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.

360 lines
14 KiB

"""Generates the specialized gemm functions."""
import mul_Nx8_Mx8_neon
import qnt_Nx8_neon
import zip_Nx8_neon
_QUANTIZED_8BIT = 'quantized_8bit'
_FULL_32BIT = 'full_32bit'
_FULL_FLOAT = 'full_float'
class Error(Exception):
"""Module level error."""
class ConfigurationError(Error):
"""Runtime configuration error."""
def GenerateCommonTempsCountersAndConsts(emitter, rows):
emitter.EmitDeclare('const std::int32_t', 'row_chunks', 'm / 3')
emitter.EmitDeclare('const std::int32_t', 'col_chunks', 'n / 3')
emitter.EmitDeclare('const std::int32_t', 'padded_k', '((k + 7) / 8) * 8')
emitter.EmitDeclare('const std::int32_t', 'chunk_size', 'k * 3')
emitter.EmitDeclare('const std::int32_t', 'zipped_chunk_size',
'(padded_k + 16) * 3')
emitter.EmitDeclare('const std::int32_t', 'zipped_rhs_size',
'(padded_k + 16) * n')
emitter.EmitDeclare('const std::uint8_t*', 'lhs_chunk', 'lhs')
emitter.EmitDeclare('const std::uint8_t*', 'rhs_chunk', 'rhs')
emitter.EmitDeclare('std::uint8_t*', 'zipped_lhs', 'scratch')
emitter.EmitDeclare(
'std::int32_t*', 'zipped_lhs_3_offsets',
'reinterpret_cast<std::int32_t*>(zipped_lhs + padded_k * 3)')
if rows is not 0:
emitter.EmitDeclare(
'std::int32_t*', 'zipped_lhs_%d_offsets' % rows,
'reinterpret_cast<std::int32_t*>(zipped_lhs + padded_k * %d)' % rows)
emitter.EmitDeclare('std::uint8_t*', 'zipped_rhs',
'scratch + zipped_chunk_size')
emitter.EmitDeclare('std::uint8_t*', 'zipped_rhs_chunk', 'zipped_rhs')
emitter.EmitDeclare('const std::int32_t', 'result_chunk_stride',
'result_stride * 3')
emitter.EmitNewline()
def GenerateQuantized8BitTempsCountersAndConsts(emitter, rows):
"""Generates all the boilerplate variables for the q8 gemm function."""
GenerateCommonTempsCountersAndConsts(emitter, rows)
emitter.EmitDeclare('const std::int32_t', 'const_offset',
'lhs_offset * rhs_offset * k + result_offset')
emitter.EmitDeclare('const std::int32_t', 'rounding_offset',
'(1 << (shift - 1))')
emitter.EmitDeclare('std::int32_t*', 'temp_result',
'reinterpret_cast<std::int32_t*>('
'scratch + zipped_chunk_size + zipped_rhs_size)')
emitter.EmitDeclare('std::uint8_t*', 'result_chunk', 'result')
emitter.EmitDeclare('std::int32_t*', 'mul_result_chunk', 'temp_result')
emitter.EmitDeclare('const std::int32_t', 'mul_result_chunk_stride_bytes',
'((n * 4 + 7) / 8) * 8')
emitter.EmitNewline()
def GenerateFullTempsCountersAndConsts(emitter, result_type, rows):
"""Generates all the boilerplate variables for the int32 and float gemms."""
GenerateCommonTempsCountersAndConsts(emitter, rows)
emitter.EmitDeclare('const std::int32_t', 'const_offset',
'lhs_offset * rhs_offset * k')
emitter.EmitDeclare(result_type, 'result_chunk', 'result')
emitter.EmitDeclare(result_type, 'mul_result_chunk', 'result')
emitter.EmitDeclare('const std::int32_t', 'mul_result_chunk_stride_bytes',
'result_stride * 4')
emitter.EmitNewline()
def ZipName(rows, leftovers, aligned):
return zip_Nx8_neon.BuildName(rows, leftovers, aligned)
def GenerateZipRhs(emitter, aligned, cols, leftovers):
"""Emits the code responsible for zipping the rhs matrix."""
emitter.EmitOpenBracket('for (int i = 0; i < col_chunks; ++i)')
emitter.EmitCall(
ZipName(3, leftovers, aligned),
['rhs_chunk', 'k', 'k', 'zipped_rhs_chunk', 'lhs_offset', 0])
emitter.EmitAssignIncrement('rhs_chunk', 'chunk_size')
emitter.EmitAssignIncrement('zipped_rhs_chunk', 'zipped_chunk_size')
emitter.EmitCloseBracket()
if cols is not 0:
emitter.EmitCall(
ZipName(cols, leftovers, aligned),
['rhs_chunk', 'k', 'k', 'zipped_rhs_chunk', 'lhs_offset', 0])
emitter.EmitNewline()
def MulName(result_type, lhs_add, rhs_add, rows, cols):
return mul_Nx8_Mx8_neon.BuildName(result_type, lhs_add, rhs_add, rows, cols)
def GetMulParams(result_type):
params = ['zipped_lhs', 'zipped_rhs_chunk', 'padded_k', 'mul_result_chunk',
'mul_result_chunk_stride_bytes']
if result_type is 'float':
params.append('result_scale')
return params
def GenerateMulRows(emitter, result, result_type, lhs_add, rhs_add, aligned,
rows, cols, leftovers):
"""Emits code responsible for multiplication of one horizontal lhs strip."""
emitter.EmitCall(
ZipName(rows, leftovers, aligned),
['lhs_chunk', 'k', 'k', 'zipped_lhs', 'rhs_offset', 'const_offset'])
emitter.EmitAssign('zipped_rhs_chunk', 'zipped_rhs')
emitter.EmitAssign('mul_result_chunk', result)
emitter.EmitOpenBracket('for (int j = 0; j < col_chunks; ++j)')
emitter.EmitCall(
MulName(result_type, lhs_add, rhs_add, rows, 3),
GetMulParams(result_type))
emitter.EmitAssignIncrement('zipped_rhs_chunk', 'zipped_chunk_size')
emitter.EmitAssignIncrement('mul_result_chunk', 3)
emitter.EmitCloseBracket()
if cols is not 0:
emitter.EmitCall(
MulName(result_type, lhs_add, rhs_add, rows, cols),
GetMulParams(result_type))
def GenerateQuantized8BitMul(emitter, aligned, rows, cols, leftovers):
"""Emits code for all lhs strips & leftover rows. Quantize after mul code."""
emitter.EmitOpenBracket('for (int i = 0; i < row_chunks; ++i)')
GenerateMulRows(emitter, 'temp_result', 'int32', False, True, aligned, 3,
cols, leftovers)
emitter.EmitCall(
qnt_Nx8_neon.BuildMultiQuantizeName(aligned, 3),
['temp_result', 'n', 'mul_result_chunk_stride_bytes',
'zipped_lhs_3_offsets', 'result_chunk', 'result_stride',
'multiplicative_offset', 'rounding_offset', '-shift'])
emitter.EmitAssignIncrement('lhs_chunk', 'chunk_size')
emitter.EmitAssignIncrement('result_chunk', 'result_chunk_stride')
emitter.EmitCloseBracket()
emitter.EmitNewline()
if rows is not 0:
GenerateMulRows(emitter, 'temp_result', 'int32', False, True, aligned, rows,
cols, leftovers)
emitter.EmitCall(
qnt_Nx8_neon.BuildMultiQuantizeName(aligned, rows),
['temp_result', 'n', 'mul_result_chunk_stride_bytes',
'zipped_lhs_%d_offsets' % rows, 'result_chunk', 'result_stride',
'multiplicative_offset', 'rounding_offset', '-shift'])
def GenerateFullMul(emitter, result_type, aligned, rows, cols, leftovers):
emitter.EmitOpenBracket('for (int i = 0; i < row_chunks; ++i)')
GenerateMulRows(emitter, 'result_chunk', result_type, True, True, aligned, 3,
cols, leftovers)
emitter.EmitAssignIncrement('lhs_chunk', 'chunk_size')
emitter.EmitAssignIncrement('result_chunk', 'result_chunk_stride')
emitter.EmitCloseBracket()
emitter.EmitNewline()
if rows is not 0:
GenerateMulRows(emitter, 'result_chunk', result_type, True, True, aligned,
rows, cols, leftovers)
def BuildName(output_type, aligned, rows, cols, leftover):
name = BuildMainGemmName(output_type) + '_%d_%d_%d' % (rows, cols, leftover)
if aligned:
name += '_aligned'
return name
def GetCommonGemmParameters():
return [['std::uint8_t*', 'scratch'], ['const std::uint8_t*', 'lhs'],
['const std::uint8_t*', 'rhs'], ['std::int32_t', 'm'],
['std::int32_t', 'n'], ['std::int32_t', 'k'],
['std::int32_t', 'lhs_offset'], ['std::int32_t', 'rhs_offset']]
def GetGemmParameters(output_type, extra_params=None):
"""Prepares a (type, parameter) array for the gemm functions."""
if extra_params is None:
extra_params = []
params = GetCommonGemmParameters()
if output_type is _QUANTIZED_8BIT:
params += [['std::int32_t', 'result_offset'],
['std::int32_t', 'multiplicative_offset'],
['std::int32_t', 'shift'], ['std::uint8_t*', 'result']]
elif output_type is _FULL_32BIT:
params += [['std::int32_t*', 'result']]
elif output_type is _FULL_FLOAT:
params += [['float', 'result_scale'], ['float*', 'result']]
else:
raise ConfigurationError('Unsupported output type: %s' % output_type)
return params + extra_params
def GetStridedGemmParameters(output_type):
return GetGemmParameters(output_type, [['std::int32_t', 'result_stride']])
def GenerateGemm(emitter, output_type, aligned, rows, cols, leftovers):
"""Build one gemm function for given row, col, and depth leftovers."""
emitter.EmitFunctionBeginA(
BuildName(output_type, aligned, rows, cols, leftovers),
GetStridedGemmParameters(output_type), 'void')
emitter.EmitAssert('m %% 3 == %d' % rows)
emitter.EmitAssert('n %% 3 == %d' % cols)
emitter.EmitAssert('k %% 8 == %d' % leftovers)
if output_type is _QUANTIZED_8BIT:
GenerateQuantized8BitTempsCountersAndConsts(emitter, rows)
GenerateZipRhs(emitter, aligned, cols, leftovers)
GenerateQuantized8BitMul(emitter, aligned, rows, cols, leftovers)
elif output_type is _FULL_32BIT:
GenerateFullTempsCountersAndConsts(emitter, 'std::int32_t*', rows)
GenerateZipRhs(emitter, aligned, cols, leftovers)
GenerateFullMul(emitter, 'int32', aligned, rows, cols, leftovers)
elif output_type is _FULL_FLOAT:
GenerateFullTempsCountersAndConsts(emitter, 'float*', rows)
GenerateZipRhs(emitter, aligned, cols, leftovers)
GenerateFullMul(emitter, 'float', aligned, rows, cols, leftovers)
else:
raise ConfigurationError('Unknown output type: %s' % output_type)
emitter.EmitFunctionEnd()
def GenerateGemmCall(emitter, output_type, aligned, m_mod, n_mod, leftovers):
emitter.EmitCall(
emitter.Scope('internal',
BuildName(output_type, aligned, m_mod, n_mod, leftovers)),
[p for (unused_t, p) in GetStridedGemmParameters(output_type)])
def GenerateGemmSwitch3(emitter, output_type, aligned, m_mod, n_mod):
"""Third level of main switch, choose optimized version on depth leftover."""
emitter.EmitSwitch('k % 8')
for leftovers in range(0, 8):
emitter.EmitCase(leftovers)
emitter.PushIndent()
GenerateGemmCall(emitter, output_type, aligned, m_mod, n_mod, leftovers)
emitter.EmitBreak()
emitter.PopIndent()
emitter.EmitSwitchEnd()
def GenerateGemmSwitch2(emitter, output_type, aligned, m_mod):
"""Second level of main switch, choose optimized version on cols leftover."""
emitter.EmitSwitch('n % 3')
for n_mod in range(0, 3):
emitter.EmitCase(n_mod)
emitter.PushIndent()
GenerateGemmSwitch3(emitter, output_type, aligned, m_mod, n_mod)
emitter.EmitBreak()
emitter.PopIndent()
emitter.EmitSwitchEnd()
def GenerateGemmSwitch1(emitter, output_type, aligned):
"""First level of main switch, choose optimized version on rows leftover."""
emitter.EmitSwitch('m % 3')
for m_mod in range(0, 3):
emitter.EmitCase(m_mod)
emitter.PushIndent()
GenerateGemmSwitch2(emitter, output_type, aligned, m_mod)
emitter.EmitBreak()
emitter.PopIndent()
emitter.EmitSwitchEnd()
def BuildMainGemmName(output_type):
if output_type is _QUANTIZED_8BIT:
return 'gemm_q8'
elif output_type is _FULL_32BIT:
return 'gemm_i32'
elif output_type is _FULL_FLOAT:
return 'gemm_f'
else:
raise ConfigurationError('Unsupported output type: %s' % output_type)
def BuildStridedMainGemmName(output_type):
return BuildMainGemmName(output_type) + '_strided'
def GenerateMainGemmFunction(emitter, output_type):
"""Emit high level gemm function that switches between optimized versions."""
emitter.EmitFunctionBeginA(
BuildStridedMainGemmName(output_type),
GetStridedGemmParameters(output_type), 'void')
emitter.EmitDeclare('const bool', 'lhs_aligned',
'((reinterpret_cast<std::uintptr_t>(lhs) % 8) == 0)')
emitter.EmitDeclare('const bool', 'rhs_aligned',
'((reinterpret_cast<std::uintptr_t>(rhs) % 8) == 0)')
emitter.EmitDeclare('const bool', 'k_aligned', '((k % 8) == 0)')
if output_type is _QUANTIZED_8BIT:
emitter.EmitDeclare('const bool', 'result_aligned',
'((reinterpret_cast<std::uintptr_t>(result) % 8) == 0)')
emitter.EmitDeclare('const bool', 'result_stride_aligned',
'((result_stride % 8) == 0)')
emitter.EmitDeclare('const bool', 'aligned',
'lhs_aligned && rhs_aligned && result_aligned '
'&& k_aligned && result_stride_aligned')
else:
emitter.EmitDeclare('const bool', 'aligned',
'lhs_aligned && rhs_aligned && k_aligned')
emitter.EmitIf('aligned')
GenerateGemmSwitch1(emitter, output_type, True)
emitter.EmitElse()
GenerateGemmSwitch1(emitter, output_type, False)
emitter.EmitEndif()
emitter.EmitFunctionEnd()
def GenerateWrapperGemmFunction(emitter, output_type):
emitter.EmitFunctionBeginA(
BuildMainGemmName(output_type), GetGemmParameters(output_type), 'void')
emitter.EmitCall(
BuildStridedMainGemmName(output_type),
[p for (unused_t, p) in GetGemmParameters(output_type)] + ['n'])
emitter.EmitFunctionEnd()
def GenerateInternalFunctions(emitter):
"""Generate all the functions hidden in the internal namespace."""
for output_type in [_QUANTIZED_8BIT, _FULL_32BIT, _FULL_FLOAT]:
for aligned in [True, False]:
for rows in range(0, 3):
for cols in range(0, 3):
for leftover in range(0, 8):
GenerateGemm(emitter, output_type, aligned, rows, cols, leftover)
emitter.EmitNewline()
def GeneratePublicFunctions(emitter):
for output_type in [_QUANTIZED_8BIT, _FULL_32BIT, _FULL_FLOAT]:
GenerateMainGemmFunction(emitter, output_type)
emitter.EmitNewline()
GenerateWrapperGemmFunction(emitter, output_type)
emitter.EmitNewline()