leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

CUTLASS 3.2 (#1024)

Browse Source 
* CUTLASS 3.2
This commit is contained in:
ANIKET SHIVAM
2023-08-07 14:50:32 -10:00
committed by GitHub
parent a0d787b746
commit 4575443d44
392 changed files with 47559 additions and 7940 deletions
Download Patch File Download Diff File Expand all files Collapse all files

150
test/python/gemm/gemm_f16_sm80.py
View File

@@ -37,82 +37,16 @@ Low-level functionality tests for GEMM with F16 operands on SM80
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 80
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.f16)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
threadblock_shape, warp_count, stages, opclass, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param swizzle: threadblock swizzling functor
"""
cluster_shape = [1, 1, 1]
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.f16
element_B = cutlass.DataType.f16
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator,
kernel_cc=cc)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.warp_count = warp_count
td.cluster_shape = cluster_shape
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages, opclass=binding_opclass(opclass))
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM80 tests.')
class GemmF16Sm80(unittest.TestCase):
"""
@@ -128,40 +62,64 @@ class GemmF16Sm80StreamK(unittest.TestCase):
"""
pass
add_test_specialized = partial(add_test_gemm, element=cutlass.DataType.f16, cc=cc, cluster_shape=[1, 1, 1])
# Tests using TensorOp
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop(GemmF16Sm80, LayoutCombination.NNN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.NNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.NTN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.NTT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TTN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TTT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [64, 128, 32], [1, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 64, 32], [2, 1, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [64, 64, 64], [1, 1, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [4, 4, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [4, 4, 8], cutlass.DataType.f16, cutlass.DataType.f16, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f16, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [64, 64, 64], [1, 1, 1], 5)
add_test_tensorop(GemmF16Sm80, LayoutCombination.TNT, [2, 2, 2], cutlass.DataType.f16, cutlass.DataType.f16, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.NNN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.NNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.NTN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.NTT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TTN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TTT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 128, 32], warp_count=[1, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 64, 32], warp_count=[2, 1, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 64], warp_count=[1, 1, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[4, 4, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[4, 4, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 64], warp_count=[1, 1, 1], stages=5)
add_test_tensorop(cls=GemmF16Sm80, layouts=LayoutCombination.TNT, alignments=[2, 2, 2], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
# Tests using SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmF16Sm80, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(GemmF16Sm80, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [64, 128, 8], [1, 2, 1], 2)
add_test_simt(GemmF16Sm80, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [128, 64, 8], [2, 1, 1], 2)
add_test_simt(GemmF16Sm80, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [64, 64, 8], [1, 1, 1], 2)
add_test_simt(GemmF16Sm80, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f16, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(cls=GemmF16Sm80, layouts=LayoutCombination.NNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
add_test_simt(cls=GemmF16Sm80, layouts=LayoutCombination.TNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 128, 8], warp_count=[1, 2, 1], stages=2)
add_test_simt(cls=GemmF16Sm80, layouts=LayoutCombination.NTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 64, 8], warp_count=[2, 1, 1], stages=2)
add_test_simt(cls=GemmF16Sm80, layouts=LayoutCombination.TTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 8], warp_count=[1, 1, 1], stages=2)
add_test_simt(cls=GemmF16Sm80, layouts=LayoutCombination.NNT, alignments=[1, 1, 1], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
# Stream K tests
add_test_streamk = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(GemmF16Sm80StreamK, LayoutCombination.NNN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_streamk(GemmF16Sm80StreamK, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [64, 64, 64], [1, 1, 1], 5)
add_test_streamk = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(cls=GemmF16Sm80StreamK, layouts=LayoutCombination.NNN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_streamk(cls=GemmF16Sm80StreamK, layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 64], warp_count=[1, 1, 1], stages=5)
if __name__ == '__main__':
unittest.main()

183
test/python/gemm/gemm_f16_sm90.py
View File

@@ -37,87 +37,16 @@ Low-level functionality tests for GEMM with F16 operands on SM90
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 90
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.f16)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
cluster_shape, threadblock_shape, stages, opclass,
kernel_schedule=cutlass.KernelScheduleType.ScheduleAuto,
swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param cluster_shape: dimensions of threadblock cluster
:type cluster_shape: list or tuple
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param kernel_schedule: kernel schedule type
:type kernel_schedule: cutlass.KernelScheduleType
:param swizzle: threadblock swizzling functor
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.f16
element_B = cutlass.DataType.f16
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.cluster_shape = cluster_shape
td.kernel_schedule = kernel_schedule
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages,
opclass=binding_opclass(opclass), kernel_schedule=kernel_schedule)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM90 tests.')
class GemmF16Sm90(unittest.TestCase):
"""
@@ -126,47 +55,85 @@ class GemmF16Sm90(unittest.TestCase):
pass
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_specialized = partial(add_test_gemm, cls=GemmF16Sm90, element=cutlass.DataType.f16,
warp_count=None, compilation_modes=['nvcc'])
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
# Tests with 1x1x1 clusters
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], 3)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [4, 4, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [4, 4, 8], cutlass.DataType.f16, cutlass.DataType.f16, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f16, [1, 1, 1], [128, 128, 32], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [64, 64, 64], 5)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNT, [2, 2, 2], cutlass.DataType.f16, cutlass.DataType.f16, [1, 1, 1], [128, 128, 32], None)
add_test_unit_cluster = partial(add_test_tensorop, cluster_shape=[1, 1, 1])
add_test_unit_cluster(layouts=LayoutCombination.NNN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=3)
add_test_unit_cluster(layouts=LayoutCombination.NNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.NTN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.NTT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.TNN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.TNT, alignments=[4, 4, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.TNT, alignments=[4, 4, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], stages=None)
add_test_unit_cluster(layouts=LayoutCombination.TNT, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 64], stages=5)
add_test_unit_cluster(layouts=LayoutCombination.TNT, alignments=[2, 2, 2], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 32], stages=None)
# Tests with different cluster shapes
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 8], cutlass.DataType.f16, cutlass.DataType.f16, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TNN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.NNN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [2, 2, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [1, 4, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [2, 4, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [4, 1, 1], [64, 128, 64], None)
add_test_tensorop(GemmF16Sm90, LayoutCombination.TTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, [4, 2, 1], [64, 128, 64], None)
add_test_cluster_shape = partial(add_test_tensorop, threadblock_shape=[64, 128, 64], stages=None)
add_test_cluster_shape(layouts=LayoutCombination.TTN, alignments=[8, 8, 8], element_output=cutlass.DataType.f16,
element_accumulator=cutlass.DataType.f16, cluster_shape=[2, 2, 1])
add_test_cluster_shape(layouts=LayoutCombination.TNN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[2, 2, 1])
add_test_cluster_shape(layouts=LayoutCombination.NTN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[2, 2, 1])
add_test_cluster_shape(layouts=LayoutCombination.NNN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[2, 2, 1])
add_test_cluster_shape(layouts=LayoutCombination.TTN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[1, 4, 1])
add_test_cluster_shape(layouts=LayoutCombination.TTN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[2, 4, 1])
add_test_cluster_shape(layouts=LayoutCombination.TTN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[4, 1, 1])
add_test_cluster_shape(layouts=LayoutCombination.TTN, alignments=[8, 8, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, cluster_shape=[4, 2, 1])
# Tests for different schedule modes
add_test_schedule = partial(add_test, GemmF16Sm90, LayoutCombination.TTN, [8, 8, 4], cutlass.DataType.f32, cutlass.DataType.f32, opclass=cutlass.OpcodeClass.TensorOp)
add_test_schedule([1, 1, 1], [128, 128, 64], None, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong)
add_test_schedule([1, 1, 1], [128, 128, 64], None, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative)
add_test_schedule([2, 1, 1], [128, 128, 64], None, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong)
add_test_schedule([2, 1, 1], [128, 128, 64], None, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative)
add_test_schedule([2, 1, 1], [256, 128, 64], None, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative)
add_test_schedule([2, 1, 1], [128, 128, 64], 5, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong)
add_test_schedule([2, 1, 1], [128, 128, 64], 5, kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative)
add_test_schedule = partial(add_test_specialized, layouts=LayoutCombination.TTN, alignments=[8, 8, 4],
element_output=cutlass.DataType.f32, element_accumulator=cutlass.DataType.f32,
opclass=cutlass.OpcodeClass.TensorOp, threadblock_shape=[128, 128, 64], stages=None)
add_test_schedule(
cluster_shape=[1, 1, 1],
kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong,
epilogue_schedule=cutlass.EpilogueScheduleType.TmaWarpSpecialized
)
add_test_schedule(
cluster_shape=[1, 1, 1],
kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative,
epilogue_schedule=cutlass.EpilogueScheduleType.TmaWarpSpecializedCooperative
)
add_test_schedule(
cluster_shape=[2, 1, 1],
kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong,
epilogue_schedule=cutlass.EpilogueScheduleType.TmaWarpSpecialized
)
add_test_schedule(
cluster_shape=[2, 1, 1],
kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedCooperative,
epilogue_schedule=cutlass.EpilogueScheduleType.TmaWarpSpecializedCooperative
)
# Tests using SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmF16Sm90, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 128, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [64, 128, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [128, 64, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f32, [1, 1, 1], [64, 64, 8], 2)
add_test_simt(GemmF16Sm90, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.f16, cutlass.DataType.f16, [1, 1, 1], [128, 128, 8], 2)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt, alignments=[1, 1, 1], cluster_shape=[1, 1, 1], stages=2)
add_test_simt(layouts=LayoutCombination.NNN, element_output=cutlass.DataType.f16, element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 8])
add_test_simt(layouts=LayoutCombination.TNN, element_output=cutlass.DataType.f16, element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 128, 8])
add_test_simt(layouts=LayoutCombination.NTN, element_output=cutlass.DataType.f16, element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 64, 8])
add_test_simt(layouts=LayoutCombination.TTN, element_output=cutlass.DataType.f16, element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 8])
add_test_simt(layouts=LayoutCombination.NNT, element_output=cutlass.DataType.f16, element_accumulator=cutlass.DataType.f16, threadblock_shape=[128, 128, 8])
if __name__ == '__main__':

113
test/python/gemm/gemm_f32_sm80.py
View File

@@ -37,83 +37,16 @@ Low-level functionality tests for GEMM with F32 operands on SM80
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 80
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.f32)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
threadblock_shape, warp_count, stages, opclass, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param swizzle: threadblock swizzling functor
"""
cluster_shape = [1, 1, 1]
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.f32
element_B = cutlass.DataType.f32
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator,
kernel_cc=cc)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.warp_count = warp_count
td.cluster_shape = cluster_shape
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages, opclass=binding_opclass(opclass))
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM80 tests.')
class GemmF32Sm80(unittest.TestCase):
"""
@@ -130,25 +63,37 @@ class GemmF32Sm80StreamK(unittest.TestCase):
pass
add_test_specialized = partial(add_test_gemm, element=cutlass.DataType.f32, cc=cc, cluster_shape=[1, 1, 1])
# Tests using TensorOp
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop(GemmF32Sm80, LayoutCombination.NNN, [4, 4, 4], cutlass.DataType.f32, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF32Sm80, LayoutCombination.NNT, [4, 4, 4], cutlass.DataType.f32, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_tensorop(GemmF32Sm80, LayoutCombination.NTN, [4, 4, 4], cutlass.DataType.f32, cutlass.DataType.f32, [64, 128, 32], [1, 2, 1], 3)
add_test_tensorop(GemmF32Sm80, LayoutCombination.NTN, [4, 4, 4], cutlass.DataType.f32, cutlass.DataType.f32, [64, 64, 32], [1, 1, 1], 4)
add_test_tensorop(cls=GemmF32Sm80, layouts=LayoutCombination.NNN, alignments=[4, 4, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF32Sm80, layouts=LayoutCombination.NNT, alignments=[4, 4, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
add_test_tensorop(cls=GemmF32Sm80, layouts=LayoutCombination.NTN, alignments=[4, 4, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 128, 32], warp_count=[1, 2, 1], stages=3)
add_test_tensorop(cls=GemmF32Sm80, layouts=LayoutCombination.NTN, alignments=[4, 4, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 32], warp_count=[1, 1, 1], stages=4)
# Tests using SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmF32Sm80, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f32, cutlass.DataType.f32, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(GemmF32Sm80, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.f32, cutlass.DataType.f32, [64, 128, 8], [1, 2, 1], 2)
add_test_simt(GemmF32Sm80, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.f32, cutlass.DataType.f32, [128, 64, 8], [2, 1, 1], 2)
add_test_simt(GemmF32Sm80, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.f32, cutlass.DataType.f32, [64, 64, 8], [1, 1, 1], 2)
add_test_simt(GemmF32Sm80, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.f32, cutlass.DataType.f32, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(cls=GemmF32Sm80, layouts=LayoutCombination.NNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
add_test_simt(cls=GemmF32Sm80, layouts=LayoutCombination.TNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 128, 8], warp_count=[1, 2, 1], stages=2)
add_test_simt(cls=GemmF32Sm80, layouts=LayoutCombination.NTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 64, 8], warp_count=[2, 1, 1], stages=2)
add_test_simt(cls=GemmF32Sm80, layouts=LayoutCombination.TTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[ 64, 64, 8], warp_count=[1, 1, 1], stages=2)
add_test_simt(cls=GemmF32Sm80, layouts=LayoutCombination.NNT, alignments=[1, 1, 1], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
# Stream K tests
add_test_streamk = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(GemmF32Sm80StreamK, LayoutCombination.TTN, [4, 4, 4], cutlass.DataType.f32, cutlass.DataType.f32, [128, 128, 32], [2, 2, 1], 3)
add_test_streamk = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(cls=GemmF32Sm80StreamK, layouts=LayoutCombination.TTN, alignments=[4, 4, 4], element_output=cutlass.DataType.f32,
element_accumulator=cutlass.DataType.f32, threadblock_shape=[128, 128, 32], warp_count=[2, 2, 1], stages=3)
if __name__ == '__main__':

113
test/python/gemm/gemm_f64_sm80.py
View File

@@ -37,84 +37,16 @@ Low-level functionality tests for GEMM with F64 operands on SM80
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 80
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.f64)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
threadblock_shape, warp_count, stages, opclass, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param swizzle: threadblock swizzling functor
"""
cluster_shape = [1, 1, 1]
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.f64
element_B = cutlass.DataType.f64
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator,
kernel_cc=cc)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.warp_count = warp_count
td.cluster_shape = cluster_shape
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages, opclass=binding_opclass(opclass))
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM80 tests.')
class GemmF64Sm80(unittest.TestCase):
"""
@@ -131,25 +63,36 @@ class GemmF64Sm80StreamK(unittest.TestCase):
pass
# Tests using TensorOp
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_specialized = partial(add_test_gemm, element=cutlass.DataType.f64, cc=cc, cluster_shape=[1, 1, 1])
add_test_tensorop(GemmF64Sm80, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [128, 128, 16], [4, 2, 1], 3)
add_test_tensorop(GemmF64Sm80, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [64, 64, 16], [2, 2, 1], 4)
add_test_tensorop(GemmF64Sm80, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [32, 32, 16], [2, 1, 1], 5)
# Tests using TensorOp
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop(cls=GemmF64Sm80, layouts=LayoutCombination.NNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[128, 128, 16], warp_count=[4, 2, 1], stages=3)
add_test_tensorop(cls=GemmF64Sm80, layouts=LayoutCombination.NTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[ 64, 64, 16], warp_count=[2, 2, 1], stages=4)
add_test_tensorop(cls=GemmF64Sm80, layouts=LayoutCombination.TTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[ 32, 32, 16], warp_count=[2, 1, 1], stages=5)
# Tests using SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmF64Sm80, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(GemmF64Sm80, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [64, 128, 8], [1, 2, 1], 2)
add_test_simt(GemmF64Sm80, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [128, 64, 8], [2, 1, 1], 2)
add_test_simt(GemmF64Sm80, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [64, 64, 8], [1, 1, 1], 2)
add_test_simt(GemmF64Sm80, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(cls=GemmF64Sm80, layouts=LayoutCombination.NNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
add_test_simt(cls=GemmF64Sm80, layouts=LayoutCombination.TNN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[ 64, 128, 8], warp_count=[1, 2, 1], stages=2)
add_test_simt(cls=GemmF64Sm80, layouts=LayoutCombination.NTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[128, 64, 8], warp_count=[2, 1, 1], stages=2)
add_test_simt(cls=GemmF64Sm80, layouts=LayoutCombination.TTN, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[ 64, 64, 8], warp_count=[1, 1, 1], stages=2)
add_test_simt(cls=GemmF64Sm80, layouts=LayoutCombination.NNT, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
# Stream K tests
add_test_streamk = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(GemmF64Sm80StreamK, LayoutCombination.NTT, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [128, 128, 16], [4, 2, 1], 3)
add_test_streamk = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(cls=GemmF64Sm80StreamK, layouts=LayoutCombination.NTT, alignments=[1, 1, 1], element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, threadblock_shape=[128, 128, 16], warp_count=[4, 2, 1], stages=3)
if __name__ == '__main__':

95
test/python/gemm/gemm_f64_sm90.py
View File

@@ -37,90 +37,16 @@ Low-level functionality tests for GEMM with F64 operands on SM90
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 90
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.f64)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
cluster_shape, threadblock_shape, stages, opclass, persistent=False, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param cluster_shape: dimensions of threadblock cluster
:type cluster_shape: list or tuple
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param persistent: whether this is a persistent warp-specialized kernel
:type persistent: bool
:param swizzle: threadblock swizzling functor
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.f64
element_B = cutlass.DataType.f64
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.cluster_shape = cluster_shape
td.persistent = persistent
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
if persistent:
suffix = "_persistent"
else:
suffix = ""
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages,
opclass=binding_opclass(opclass), suffix=suffix)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM90 tests.')
class GemmF64Sm90(unittest.TestCase):
"""
@@ -129,13 +55,14 @@ class GemmF64Sm90(unittest.TestCase):
pass
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_specialized = partial(add_test_gemm, cls=GemmF64Sm90, alignments=[1, 1, 1], cluster_shape=[1, 1, 1],
element=cutlass.DataType.f64, element_output=cutlass.DataType.f64,
element_accumulator=cutlass.DataType.f64, compilation_modes=['nvcc'])
add_test_tensorop(GemmF64Sm90, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [1, 1, 1], [128, 128, 32], 3)
add_test_tensorop(GemmF64Sm90, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [1, 1, 1], [128, 128, 32], 3)
add_test_simt(GemmF64Sm90, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [1, 1, 1], [128, 128, 8], 2)
add_test_simt(GemmF64Sm90, LayoutCombination.TTT, [1, 1, 1], cutlass.DataType.f64, cutlass.DataType.f64, [1, 1, 1], [64, 128, 8], 2)
add_test_specialized(opclass=cutlass.OpcodeClass.TensorOp, layouts=LayoutCombination.NNT, threadblock_shape=[128, 128, 32], stages=3)
add_test_specialized(opclass=cutlass.OpcodeClass.TensorOp, layouts=LayoutCombination.TNN, threadblock_shape=[128, 128, 32], stages=3)
add_test_specialized( opclass=cutlass.OpcodeClass.Simt, layouts=LayoutCombination.NNN, threadblock_shape=[128, 128, 8], stages=2)
add_test_specialized( opclass=cutlass.OpcodeClass.Simt, layouts=LayoutCombination.TTT, threadblock_shape=[ 64, 128, 8], stages=2)
if __name__ == '__main__':

113
test/python/gemm/gemm_s8_sm80.py
View File

@@ -37,84 +37,16 @@ Low-level functionality tests for GEMM with S8 operands on SM80
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 80
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.s8)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
threadblock_shape, warp_count, stages, opclass, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param swizzle: threadblock swizzling functor
"""
cluster_shape = [1, 1, 1]
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.s8
element_B = cutlass.DataType.s8
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator,
kernel_cc=cc)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.warp_count = warp_count
td.cluster_shape = cluster_shape
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages, opclass=binding_opclass(opclass))
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM80 tests.')
class GemmS8Sm80(unittest.TestCase):
"""
@@ -131,25 +63,36 @@ class GemmS8Sm80StreamK(unittest.TestCase):
pass
# Tests using TensorOp
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_specialized = partial(add_test_gemm, element=cutlass.DataType.s8, cc=cc, cluster_shape=[1, 1, 1])
add_test_tensorop(GemmS8Sm80, LayoutCombination.TNN, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [256, 128, 64], [4, 2, 1], 3)
add_test_tensorop(GemmS8Sm80, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [128, 256, 64], [2, 4, 1], 3)
add_test_tensorop(GemmS8Sm80, LayoutCombination.TNN, [16, 16, 4], cutlass.DataType.s32, cutlass.DataType.s32, [64, 64, 64], [1, 1, 1], 4)
# Tests using TensorOp
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
add_test_tensorop(cls=GemmS8Sm80, layouts=LayoutCombination.TNN, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[256, 128, 64], warp_count=[4, 2, 1], stages=3)
add_test_tensorop(cls=GemmS8Sm80, layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[128, 256, 64], warp_count=[2, 4, 1], stages=3)
add_test_tensorop(cls=GemmS8Sm80, layouts=LayoutCombination.TNN, alignments=[16, 16, 4], element_output=cutlass.DataType.s32,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[ 64, 64, 64], warp_count=[1, 1, 1], stages=4)
# Tests using SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmS8Sm80, LayoutCombination.NNN, [1, 1, 1], cutlass.DataType.s8, cutlass.DataType.s32, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(GemmS8Sm80, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.s8, cutlass.DataType.s32, [64, 128, 8], [1, 2, 1], 2)
add_test_simt(GemmS8Sm80, LayoutCombination.NTN, [1, 1, 1], cutlass.DataType.s8, cutlass.DataType.s32, [128, 64, 8], [2, 1, 1], 2)
add_test_simt(GemmS8Sm80, LayoutCombination.TTN, [1, 1, 1], cutlass.DataType.s32, cutlass.DataType.s32, [64, 64, 8], [1, 1, 1], 2)
add_test_simt(GemmS8Sm80, LayoutCombination.NNT, [1, 1, 1], cutlass.DataType.s32, cutlass.DataType.s32, [128, 128, 8], [2, 2, 1], 2)
add_test_simt(cls=GemmS8Sm80, layouts=LayoutCombination.NNN, alignments=[1, 1, 1], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
add_test_simt(cls=GemmS8Sm80, layouts=LayoutCombination.TNN, alignments=[1, 1, 1], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[ 64, 128, 8], warp_count=[1, 2, 1], stages=2)
add_test_simt(cls=GemmS8Sm80, layouts=LayoutCombination.NTN, alignments=[1, 1, 1], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[128, 64, 8], warp_count=[2, 1, 1], stages=2)
add_test_simt(cls=GemmS8Sm80, layouts=LayoutCombination.TTN, alignments=[1, 1, 1], element_output=cutlass.DataType.s32,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[ 64, 64, 8], warp_count=[1, 1, 1], stages=2)
add_test_simt(cls=GemmS8Sm80, layouts=LayoutCombination.NNT, alignments=[1, 1, 1], element_output=cutlass.DataType.s32,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[128, 128, 8], warp_count=[2, 2, 1], stages=2)
# Stream K tests
add_test_streamk = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(GemmS8Sm80StreamK, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [128, 256, 64], [2, 4, 1], 3)
add_test_streamk = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp, swizzle=cutlass.swizzle.ThreadblockSwizzleStreamK)
add_test_streamk(cls=GemmS8Sm80StreamK, layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, threadblock_shape=[128, 256, 64], warp_count=[2, 4, 1], stages=3)
if __name__ == '__main__':

122
test/python/gemm/gemm_s8_sm90.py
View File

@@ -37,90 +37,16 @@ Low-level functionality tests for GEMM with S8 operands on SM90
from functools import partial
import cutlass
from cutlass.utils.datatypes import binding_opclass, binding_type
from cutlass.backend.test.gemm_testbed import test_all_gemm
import logging
import unittest
from cutlass.backend.test.utils import LayoutCombination, get_name
from cutlass.backend.test.utils import LayoutCombination, add_test_gemm
from cutlass.backend.utils.device import device_cc
cutlass.set_log_level(logging.WARNING)
cc = 90
# Partial specialziation for naming tests
bound_type = binding_type(cutlass.DataType.s8)
name_fn = partial(get_name, element_a=bound_type, element_b=bound_type, arch=cc)
def add_test(cls, layouts, alignments, element_output, element_accumulator,
cluster_shape, threadblock_shape, stages, opclass, persistent=False, swizzle=None):
"""
Create a test-running function with the given specification and set it as a method of `cls`.
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: layouts of A, B, and C operands
:type layouts: list or tuple
:param alignments: alingments of A, B, and C operands
:type alignments: list or tuple
:param element_output: data type of the output element
:type element_output: cutlass.DataType
:param element_accumulator: data type used in accumulation
:type element_accumulator: cutlass.DataType
:param cluster_shape: dimensions of threadblock cluster
:type cluster_shape: list or tuple
:param threadblock_shape: dimensions of threadblock tiles
:type threadblock_shape: list or tuple
:param warp_count: warps to be launched per threadblock dimension
:type warp_count: list or tuple
:param stages: number of pipeline stages to use in the kernel
:type stages: int
:param opclass: class of operation being performed (e.g., SIMT, Tensor Core)
:type opclass: cutlass.OpClass
:param persistent: whether this is a persistent warp-specialized kernel
:type persistent: bool
:param swizzle: threadblock swizzling functor
"""
def run(self):
"""
Dynamically-generated function that constructs a GEMM operation and verifies it against
multiple test cases.
"""
element_A = cutlass.DataType.s8
element_B = cutlass.DataType.s8
layout_A, layout_B, layout_C = layouts
alignment_A, alignment_B, alignment_C = alignments
plan = cutlass.op.Gemm(element_A=element_A, element_B=element_B,
element_C=element_output, element_D=element_output,
layout_A=layout_A, layout_B=layout_B, layout_C=layout_C,
element_accumulator=element_accumulator)
plan.opclass = opclass
if swizzle is not None:
plan.swizzling_functor = swizzle
td = plan.tile_descriptions()[0]
td.threadblock_shape = threadblock_shape
td.stages = stages
td.cluster_shape = cluster_shape
td.persistent = persistent
op = plan.construct(tile_description=td, alignment_A=alignment_A, alignment_B=alignment_B, alignment_C=alignment_C)
self.assertTrue(test_all_gemm(op, 'universal'))
if persistent:
suffix = "_persistent"
else:
suffix = ""
element_epilogue = element_accumulator
name = name_fn(layouts, alignments, binding_type(element_output), binding_type(element_accumulator),
binding_type(element_epilogue), cluster_shape, threadblock_shape, stages,
opclass=binding_opclass(opclass), suffix=suffix)
setattr(cls, name, run)
return run
@unittest.skipIf(device_cc() < cc, 'Device compute capability is insufficient for SM90 tests.')
class GemmS8Sm90(unittest.TestCase):
"""
@@ -129,26 +55,40 @@ class GemmS8Sm90(unittest.TestCase):
pass
add_test_tensorop = partial(add_test, opclass=cutlass.OpcodeClass.TensorOp)
add_test_specialized = partial(add_test_gemm, cls=GemmS8Sm90, element=cutlass.DataType.s8, compilation_modes=['nvcc'])
add_test_tensorop = partial(add_test_specialized, opclass=cutlass.OpcodeClass.TensorOp)
# Tests with 1x1x1 clusters
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNN, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [128, 128, 128], 3)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 8], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [64, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [128, 64, 32], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [4, 4, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [128, 128, 128], None)
add_test_tensorop(layouts=LayoutCombination.TNN, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[128, 128, 128], stages=3)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[128, 128, 128], stages=None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 8], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[128, 128, 128], stages=None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[64, 128, 128], stages=None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[128, 64, 32], stages=None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[ 4, 4, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[128, 128, 128], stages=None)
# Tests with different cluster shapes
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [2, 2, 1], [128, 128, 128], None)
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [1, 4, 1], [128, 128, 128], None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[2, 2, 1], threadblock_shape=[128, 128, 128], stages=None)
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 4, 1], threadblock_shape=[128, 128, 128], stages=None)
# Tests with persistent warp-specialized threadblocks
add_test_tensorop(GemmS8Sm90, LayoutCombination.TNT, [16, 16, 16], cutlass.DataType.s8, cutlass.DataType.s32, [2, 1, 1], [128, 128, 128], None, persistent=True)
# Tests with warp-specialized ping-pong schedule
add_test_tensorop(layouts=LayoutCombination.TNT, alignments=[16, 16, 16], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[2, 1, 1], threadblock_shape=[128, 128, 128], stages=None,
kernel_schedule=cutlass.KernelScheduleType.TmaWarpSpecializedPingpong,
epilogue_schedule=cutlass.EpilogueScheduleType.TmaWarpSpecialized)
# Tests for SIMT
add_test_simt = partial(add_test, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(GemmS8Sm90, LayoutCombination.TNN, [1, 1, 1], cutlass.DataType.s8, cutlass.DataType.s32, [1, 1, 1], [64, 32, 8], 2)
add_test_simt = partial(add_test_specialized, opclass=cutlass.OpcodeClass.Simt)
add_test_simt(layouts=LayoutCombination.TNN, alignments=[1, 1, 1], element_output=cutlass.DataType.s8,
element_accumulator=cutlass.DataType.s32, cluster_shape=[1, 1, 1], threadblock_shape=[64, 32, 8], stages=2)
if __name__ == '__main__':