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

Adding more Threadblock Tiles for Mixed-input TensorOp (BF16 * S8) in cutlass_library (#1132)

Browse Source 
* Adding more tiles in the cutlass_library for mixed-input support.

* fix rebase issue

* more tiles to upcast a
This commit is contained in:
Manish Gupta
2023-10-13 11:33:15 -04:00
committed by GitHub
parent fa8dfe631f
commit 757275f279
14 changed files with 830 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/gemm/warp/gemm_mixed_input_sm80.cu
+3 -4
View File
@@ -75,7 +75,6 @@ TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_f16_i8, 128x128x64_64x64x64_
.run();
}
TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_f16_i8, 64x64x64_64x64x64_16x8x16) {
using Shape = cutlass::gemm::GemmShape<64, 64, 64>;
using InstructionShape = cutlass::gemm::GemmShape<16, 8, 16>;
@@ -140,7 +139,6 @@ TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_i8_f16, 64x64x64_64x64x64_16
.run();
}
////////////////////////////////////////////////////////////////////////////////
/// F32 <= F16 * U8 + F32 (Upcast on Operand B)
////////////////////////////////////////////////////////////////////////////////
@@ -227,6 +225,7 @@ TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_u8_f16, 128x128x64_64x64x64_
.run();
}
////////////////////////////////////////////////////////////////////////////////
/// F32 <= B16 * U8 + F32 (Upcast on Operand B)
////////////////////////////////////////////////////////////////////////////////
@@ -251,7 +250,7 @@ TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_bf16_u8, 64x64x64_64x64x64_1
}
////////////////////////////////////////////////////////////////////////////////
/// F32 <= B16 * U8 + F32 (Upcast on Operand B)
/// F32 <= U8 * BF16 + F32 (Upcast on Operand A)
////////////////////////////////////////////////////////////////////////////////
TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_u8_bf16, 64x64x64_64x64x64_16x8x16) {
using Shape = cutlass::gemm::GemmShape<64, 64, 64>;
@@ -297,7 +296,7 @@ TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_bf16_i8, 64x64x64_64x64x64_1
}
////////////////////////////////////////////////////////////////////////////////
/// F32 <= B16 * I8 + F32 (Upcast on Operand B)
/// F32 <= I8 * BF16 + F32 (Upcast on Operand A)
////////////////////////////////////////////////////////////////////////////////
TEST(SM80_warp_gemm_mixed_input_tensor_op_crosswise_i8_bf16, 64x64x64_64x64x64_16x8x16) {
using Shape = cutlass::gemm::GemmShape<64, 64, 64>;