Skip to content

Commit ca2ee0e

Browse files
jefyang1illsilin
jefyang1
and
illsilin
authored
Fix test_gemm_multiply_multiply_wp_xdl_fp8 on gfx950 (#3191)
Co-authored-by: Illia Silin <[email protected]>
1 parent 8d50001 commit ca2ee0e

File tree

3 files changed

+27
-1
lines changed

3 files changed

+27
-1
lines changed

‎library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn.hpp‎

Lines changed: 14 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -147,11 +147,24 @@ using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_
147147
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 256, 512, 16, 16, 16, 16, 4, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
148148
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 256, 512, 16, 16, 16, 16, 2, 4, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<32, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
149149
// N 512
150-
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 16, 16, 4, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>,
151150
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 32, 512, 256, 16, 16, 16, 16, 2, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
152151
// clang-format on
153152
>;
154153

154+
// separate instances from p3 instances that may not work on gfx950
155+
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
156+
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances_part2 =
157+
std::tuple<
158+
// clang-format off
159+
//##########################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
160+
//##########################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| Pipeline| Pipeline|
161+
//##########################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NWaveNPerXdl| Scheduler| Verision|
162+
//##########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
163+
// known issue: v1 of this instance with kbatch > 1 failed verification on test_gemm_multiply_multiply_wp_xdl_fp8
164+
DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 512, 256, 16, 16, 16, 16, 4, 8, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlkGemmPipeVer, F8>
165+
// clang-format on
166+
>;
167+
155168
template <BlockGemmPipelineVersion BlkGemmPipeVer, GemmSpecialization GemmSpec>
156169
using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p4_instances =
157170
std::tuple<

‎library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,14 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_
2626
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances<
2727
v1,
2828
GemmDefault>{});
29+
if(ck::get_device_name() != "gfx950")
30+
{
31+
add_device_operation_instances(
32+
instances,
33+
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances_part2<
34+
v1,
35+
GemmDefault>{});
36+
}
2937
}
3038

3139
} // namespace instance

‎library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_wp/f8_f8_bf16/device_gemm_multiply_multiply_wp_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance_v2.cpp‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,11 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_
2626
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances<
2727
v2,
2828
GemmDefault>{});
29+
add_device_operation_instances(
30+
instances,
31+
device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances_part2<
32+
v2,
33+
GemmDefault>{});
2934
}
3035

3136
} // namespace instance

0 commit comments

Comments
 (0)