leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 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
examples/13_two_tensor_op_fusion/device/b2b_gemm.h
-8
View File
@@ -390,14 +390,6 @@ public:
if (result != cudaSuccess) {
return Status::kErrorInternal;
}
result = cudaFuncSetAttribute(
Kernel<B2bGemmKernel>,
cudaFuncAttributePreferredSharedMemoryCarveout, 100);
if (result != cudaSuccess) {
return Status::kErrorInternal;
}
}
cutlass::Kernel<B2bGemmKernel><<<grid, block, smem_size, stream>>>(params_);
examples/13_two_tensor_op_fusion/device/b2b_implicit_gemm_convolution.h
-8
View File
@@ -197,14 +197,6 @@ public:
if (result != cudaSuccess) {
return Status::kErrorInternal;
}
result = cudaFuncSetAttribute(
cutlass::Kernel<B2bImplicitGemmKernel>,
cudaFuncAttributePreferredSharedMemoryCarveout, 100);
if (result != cudaSuccess) {
return Status::kErrorInternal;
}
}
return Status::kSuccess;
examples/13_two_tensor_op_fusion/threadblock/b2b_mma_multistage.h
+9 -21
View File
@@ -395,10 +395,8 @@ public:
for (int stage = 0; stage < Base::kStages - 1;
++stage, --gemm_k_iterations_0) {
if (gemm_k_iterations_0 == 0) {
iterator_A0.clear_mask();
iterator_B0.clear_mask();
}
iterator_A0.clear_mask(gemm_k_iterations_0 == 0);
iterator_B0.clear_mask(gemm_k_iterations_0 == 0);
iterator_A0.set_iteration_index(0);
this->smem_iterator_A0_.set_iteration_index(0);
@@ -490,10 +488,8 @@ public:
++this->warp_tile_iterator_A0_;
++this->warp_tile_iterator_B0_;
if (gemm_k_iterations_0 == 0) {
iterator_A0.clear_mask();
iterator_B0.clear_mask();
}
iterator_A0.clear_mask(gemm_k_iterations_0 == 0);
iterator_B0.clear_mask(gemm_k_iterations_0 == 0);
int smem_write_stage_idx = Base::kStages - 1;
int smem_read_stage_idx = 0;
@@ -601,10 +597,8 @@ public:
}
--gemm_k_iterations_0;
if (gemm_k_iterations_0 == 0) {
iterator_A0.clear_mask();
iterator_B0.clear_mask();
}
iterator_A0.clear_mask(gemm_k_iterations_0 == 0);
iterator_B0.clear_mask(gemm_k_iterations_0 == 0);
}
// Do any conversions feeding the first stage at the end of the loop so
@@ -634,9 +628,7 @@ public:
for (int stage = 0; stage < Base::kStages - 1;
++stage, --gemm_k_iterations_1) {
if (gemm_k_iterations_1 == 0) {
iterator_B1.clear_mask();
}
iterator_B1.clear_mask(gemm_k_iterations_1 == 0);
iterator_B1.set_iteration_index(0);
this->smem_iterator_B1_.set_iteration_index(0);
@@ -694,9 +686,7 @@ public:
++warp_tile_iterator_A1_;
++this->warp_tile_iterator_B1_;
if (gemm_k_iterations_1 == 0) {
iterator_B1.clear_mask();
}
iterator_B1.clear_mask(gemm_k_iterations_1 == 0);
smem_write_stage_idx = Base::kStages - 1;
smem_read_stage_idx = 0;
@@ -793,9 +783,7 @@ public:
++smem_read_stage_idx;
}
if (gemm_k_iterations_1 == 1) {
iterator_B1.clear_mask();
}
iterator_B1.clear_mask(gemm_k_iterations_1 == 1);
}
// Do any conversions feeding the first stage at the end of the loop so