CUTLASS 2.0 (#62)
CUTLASS 2.0 Substantially refactored for - Better performance, particularly for native Turing Tensor Cores - Robust and durable templates spanning the design space - Encapsulated functionality embodying modern C++11 programming techniques - Optimized containers and data types for efficient, generic, portable device code Updates to: - Quick start guide - Documentation - Utilities - CUTLASS Profiler Native Turing Tensor Cores - Efficient GEMM kernels targeting Turing Tensor Cores - Mixed-precision floating point, 8-bit integer, 4-bit integer, and binarized operands Coverage of existing CUTLASS functionality: - GEMM kernels targeting CUDA and Tensor Cores in NVIDIA GPUs - Volta Tensor Cores through native mma.sync and through WMMA API - Optimizations such as parallel reductions, threadblock rasterization, and intra-threadblock reductions - Batched GEMM operations - Complex-valued GEMMs Note: this commit and all that follow require a host compiler supporting C++11 or greater.
This commit is contained in:
@@ -0,0 +1,450 @@
|
||||
#
|
||||
# \file generator.py
|
||||
#
|
||||
# \brief Generates the CUTLASS Library's instances
|
||||
#
|
||||
|
||||
import enum
|
||||
import os.path
|
||||
import shutil
|
||||
import functools
|
||||
import operator
|
||||
|
||||
from library import *
|
||||
|
||||
|
||||
###################################################################################################
|
||||
#
|
||||
# Data structure modeling a GEMM operation
|
||||
#
|
||||
###################################################################################################
|
||||
|
||||
#
|
||||
class GemmOperation:
|
||||
#
|
||||
def __init__(self, gemm_kind, arch, tile_description, A, B, C, element_epilogue):
|
||||
self.operation_kind = OperationKind.Gemm
|
||||
self.arch = arch
|
||||
self.tile_description = tile_description
|
||||
self.gemm_kind = gemm_kind
|
||||
self.A = A
|
||||
self.B = B
|
||||
self.C = C
|
||||
self.element_epilogue = element_epilogue
|
||||
|
||||
#
|
||||
def core_name(self):
|
||||
''' The basic operation kind is prefixed with a letter indicating the accumulation type. '''
|
||||
if self.tile_description.math_instruction.opcode_class == OpcodeClass.TensorOp or \
|
||||
self.tile_description.math_instruction.opcode_class == OpcodeClass.WmmaTensorOp:
|
||||
inst_shape = "%d%d%d" % tuple(self.tile_description.math_instruction.instruction_shape)
|
||||
else:
|
||||
inst_shape = ''
|
||||
|
||||
return "%s%s%s" % (ShortDataTypeNames[self.tile_description.math_instruction.element_accumulator], inst_shape, GemmKindNames[self.gemm_kind])
|
||||
|
||||
#
|
||||
def extended_name(self):
|
||||
''' Append data types if they differ from compute type. '''
|
||||
if self.C.element != self.tile_description.math_instruction.element_accumulator and \
|
||||
self.A.element != self.tile_description.math_instruction.element_accumulator:
|
||||
extended_name = "${element_c}_${core_name}_${element_a}"
|
||||
elif self.C.element == self.tile_description.math_instruction.element_accumulator and \
|
||||
self.A.element != self.tile_description.math_instruction.element_accumulator:
|
||||
extended_name = "${core_name}_${element_a}"
|
||||
else:
|
||||
extended_name = "${core_name}"
|
||||
|
||||
extended_name = SubstituteTemplate(extended_name, {
|
||||
'element_a': DataTypeNames[self.A.element],
|
||||
'element_c': DataTypeNames[self.C.element],
|
||||
'core_name': self.core_name()
|
||||
})
|
||||
|
||||
return extended_name
|
||||
|
||||
#
|
||||
def procedural_name(self):
|
||||
''' The full procedural name indicates architecture, extended name, tile size, and layout. '''
|
||||
if self.tile_description.stages > 2:
|
||||
threadblock = "%dx%d_%dx%d" % (
|
||||
self.tile_description.threadblock_shape[0],
|
||||
self.tile_description.threadblock_shape[1],
|
||||
self.tile_description.threadblock_shape[2],
|
||||
self.tile_description.stages
|
||||
)
|
||||
else:
|
||||
threadblock = "%dx%d" % (self.tile_description.threadblock_shape[0], self.tile_description.threadblock_shape[1])
|
||||
|
||||
opcode_class_name = OpcodeClassNames[self.tile_description.math_instruction.opcode_class]
|
||||
|
||||
return SubstituteTemplate(
|
||||
"cutlass_${opcode_class}_${extended_name}_${threadblock}_${layout}",
|
||||
{
|
||||
'opcode_class': opcode_class_name,
|
||||
'extended_name': self.extended_name(),
|
||||
'threadblock': threadblock,
|
||||
'layout': "%s%s" % (ShortLayoutTypeNames[self.A.layout], ShortLayoutTypeNames[self.B.layout]),
|
||||
}
|
||||
)
|
||||
|
||||
#
|
||||
def configuration_name(self):
|
||||
''' The full procedural name indicates architecture, extended name, tile size, and layout. '''
|
||||
return self.procedural_name()
|
||||
|
||||
###################################################################################################
|
||||
#
|
||||
# Emits single instances of a CUTLASS device-wide operator
|
||||
#
|
||||
###################################################################################################
|
||||
|
||||
#
|
||||
class EmitGemmInstance:
|
||||
''' Responsible for emitting a CUTLASS template definition'''
|
||||
|
||||
def __init__(self):
|
||||
self.template = """
|
||||
// Gemm operator ${operation_name}
|
||||
using Operation_${operation_name} = cutlass::gemm::device::Gemm<
|
||||
${element_a}, ${layout_a},
|
||||
${element_b}, ${layout_b},
|
||||
${element_c}, ${layout_c},
|
||||
${element_accumulator},
|
||||
${opcode_class},
|
||||
${arch},
|
||||
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
|
||||
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
|
||||
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
${element_c},
|
||||
${epilogue_vector_length},
|
||||
${element_accumulator},
|
||||
${element_epilogue}
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle,
|
||||
${stages}
|
||||
>;
|
||||
"""
|
||||
|
||||
def emit(self, operation):
|
||||
|
||||
warp_shape = [operation.tile_description.threadblock_shape[idx] // operation.tile_description.warp_count[idx] for idx in range(3)]
|
||||
#warp_shape[2] = operation.tile_description.math_instruction.instruction_shape[2]
|
||||
warp_shape[2] = operation.tile_description.threadblock_shape[2]
|
||||
|
||||
epilogue_vector_length = int(min(operation.C.alignment * DataTypeSize[operation.C.element], 128) / DataTypeSize[operation.C.element])
|
||||
|
||||
values = {
|
||||
'operation_name': operation.procedural_name(),
|
||||
'element_a': DataTypeTag[operation.A.element],
|
||||
'layout_a': LayoutTag[operation.A.layout],
|
||||
'element_b': DataTypeTag[operation.B.element],
|
||||
'layout_b': LayoutTag[operation.B.layout],
|
||||
'element_c': DataTypeTag[operation.C.element],
|
||||
'layout_c': LayoutTag[operation.C.layout],
|
||||
'element_accumulator': DataTypeTag[operation.tile_description.math_instruction.element_accumulator],
|
||||
'opcode_class': OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
|
||||
'arch': "cutlass::arch::Sm%d" % operation.arch,
|
||||
'threadblock_shape_m': str(operation.tile_description.threadblock_shape[0]),
|
||||
'threadblock_shape_n': str(operation.tile_description.threadblock_shape[1]),
|
||||
'threadblock_shape_k': str(operation.tile_description.threadblock_shape[2]),
|
||||
'warp_shape_m': str(warp_shape[0]),
|
||||
'warp_shape_n': str(warp_shape[1]),
|
||||
'warp_shape_k': str(warp_shape[2]),
|
||||
'instruction_shape_m': str(operation.tile_description.math_instruction.instruction_shape[0]),
|
||||
'instruction_shape_n': str(operation.tile_description.math_instruction.instruction_shape[1]),
|
||||
'instruction_shape_k': str(operation.tile_description.math_instruction.instruction_shape[2]),
|
||||
'epilogue_vector_length': str(epilogue_vector_length),
|
||||
'element_epilogue': str(DataTypeTag[operation.element_epilogue]),
|
||||
'stages': str(operation.tile_description.stages)
|
||||
}
|
||||
|
||||
return SubstituteTemplate(self.template, values)
|
||||
|
||||
###################################################################################################
|
||||
|
||||
#
|
||||
class EmitGemmBatchedInstance:
|
||||
''' Responsible for emitting a CUTLASS template definition'''
|
||||
|
||||
def __init__(self):
|
||||
self.template = """
|
||||
// Gemm operator ${operation_name}
|
||||
using Operation_${operation_name} = cutlass::gemm::device::GemmBatched<
|
||||
${element_a}, ${layout_a},
|
||||
${element_b}, ${layout_b},
|
||||
${element_c}, ${layout_c},
|
||||
${element_accumulator},
|
||||
${opcode_class},
|
||||
${arch},
|
||||
cutlass::gemm::GemmShape<${threadblock_shape_m}, ${threadblock_shape_n}, ${threadblock_shape_k}>,
|
||||
cutlass::gemm::GemmShape<${warp_shape_m}, ${warp_shape_n}, ${warp_shape_k}>,
|
||||
cutlass::gemm::GemmShape<${instruction_shape_m}, ${instruction_shape_n}, ${instruction_shape_k}>,
|
||||
cutlass::epilogue::thread::LinearCombination<
|
||||
${element_c},
|
||||
${epilogue_vector_length},
|
||||
${element_accumulator},
|
||||
${element_epilogue}
|
||||
>,
|
||||
cutlass::gemm::threadblock::GemmBatchedIdentityThreadblockSwizzle,
|
||||
${stages},
|
||||
${align_a},
|
||||
${align_b}
|
||||
>;
|
||||
"""
|
||||
|
||||
def emit(self, operation):
|
||||
|
||||
warp_shape = [operation.tile_description.threadblock_shape[idx] // operation.tile_description.warp_count[idx] for idx in range(3)]
|
||||
#warp_shape[2] = operation.tile_description.math_instruction.instruction_shape[2]
|
||||
warp_shape[2] = operation.tile_description.threadblock_shape[2]
|
||||
|
||||
epilogue_vector_length = int(min(operation.C.alignment * DataTypeSize[operation.C.element], 128) / DataTypeSize[operation.C.element])
|
||||
|
||||
values = {
|
||||
'operation_name': operation.procedural_name(),
|
||||
'element_a': DataTypeTag[operation.A.element],
|
||||
'layout_a': LayoutTag[operation.A.layout],
|
||||
'element_b': DataTypeTag[operation.B.element],
|
||||
'layout_b': LayoutTag[operation.B.layout],
|
||||
'element_c': DataTypeTag[operation.C.element],
|
||||
'layout_c': LayoutTag[operation.C.layout],
|
||||
'element_accumulator': DataTypeTag[operation.tile_description.math_instruction.element_accumulator],
|
||||
'opcode_class': OpcodeClassTag[operation.tile_description.math_instruction.opcode_class],
|
||||
'arch': "cutlass::arch::Sm%d" % operation.arch,
|
||||
'threadblock_shape_m': str(operation.tile_description.threadblock_shape[0]),
|
||||
'threadblock_shape_n': str(operation.tile_description.threadblock_shape[1]),
|
||||
'threadblock_shape_k': str(operation.tile_description.threadblock_shape[2]),
|
||||
'warp_shape_m': str(warp_shape[0]),
|
||||
'warp_shape_n': str(warp_shape[1]),
|
||||
'warp_shape_k': str(warp_shape[2]),
|
||||
'instruction_shape_m': str(operation.tile_description.math_instruction.instruction_shape[0]),
|
||||
'instruction_shape_n': str(operation.tile_description.math_instruction.instruction_shape[1]),
|
||||
'instruction_shape_k': str(operation.tile_description.math_instruction.instruction_shape[2]),
|
||||
'epilogue_vector_length': str(epilogue_vector_length),
|
||||
'element_epilogue': str(DataTypeTag[operation.element_epilogue]),
|
||||
'stages': str(operation.tile_description.stages),
|
||||
'align_a': str(operation.A.alignment),
|
||||
'align_b': str(operation.B.alignment),
|
||||
}
|
||||
|
||||
return SubstituteTemplate(self.template, values)
|
||||
|
||||
###################################################################################################
|
||||
#
|
||||
# Generator functions for all layouts
|
||||
#
|
||||
###################################################################################################
|
||||
|
||||
#
|
||||
def GenerateGemmSimt(gemm_kind, manifest, tile_descriptions, min_cc):
|
||||
layouts = [
|
||||
(LayoutType.ColumnMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.ColumnMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
]
|
||||
|
||||
# for each tile configuration, emit a GEMM
|
||||
for tile in tile_descriptions:
|
||||
for layout in layouts:
|
||||
|
||||
A = TensorDescription(tile.math_instruction.element_a, layout[0], 1)
|
||||
B = TensorDescription(tile.math_instruction.element_b, layout[1], 1)
|
||||
C = TensorDescription(tile.math_instruction.element_accumulator, layout[2], 1)
|
||||
|
||||
manifest.append(GemmOperation(gemm_kind, 50, tile, A, B, C, tile.math_instruction.element_accumulator))
|
||||
|
||||
#
|
||||
def GenerateGemmTensorOp(gemm_kind, manifest, tile_descriptions, min_cc, minimum_alignment = [128,]):
|
||||
|
||||
# Canonical matrix layouts
|
||||
canonical_layouts = [
|
||||
(LayoutType.ColumnMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.ColumnMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
]
|
||||
|
||||
# Interleaved matrix layouts
|
||||
interleaved_layouts = {
|
||||
8: [
|
||||
#(LayoutType.ColumnMajorInterleaved32, LayoutType.RowMajorInterleaved32, LayoutType.ColumnMajorInterleaved32),
|
||||
(LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
],
|
||||
4: [
|
||||
#(LayoutType.ColumnMajorInterleaved64, LayoutType.RowMajorInterleaved64, LayoutType.ColumnMajorInterleaved64),
|
||||
(LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
]
|
||||
}
|
||||
|
||||
# for each tile configuration, emit a GEMM
|
||||
for align in minimum_alignment:
|
||||
for tile in tile_descriptions:
|
||||
|
||||
min_input_size = min(DataTypeSize[tile.math_instruction.element_a], DataTypeSize[tile.math_instruction.element_a])
|
||||
|
||||
# If the data type is large enough, use canonical layouts.
|
||||
if min_input_size >= 16:
|
||||
layouts = canonical_layouts
|
||||
else:
|
||||
layouts = interleaved_layouts[min_input_size]
|
||||
|
||||
for layout in layouts:
|
||||
|
||||
#
|
||||
output_types = [tile.math_instruction.element_a, tile.math_instruction.element_accumulator] \
|
||||
if DataTypeSize[tile.math_instruction.element_accumulator] == 32 \
|
||||
else [tile.math_instruction.element_accumulator,]
|
||||
|
||||
align_a = align // DataTypeSize[tile.math_instruction.element_a]
|
||||
align_b = align // DataTypeSize[tile.math_instruction.element_b]
|
||||
|
||||
|
||||
for output_type in output_types:
|
||||
|
||||
rows_per_warp = 8 // tile.warp_count[1]
|
||||
align_c = min(int(align / DataTypeSize[output_type]), tile.threadblock_shape[1] * rows_per_warp // 32)
|
||||
|
||||
A = TensorDescription(tile.math_instruction.element_a, layout[0], align_a)
|
||||
B = TensorDescription(tile.math_instruction.element_b, layout[1], align_b)
|
||||
C = TensorDescription(output_type, layout[2], max(1, align_c))
|
||||
|
||||
element_epilogue = DataType.f32 if tile.math_instruction.element_accumulator == DataType.s32 \
|
||||
else tile.math_instruction.element_accumulator
|
||||
|
||||
manifest.append(GemmOperation(gemm_kind, min_cc, tile, A, B, C, element_epilogue))
|
||||
|
||||
|
||||
#
|
||||
def GenerateGemmWmmaTensorOp(gemm_kind, manifest, tile_descriptions, min_cc, minimum_alignment = [128,]):
|
||||
|
||||
# Wmma supported matrix layouts
|
||||
layouts = [
|
||||
(LayoutType.ColumnMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.ColumnMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.ColumnMajor, LayoutType.ColumnMajor),
|
||||
(LayoutType.RowMajor, LayoutType.RowMajor, LayoutType.ColumnMajor),
|
||||
]
|
||||
|
||||
# for each tile configuration, emit a GEMM
|
||||
for align in minimum_alignment:
|
||||
for tile in tile_descriptions:
|
||||
for layout in layouts:
|
||||
|
||||
#
|
||||
output_types = [tile.math_instruction.element_a, tile.math_instruction.element_accumulator] \
|
||||
if DataTypeSize[tile.math_instruction.element_accumulator] == 32 \
|
||||
else [tile.math_instruction.element_accumulator,]
|
||||
|
||||
align_a = align // DataTypeSize[tile.math_instruction.element_a]
|
||||
align_b = align // DataTypeSize[tile.math_instruction.element_b]
|
||||
|
||||
|
||||
for output_type in output_types:
|
||||
|
||||
rows_per_warp = 8 // tile.warp_count[1]
|
||||
align_c = min(int(align / DataTypeSize[output_type]), tile.threadblock_shape[1] * rows_per_warp // 32)
|
||||
|
||||
A = TensorDescription(tile.math_instruction.element_a, layout[0], align_a)
|
||||
B = TensorDescription(tile.math_instruction.element_b, layout[1], align_b)
|
||||
C = TensorDescription(output_type, layout[2], max(1, align_c))
|
||||
|
||||
element_epilogue = DataType.f32 if tile.math_instruction.element_accumulator == DataType.s32 \
|
||||
else tile.math_instruction.element_accumulator
|
||||
|
||||
manifest.append(GemmOperation(gemm_kind, min_cc, tile, A, B, C, element_epilogue))
|
||||
|
||||
###################################################################################################
|
||||
#
|
||||
# Emitters functions for all targets
|
||||
#
|
||||
###################################################################################################
|
||||
|
||||
class EmitGemmConfigurationLibrary:
|
||||
def __init__(self, operation_path, configuration_name):
|
||||
self.configuration_name = configuration_name
|
||||
self.configuration_path = os.path.join(operation_path, "%s.cu" % configuration_name).replace('\\', '/')
|
||||
|
||||
self.instance_emitter = {
|
||||
GemmKind.Gemm: EmitGemmInstance,
|
||||
GemmKind.Batched: EmitGemmBatchedInstance
|
||||
}
|
||||
|
||||
self.gemm_kind_wrappers = {
|
||||
GemmKind.Gemm: 'GemmOperation',
|
||||
GemmKind.Batched: 'GemmBatchedOperation',
|
||||
}
|
||||
|
||||
self.wmma_guard_start = "#if defined(CUTLASS_ARCH_WMMA_SM${sm_number}_ENABLED)"
|
||||
|
||||
self.instance_template = """
|
||||
${compile_guard_start}
|
||||
manifest.append(new ${gemm_kind}<Operation_${operation_name}>("${operation_name}"));
|
||||
${compile_guard_end}
|
||||
"""
|
||||
self.header_template = """
|
||||
/*
|
||||
Generated by gemm_operation.py - Do not edit.
|
||||
*/
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
#include "cutlass/arch/wmma.h"
|
||||
#include "cutlass/cutlass.h"
|
||||
#include "cutlass/library/library.h"
|
||||
#include "cutlass/library/manifest.h"
|
||||
|
||||
#include "library_internal.h"
|
||||
#include "gemm_operation.h"
|
||||
|
||||
namespace cutlass {
|
||||
namespace library {
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
void initialize_${configuration_name}(Manifest &manifest) {
|
||||
|
||||
"""
|
||||
self.epilogue_template = """
|
||||
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace library
|
||||
} // namespace cutlass
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
"""
|
||||
|
||||
def __enter__(self):
|
||||
self.configuration_file = open(self.configuration_path, "w")
|
||||
self.configuration_file.write(SubstituteTemplate(self.header_template, {
|
||||
'configuration_name': self.configuration_name
|
||||
}))
|
||||
self.operations = []
|
||||
return self
|
||||
|
||||
def emit(self, operation):
|
||||
emitter = self.instance_emitter[operation.gemm_kind]()
|
||||
|
||||
self.operations.append(operation)
|
||||
self.configuration_file.write(emitter.emit(operation))
|
||||
self.configuration_file.write(SubstituteTemplate(self.instance_template, {
|
||||
'configuration_name': self.configuration_name,
|
||||
'operation_name': operation.procedural_name(),
|
||||
'gemm_kind': self.gemm_kind_wrappers[operation.gemm_kind],
|
||||
'compile_guard_start': SubstituteTemplate(self.wmma_guard_start, {'sm_number': str(operation.arch)}) \
|
||||
if operation.tile_description.math_instruction.opcode_class == OpcodeClass.WmmaTensorOp else "",
|
||||
'compile_guard_end': "#endif" \
|
||||
if operation.tile_description.math_instruction.opcode_class == OpcodeClass.WmmaTensorOp else ""
|
||||
}))
|
||||
|
||||
def __exit__(self, exception_type, exception_value, traceback):
|
||||
self.configuration_file.write(self.epilogue_template)
|
||||
self.configuration_file.close()
|
||||
|
||||
###################################################################################################
|
||||
###################################################################################################
|
||||
Reference in New Issue
Block a user