available(device, "cl_khr_subgroups");
bool ifpSupport = true;
if (!hasExtension)
{
log_info(
"Device does not support 'cl_khr_subgroups'. Skipping the test.\n");
return TEST_SKIPPED_ITSELF;
}
// ifp only in subgroup functions tests:
test_status error;
error = checkIFPSupport(device, ifpSupport);
if (error != TEST_PASS)
{
return error;
}
if (ifpSupport == false)
{
log_info(
"Error reason: the extension cl_khr_subgroups requires that "
"Independed forward progress has to be supported by device.\n");
return TEST_FAIL;
}
return test_ifp(device, context, queue, num_elements, false);
}