leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 26 Packages Projects Releases Wiki Activity

Shard gemm reference templates into multiple TUs for parallel compilation (#1043)

Browse Source 
* Split apart gemm reference templates into multiple TUs for parallel compilation

* remove old files

* better balancing of ref kernels across TUs

* remove 3 new added refcheck kernels and some un-necessary fp8 library instances to reduce lib size

* remove auto fp8 kernels

* remove some redundant kernels
This commit is contained in:
Vijay Thakkar
2023-08-30 16:46:30 -04:00
committed by GitHub
parent 34fd98056b
commit e01b9b5029
18 changed files with 1498 additions and 824 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/library/src/reference/initialize_reference_operations.cu
+32 -4
View File
@@ -42,8 +42,21 @@
namespace cutlass {
namespace library {
void initialize_gemm_reference_operations(Manifest &manifest);
void initialize_gemm_fp8_reference_operations(Manifest &manifest);
// note: init methods for the same op-class may be split into multiple to parallelize compilation
void initialize_gemm_reference_operations_int4(Manifest &manifest);
void initialize_gemm_reference_operations_int8_interleaved_32(Manifest &manifest);
void initialize_gemm_reference_operations_int8_interleaved_64(Manifest &manifest);
void initialize_gemm_reference_operations_int8_canonical(Manifest &manifest);
void initialize_gemm_reference_operations_e4m3a_e4m3out(Manifest &manifest);
void initialize_gemm_reference_operations_e5m2a_e4m3out(Manifest &manifest);
void initialize_gemm_reference_operations_e4m3a_e5m2out(Manifest &manifest);
void initialize_gemm_reference_operations_e5m2a_e5m2out(Manifest &manifest);
void initialize_gemm_reference_operations_fp8in_fp16out(Manifest &manifest);
void initialize_gemm_reference_operations_fp8in_bf16out(Manifest &manifest);
void initialize_gemm_reference_operations_fp8in_fp32out(Manifest &manifest);
void initialize_gemm_reference_operations_fp32out(Manifest &manifest);
void initialize_gemm_reference_operations_fp_other(Manifest &manifest);
void initialize_conv2d_reference_operations(Manifest &manifest);
void initialize_conv3d_reference_operations(Manifest &manifest);
@@ -52,8 +65,23 @@ void initialize_conv3d_reference_operations(Manifest &manifest);
void initialize_reference_operations(Manifest &manifest) {
initialize_conv2d_reference_operations(manifest);
initialize_conv3d_reference_operations(manifest);
initialize_gemm_reference_operations(manifest);
initialize_gemm_fp8_reference_operations(manifest);
initialize_gemm_reference_operations_int4(manifest);
initialize_gemm_reference_operations_int8_interleaved_32(manifest);
initialize_gemm_reference_operations_int8_interleaved_64(manifest);
initialize_gemm_reference_operations_int8_canonical(manifest);
initialize_gemm_reference_operations_e4m3a_e4m3out(manifest);
initialize_gemm_reference_operations_e5m2a_e4m3out(manifest);
initialize_gemm_reference_operations_e4m3a_e5m2out(manifest);
initialize_gemm_reference_operations_e5m2a_e5m2out(manifest);
initialize_gemm_reference_operations_fp8in_fp16out(manifest);
initialize_gemm_reference_operations_fp8in_bf16out(manifest);
initialize_gemm_reference_operations_fp8in_fp32out(manifest);
initialize_gemm_reference_operations_fp32out(manifest);
initialize_gemm_reference_operations_fp_other(manifest);
}
///////////////////////////////////////////////////////////////////////////////////////////////////