leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

CUTLASS 2.6.1 - functional and performance enhancements to strided DGRAD, fixes, and tuning

Browse Source 
* cutlass 2.6 update

* remove debug prints

* cutlass 2.6.1 (minor update)

* Updated CHANGELOG.

* Minor edit to readme to indicate patch version.

* Minor edit to readme.

Co-authored-by:  Haicheng Wu <haichengw@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
This commit is contained in:
Manish Gupta
2021-09-03 10:26:15 -07:00
committed by GitHub
co-authored by Haicheng Wu <haichengw@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
parent a01feb93d9
commit 6c2f8f2fb8
55 changed files with 317 additions and 315 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/library/scripts/generator.py
+7 -2
View File
@@ -1312,6 +1312,7 @@ def GenerateSM80_TensorOp_16816(manifest, args):
TileDescription([256, 64, 32], 4, [4, 1, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 256, 32], 4, [1, 4, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 32], 3, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 32], 4, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 32], 5, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 64, 32], 6, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 128, 32], 6, [2, 2, 1], math_inst, min_cc, max_cc),
@@ -1392,7 +1393,7 @@ def GenerateSM80_SparseTensorOp_16832(manifest, args):
max_cc = 1024
max_cc_smem_limited = 80
alignment_constraints = [8, 4, 2]
alignment_constraints = [8]
for math_inst in math_instructions:
tile_descriptions = [
@@ -1967,6 +1968,8 @@ def GenerateSM80_TensorOp_1688(manifest, args):
TileDescription([256, 64, 16], 4, [4, 1, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 256, 16], 4, [1, 4, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 5, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 4, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 3, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 64, 16], 6, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 128, 16], 6, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 64, 16], 10, [2, 2, 1], math_inst, min_cc, max_cc),
@@ -2051,6 +2054,8 @@ def GenerateSM80_TensorOp_1688_fast_math(manifest, args):
TileDescription([256, 64, 16], 4, [4, 1, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 256, 16], 4, [1, 4, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 5, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 4, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 128, 16], 3, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([128, 64, 16], 6, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 128, 16], 6, [2, 2, 1], math_inst, min_cc, max_cc),
TileDescription([ 64, 64, 16], 10, [2, 2, 1], math_inst, min_cc, max_cc),
@@ -2100,7 +2105,7 @@ def GenerateSM80_SparseTensorOp_16816_fast_math(manifest, args):
max_cc = 1024
max_cc_smem_limited = 80
alignment_constraints = [4, 2, 1]
alignment_constraints = [4]
for math_inst in math_instructions:
tile_descriptions = [
tools/library/scripts/manifest.py
+1 -2
View File
@@ -20,7 +20,6 @@ class EmitOperationKindLibrary:
self.generated_path = generated_path
self.kind = kind
self.args = args
self.emitters = {
OperationKind.Gemm: EmitGemmConfigurationLibrary
, OperationKind.Conv2d: EmitConv2dConfigurationLibrary
@@ -347,7 +346,7 @@ class Manifest:
with interface_emitters[target](generated_path, self.operation_count, self.args) as iface_emitter:
for operation_kind, configurations in self.operations.items():
iface_emitter.emit(OperationKindNames[operation_kind])
iface_emitter.emit(OperationKindNames[operation_kind])
source_files += iface_emitter.source_files
tools/library/src/reference/gemm_reference_operation.h
+7 -13
View File
@@ -186,12 +186,6 @@ public:
GemmUniversalConfiguration const &config = *static_cast<GemmUniversalConfiguration const *>(host_workspace);
GemmUniversalArguments const &args = *static_cast<GemmUniversalArguments const *>(arguments);
ElementCompute alpha;
ElementCompute beta;
alpha = *static_cast<ElementCompute const *>(args.alpha);
beta = *static_cast<ElementCompute const *>(args.beta);
TensorRefA ref_A{static_cast<ElementA *>(const_cast<void *>(args.A)), LayoutA(int(config.lda))};
TensorRefB ref_B{static_cast<ElementB *>(const_cast<void *>(args.B)), LayoutB(int(config.ldb))};
TensorRefC ref_C{static_cast<ElementC *>(const_cast<void *>(args.C)), LayoutC(int(config.ldc))};
@@ -212,16 +206,16 @@ public:
InnerProductOp
>(
config.problem_size,
alpha,
*static_cast<ElementCompute const *>(args.alpha),
ref_A,
kTransformA,
ref_B,
kTransformB,
beta,
*static_cast<ElementCompute const *>(args.beta),
ref_C,
ref_D,
ElementAccumulator(),
config.batch_count,
((config.mode == library::GemmUniversalMode::kBatched) ? config.batch_count : 1),
args.batch_stride_A,
args.batch_stride_B,
args.batch_stride_C,
@@ -245,16 +239,16 @@ public:
InnerProductOp
>(
config.problem_size,
alpha,
*static_cast<ElementCompute const *>(args.alpha),
ref_A,
kTransformA,
ref_B,
kTransformB,
beta,
*static_cast<ElementCompute const *>(args.beta),
ref_C,
ref_D,
ElementAccumulator(),
config.batch_count,
((config.mode == library::GemmUniversalMode::kBatched) ? config.batch_count : 1),
args.batch_stride_A,
args.batch_stride_B,
args.batch_stride_C,
@@ -263,7 +257,7 @@ public:
return Status::kSuccess;
}
return Status::kErrorNotSupported;
}
};
tools/profiler/src/gemm_operation_profiler.cu
+1 -1
View File
@@ -791,7 +791,7 @@ bool GemmOperationProfiler::verify_with_reference_(
handle.set_provider(provider);
Status status = handle.gemm_universal(
library::GemmUniversalMode::kGemm,
problem_.mode,
gemm_workspace_.configuration.problem_size.m(),
gemm_workspace_.configuration.problem_size.n(),
gemm_workspace_.configuration.problem_size.k(),
tools/profiler/src/gpu_timer.h
+1
View File
@@ -29,6 +29,7 @@
#pragma once
#include <cuda_runtime.h>
#include "cutlass/cutlass.h"
namespace cutlass {
namespace profiler {
tools/profiler/src/options.cu
+3 -1
View File
@@ -425,7 +425,9 @@ void Options::Profiling::print_usage(std::ostream &out) const {
<< " Number of ms to sleep between profiling periods (ms).\n\n"
<< " --profiling-enabled=<bool> "
<< " If true, profiling is actually conducted.\n\n";
<< " If true, profiling is actually conducted.\n\n"
;
}
void Options::Profiling::print_options(std::ostream &out, int indent) const {
tools/util/include/cutlass/util/command_line.h
+2
View File
@@ -32,6 +32,8 @@
#include <cuda_runtime.h>
#include "cutlass/cutlass.h"
namespace cutlass {
/******************************************************************************