leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

v4.4 tag release update. (#3032)

Browse Source
This commit is contained in:
Junkai-Wu
2026-02-13 23:27:58 -05:00
committed by GitHub
parent 01687cfba1
commit d4bbf728ca
140 changed files with 41622 additions and 3689 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/75_blackwell_grouped_gemm/75_blackwell_grouped_gemm.cu
+13 -3
View File
@@ -251,6 +251,8 @@ struct Options {
dim3 cluster_shape = dim3(4,2,1);
dim3 cluster_shape_fallback = dim3(2,1,1);
RasterOrderOptions raster_order = RasterOrderOptions::AlongM;
char raster_char = 'M';
int swizzle = 1;
int max_sm_count = INT_MAX;
std::string benchmark_path;
std::vector<typename ProblemShape::UnderlyingProblemShape> problem_sizes_host;
@@ -294,7 +296,6 @@ struct Options {
randomize_problems(cmd);
}
char raster_char;
cmd.get_cmd_line_argument("raster", raster_char);
if (raster_char == 'N' || raster_char == 'n') {
@@ -303,6 +304,7 @@ struct Options {
else if (raster_char == 'M' || raster_char == 'm') {
raster_order = RasterOrderOptions::AlongM;
}
cmd.get_cmd_line_argument("swizzle", swizzle, 1);
}
void randomize_problems(cutlass::CommandLine &cmd) {
@@ -378,7 +380,8 @@ struct Options {
<< " --beta=<f32> Epilogue scalar beta\n\n"
<< " --cluster_m=<int> and --cluster_n=<int> Sets the X,Y dims of the preferred cluster shape\n"
<< " --cluster_fallback_m=<int> and --cluster_fallback_n=<int> Sets the X,Y dims of the fallback cluster shape\n\n"
<< " --raster=<char> CTA Rasterization direction (N for along N, M for along M)\n\n"
<< " --raster=<char> Cluster rasterization direction (N for along N, M for along M)\n"
<< " --swizzle=<int> Cluster swizzle (swizzle up to 8 and with the nearest multiple of 2)\n\n"
<< " --iterations=<int> Number of profiling iterations to perform\n\n"
<< " --benchmark=<str> Executes a benchmark problem size\n"
<< " --max_sm_count=<int> Run kernels using only these number of SMs\n"
@@ -615,6 +618,7 @@ typename Gemm::Arguments args_from_options(Options &options, bool host_problem_s
typename Gemm::GemmKernel::TileSchedulerArguments scheduler;
scheduler.raster_order = options.raster_order;
scheduler.max_swizzle_size = options.swizzle;
if (host_problem_shapes_available) {
arguments = typename Gemm::Arguments {
@@ -685,7 +689,13 @@ int run(Options &options, bool host_problem_shapes_available = true)
std::cout << " " << options.problem_sizes_host.at(i);
std::cout << ", " << alpha_host.at(i) << ", " << beta_host.at(i) << std::endl;
}
std::cout << " Groups : " << options.groups << std::endl;
std::cout << " Groups : " << options.groups << std::endl;
std::cout << " Cluster Shape : " << options.cluster_shape.x << "x" << options.cluster_shape.y << std::endl;
std::cout << " Cluster Fallback Shape : " << options.cluster_shape_fallback.x << "x" << options.cluster_shape_fallback.y << std::endl;
std::cout << " Raster Order : Along-" << options.raster_char << std::endl;
std::cout << " Max Swizzle Size : " << options.swizzle << std::endl;
// Instantiate CUTLASS kernel depending on templates
Gemm gemm;
examples/75_blackwell_grouped_gemm/75_blackwell_grouped_gemm_block_scaled.cu
+13 -3
View File
@@ -309,6 +309,8 @@ struct Options {
dim3 cluster_shape = dim3(2,1,1);
dim3 cluster_shape_fallback = dim3(2,1,1);
RasterOrderOptions raster_order = RasterOrderOptions::AlongN;
char raster_char = 'N';
int swizzle = 1;
int max_sm_count = INT_MAX;
std::string benchmark_path;
std::vector<typename ProblemShape::UnderlyingProblemShape> problem_sizes_host;
@@ -356,7 +358,6 @@ struct Options {
randomize_problems(cmd);
}
char raster_char;
cmd.get_cmd_line_argument("raster", raster_char);
if (raster_char == 'N' || raster_char == 'n') {
@@ -365,6 +366,7 @@ struct Options {
else if (raster_char == 'M' || raster_char == 'm') {
raster_order = RasterOrderOptions::AlongM;
}
cmd.get_cmd_line_argument("swizzle", swizzle, 1);
}
void randomize_problems(cutlass::CommandLine &cmd) {
@@ -441,7 +443,8 @@ struct Options {
<< " --norm_constant=<f32> Epilogue scalar normalization constant for the output matrix\n\n"
<< " --cluster_m=<int> and --cluster_n=<int> Sets the X,Y dims of the preferred cluster shape\n"
<< " --cluster_fallback_m=<int> and --cluster_fallback_n=<int> Sets the X,Y dims of the fallback cluster shape\n\n"
<< " --raster=<char> CTA Rasterization direction (N for along N, M for along M)\n\n"
<< " --raster=<char> Cluster rasterization direction (N for along N, M for along M)\n"
<< " --swizzle=<int> Cluster swizzle (swizzle up to 8 and with the nearest multiple of 2)\n\n"
<< " --iterations=<int> Number of profiling iterations to perform\n\n"
<< " --benchmark=<str> Executes a benchmark problem size\n"
<< " --max_sm_count=<int> Run kernels using only these number of SMs\n"
@@ -722,6 +725,7 @@ typename Gemm::Arguments args_from_options(Options &options, bool host_problem_s
typename Gemm::GemmKernel::TileSchedulerArguments scheduler;
scheduler.raster_order = options.raster_order;
scheduler.max_swizzle_size = options.swizzle;
if (host_problem_shapes_available) {
arguments = typename Gemm::Arguments {
@@ -813,7 +817,13 @@ int run(Options &options, bool host_problem_shapes_available = true)
std::cout << " " << options.problem_sizes_host.at(i);
std::cout << ", " << alpha_host.at(i) << ", " << beta_host.at(i) << std::endl;
}
std::cout << " Groups : " << options.groups << std::endl;
std::cout << " Groups : " << options.groups << std::endl;
std::cout << " Cluster Shape : " << options.cluster_shape.x << "x" << options.cluster_shape.y << std::endl;
std::cout << " Cluster Fallback Shape : " << options.cluster_shape_fallback.x << "x" << options.cluster_shape_fallback.y << std::endl;
std::cout << " Raster Order : Along-" << options.raster_char << std::endl;
std::cout << " Max Swizzle Size : " << options.swizzle << std::endl;
// Instantiate CUTLASS kernel depending on templates
Gemm gemm;
examples/75_blackwell_grouped_gemm/CMakeLists.txt
+4 -4
View File
@@ -31,16 +31,16 @@
set(TEST_RANDOM --iterations=0) # Random problem sizes
set(TEST_RANDOM_LARGE_GROUP --groups=50 --iterations=0) # Random problem sizes
set(TEST_RANDOM --iterations=0 --raster=M --swizzle=4) # Random problem sizes
set(TEST_RANDOM_LARGE_GROUP --groups=50 --iterations=0 --raster=M --swizzle=2) # Random problem sizes
set(TEST_EPILOGUE --alpha=0.5 --beta=0.5 --iterations=0) # Random problem sizes
set(TEST_EPILOGUE_LARGE_GROUP --alpha=1.5 --beta=2.0 --groups=50 --iterations=0) # Random problem sizes
set(TEST_EPILOGUE_OP --beta=0.5 --iterations=1) # Random problem sizes
set(TEST_EPILOGUE_OP --beta=0.5 --iterations=1 --raster=N --swizzle=8) # Random problem sizes
set(TEST_EPILOGUE_OP_LARGE_GROUP --alpha=1.5 --groups=50 --iterations=1) # Random problem sizes
set(TEST_FIXED --m=2048 --n=5120 --k=8192 --iterations=0) # Fixed problem sizes
set(TEST_FIXED --m=2048 --n=5120 --k=8192 --iterations=0 --raster=M --swizzle=8) # Fixed problem sizes
set(TEST_FIXED_LARGE_GROUP --m=2048 --n=512 --k=512 --groups=51 --iterations=0) # Fixed problem sizes
set(TEST_SMALL --m=256 --n=128 --iterations=0) # Small problem sizes