Skip to content

GitLab

"include/vscode:/vscode.git/clone" did not exist on "9707178f9ef78fbe6dcfb1a7a46127e57b69e757"
Commit ceeca94a authored by rocking's avatar rocking
Browse files

Merge commit 'd6709dc3' into gemm_layernorm_welford

parents 2732d06c d6709dc3
Hide whitespace changes
Inline Side-by-side
Showing with 653 additions and 162 deletions
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
View file @ ceeca94a
add_instance_library(device_batched_gemm_gemm_instance
device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
)
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ ceeca94a
......@@ -35,11 +35,21 @@ using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_inst
//################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
// Padded fallback kernel
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp 0 → 100644
View file @ ceeca94a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm_xdl_cshuffle.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances = std::tuple<
// clang-format off
//################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 4, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can trigger compiler crash in mainline #9110 but not in #10738
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 4, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 4, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can cause validation error on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 4, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can cause validation error on MI100
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 4, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 4, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 4, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 4, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 4, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 4, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 4, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 4, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 4, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 4, S<1, 32, 1, 8>, 8>,
// Padded fallback kernel
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 4, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 4, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
Col,
Col,
Row,
F16,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ ceeca94a
......@@ -26,6 +26,8 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded =
ck::tensor_operation::device::GemmSpecialization::MNOPadding; // Padding K is currently flawed
// c[g, m, n] = a[g, m, k] * b[g, n, k]
using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
......@@ -35,10 +37,21 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f16_f16_instance.cpp
View file @ ceeca94a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
namespace {
using F16 = ck::half_t;
using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
} // namespace
template <index_t Rank, index_t Reduce>
using device_softmax_f16_f16_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
DeviceSoftmaxImpl< F16, F32, F16, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
// clang-format on
>;
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f16_f16_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, Pass, Pass, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
}
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f16_f16_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, Pass, Pass, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
......
library/src/tensor_operation_instance/gpu/normalization/device_softmax_f32_f32_instance.cpp
View file @ ceeca94a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
namespace {
using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
} // namespace
template <index_t Rank, index_t Reduce>
using device_softmax_f32_f32_instances = std::tuple<
// clang-format off
// InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
// InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
DeviceSoftmaxImpl< F32, F32, F32, Pass, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
// clang-format on
>;
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, Pass, Pass, 3>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
}
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, Pass, Pass, 4>>& instances)
{
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
......
profiler/include/profile_batched_gemm_softmax_gemm_impl.hpp
View file @ ceeca94a
......@@ -147,9 +147,16 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
// Still unsure whether this kind of deterministic floating point accurary issue is expected
// or not. May want to try exact same approach as the GPU kernel in the host reference
// GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
// shrink the input value range as it is less likely to produce errors of around ~1e-3.
// a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
// b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
// b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
......
profiler/include/profile_normalization_impl.hpp
View file @ ceeca94a
......@@ -6,25 +6,36 @@
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/data_type.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>&);
namespace {
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
} // namespace
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f16_f16_rank3_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>&);
void add_device_softmax_f16_f16_rank4_instances(
std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>&);
void add_device_softmax_f32_f32_rank3_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>&);
void add_device_softmax_f32_f32_rank4_instances(
std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>&);
} // namespace instance
} // namespace device
......@@ -57,7 +68,7 @@ template <> std::string type_to_string<int8_t>() { return "int8"; }
template <> std::string type_to_string<int32_t>() { return "int32"; }
// clang-format on
template <typename InDataType, typename AccDataType, typename OutDataType>
template <typename InDataType, typename AccDataType, typename OutDataType, index_t Rank>
void profile_normalization_impl(int do_verification,
int init_method,
bool do_log,
......@@ -69,6 +80,11 @@ void profile_normalization_impl(int do_verification,
AccDataType beta,
NormType norm_type)
{
if(Rank != in_length.size())
{
throw std::runtime_error("Input tensor rank is different from template argument Rank!");
}
Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
: Tensor<InDataType>(in_length, in_strides);
Tensor<OutDataType> out(in.mDesc);
......@@ -99,30 +115,31 @@ void profile_normalization_impl(int do_verification,
std::vector<index_t> i_in_lengths(in.mDesc.GetLengths().begin(), in.mDesc.GetLengths().end());
std::vector<index_t> i_in_strides(in.mDesc.GetStrides().begin(), in.mDesc.GetStrides().end());
// add device normalization instances
std::vector<tensor_operation::device::DeviceNormalizationPtr> instances;
// add device softmax instances
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceOpPtr = tensor_operation::device::
DeviceSoftmaxPtr<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
std::vector<DeviceOpPtr> instances;
if(norm_type == NormType::SOFTMAX)
{
if constexpr(is_same<InDataType, half_t>::value && is_same<OutDataType, half_t>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
if constexpr(Rank == 3)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank3_instances(
instances);
if(in_length.size() == 4)
else if constexpr(Rank == 4)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank4_instances(
instances);
}
else if constexpr(is_same<InDataType, float>::value && is_same<OutDataType, float>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
if constexpr(Rank == 3)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank3_instances(
instances);
if(in_length.size() == 4)
else if constexpr(Rank == 4)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank4_instances(
instances);
}
......@@ -137,6 +154,8 @@ void profile_normalization_impl(int do_verification,
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
for(auto& inst_ptr : instances)
{
// Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
......@@ -153,7 +172,9 @@ void profile_normalization_impl(int do_verification,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
out_dev.GetDeviceBuffer(),
PassThrough{},
PassThrough{});
if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
......
profiler/src/profile_normalization.cpp
View file @ ceeca94a
......@@ -50,7 +50,7 @@ struct ArgParser
void print_help()
{
std::cout << "arg1: tensor operation (layernorm/batchnorm/softmax)\n"
std::cout << "arg1: tensor operation (batchnorm/softmax)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
......@@ -91,31 +91,73 @@ int profile_normalization(int argc, char* argv[])
arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
if(data_type == NormDataType::F16_F16)
if(length.size() == 3)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t, 3>(
do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float, 3>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else if(data_type == NormDataType::F32_F32)
else if(length.size() == 4)
{
ck::profiler::profile_normalization_impl<float, float, float>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t, 4>(
do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float, 4>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
......
script/process_perf_data.py
View file @ ceeca94a
......@@ -127,11 +127,16 @@ def parse_logfile(logfile):
lst=line.split()
res.append(lst[1])
#parse all other performance tests:
elif 'resnet50' or 'batched_gemm' or 'grouped_gemm' or 'conv_bwd_data' or 'gemm_bilinear' or 'reduction' in logfile:
elif 'resnet50' in logfile or 'batched_gemm' in logfile or 'grouped_gemm' in logfile or 'conv_bwd_data' in logfile or 'gemm_bilinear' in logfile or 'reduction' in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
res.append(lst[4])
elif 'onnx_gemm' in logfile or 'splitK_gemm' in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
res.append(lst[33])
return res
......@@ -281,6 +286,14 @@ def main():
for i in range(1,50):
testlist.append("Layer%i"%i)
table_name="ck_resnet50_N256_tflops"
if 'onnx_gemm' in filename:
for i in range(1,len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_onnx_gemm_tflops"
if 'splitK_gemm' in filename:
for i in range(1,len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_splitK_gemm_tflops"
tflops_base = get_baseline(table_name,conn)
store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, conn)
......
script/process_qa_data.sh
View file @ ceeca94a
......@@ -2,8 +2,8 @@
#
# in order to run this script you'd need the following python packages:
pip3 install --upgrade pip
pip3 install sqlalchemy pymysql pandas sshtunnel
#pip3 install --upgrade pip
#pip3 install sqlalchemy pymysql pandas sshtunnel
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
......@@ -20,3 +20,5 @@ python3 process_perf_data.py perf_conv_fwd_"$gpu_arch".log
python3 process_perf_data.py perf_conv_bwd_data_"$gpu_arch".log
python3 process_perf_data.py perf_gemm_bilinear_"$gpu_arch".log
python3 process_perf_data.py perf_reduction_"$gpu_arch".log
python3 process_perf_data.py perf_splitK_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_onnx_gemm_"$gpu_arch".log
script/profile_onnx_gemm.sh 0 → 100755
View file @ ceeca94a
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
DRIVER="../build/bin/ckProfiler"
echo $DRIVER
OP=$1
DATATYPE=$2
LAYOUT=$3
VERIFY=$4
INIT=$5
LOG=$6
TIME=$7
# GEMM kernel benchmarks used by ONNX
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideC
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 768 768 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 768 2304 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 768 3072 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 3072 768 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 1024 1024 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 1024 3072 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 1024 4096 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 384 4096 1024 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 768 768 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 768 2304 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 768 3072 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 3072 768 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 1024 1024 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 1024 3072 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 1024 4096 -1 -1 -1
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 24576 4096 1024 -1 -1 -1
script/run_full_performance_tests.sh
View file @ ceeca94a
......@@ -40,99 +40,103 @@ function print_log_header(){
#run gemm tests
export gemm_log="perf_gemm_${gpu_arch}.log"
print_log_header $gemm_log $env_type $branch $host_name
./profile_gemm.sh gemm 0 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
#run batched_gemm tests
export batched_gemm_log="perf_batched_gemm_${gpu_arch}.log"
print_log_header $batched_gemm_log $env_type $branch $host_name
./profile_batched_gemm.sh batched_gemm 0 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 0 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 1 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 2 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 3 $verify 1 0 1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 0 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 2 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
./profile_batched_gemm.sh batched_gemm 3 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
#run grouped_gemm tests
export grouped_gemm_log="perf_grouped_gemm_${gpu_arch}.log"
print_log_header $grouped_gemm_log $env_type $branch $host_name
./profile_grouped_gemm.sh grouped_gemm 1 0 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 1 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 2 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 3 $verify 1 0 1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
./profile_grouped_gemm.sh grouped_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
#run GEMM+Bilinear tests
export gemm_bilinear_log="perf_gemm_bilinear_${gpu_arch}.log"
print_log_header $gemm_bilinear_log $env_type $branch $host_name
./profile_gemm_bilinear.sh gemm_bilinear 1 0 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 1 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 2 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 3 $verify 1 0 1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
./profile_gemm_bilinear.sh gemm_bilinear 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
#run conv_fwd tests
export conv_fwd_log="perf_conv_fwd_${gpu_arch}.log"
print_log_header $conv_fwd_log $env_type $branch $host_name
./profile_conv_fwd.sh conv_fwd 0 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 1 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 2 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 3 1 $verify 1 0 1 256 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
./profile_conv_fwd.sh conv_fwd 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
#run conv_bwd_data tests
export conv_bwd_data_log="perf_conv_bwd_data_${gpu_arch}.log"
print_log_header $conv_bwd_data_log $env_type $branch $host_name
./profile_conv_bwd_data.sh conv_bwd_data 0 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 1 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 2 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 3 1 $verify 1 0 1 256 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
./profile_conv_bwd_data.sh conv_bwd_data 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
#run resnet50 tests
export resnet256_log="perf_resnet50_N256_${gpu_arch}.log"
print_log_header $resnet256_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 | tee -a $resnet256_log
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 2>&1 | tee -a $resnet256_log
export resnet4_log="perf_resnet50_N4_${gpu_arch}.log"
print_log_header $resnet4_log $env_type $branch $host_name
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 | tee -a $resnet4_log
./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 2>&1 | tee -a $resnet4_log
#run reduction tests
export reduction_log="perf_reduction_${gpu_arch}.log"
print_log_header $reduction_log $env_type $branch $host_name
./profile_reduce_with_index.sh $verify 2 10 --half | tee -a $reduction_log
./profile_reduce_no_index.sh $verify 2 10 --half | tee -a $reduction_log
./profile_reduce_with_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
./profile_reduce_no_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
#run splitK_gemm tests
export splitK_gemm_log="perf_splitK_gemm_${gpu_arch}.log"
print_log_header $splitK_gemm_log $env_type $branch $host_name
./profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 1 0 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 1 1 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 1 2 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
./profile_splitK_gemm.sh gemm_splitk 1 3 $verify 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 1 0 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 1 1 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 1 2 $verify 1 0 1 4 | tee -a $splitK_gemm_log
../script/profile_splitK_gemm.sh gemm_splitk 1 3 $verify 1 0 1 4 | tee -a $splitK_gemm_log
#run ONNX gemm tests
export onnx_log="perf_onnx_gemm_${gpu_arch}.log"
print_log_header $onnx_log $env_type $branch $host_name
./profile_onnx_gemm.sh gemm 0 0 $verify 2 0 1 2>&1 | tee -a $onnx_log
./profile_onnx_gemm.sh gemm 1 0 $verify 2 0 1 2>&1 | tee -a $onnx_log
test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
View file @ ceeca94a
......@@ -11,7 +11,8 @@ class TestBatchedGemmGemmFP16 : public TestBatchedGemmGemm<Tuple>
// clang-format off
using KernelTypes = ::testing::Types<
std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>
std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>,
std::tuple<F16, F16, F16, F16, Row, Col, Col, Row>
>;
// clang-format on
......@@ -68,7 +69,6 @@ TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddN)
this->Run();
}
// Currently expected that no kernels can support this case
TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddK)
{
this->lengths_ = std::vector<std::vector<int>>{
......@@ -140,9 +140,10 @@ TEST(TestBatchedGemmGemmInterface, GemmSpecializationSizeMismatch)
// clang-format off
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
// Kernel can't support odd K because K must be integer multiples of K1 values of either A or B
// Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
// Kernel can't support odd O size because it must satisfy SizeO % B1SrcScalarPerVector == 0
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
// Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
// clang-format on
}
test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
View file @ ceeca94a
......@@ -19,6 +19,73 @@ TYPED_TEST_SUITE(TestBatchedGemmSoftmaxGemmFP16, KernelTypes);
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16) { this->Run(); }
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadM)
{
this->lengths_ = std::vector<std::vector<int>>{
{136, 128, 32, 128, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadN)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 136, 32, 128, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadK)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 40, 128, 1},
{128, 128, 136, 128, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadO)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 32, 136, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddM)
{
this->lengths_ = std::vector<std::vector<int>>{
{129, 128, 32, 128, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddN)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 129, 32, 128, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddK)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 33, 128, 1},
{128, 128, 129, 128, 1},
};
this->Run();
}
// If kernel B1Layout is RowMajor, expect not to support odd O size
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddO)
{
this->lengths_ = std::vector<std::vector<int>>{
{128, 128, 32, 129, 1},
};
this->Run();
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16)
{
this->lengths_ = std::vector<std::vector<int>>{
......@@ -37,3 +104,58 @@ TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16)
this->verify_ = false;
this->Run();
}
using ck::tensor_operation::device::GemmSpecialization;
// TODO: enable KPadding tests when it is implemented
TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
{
int P = 120; // requires padding
int Q = 128; // do not require padding
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
// EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
// clang-format on
}
TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMismatch)
{
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
// Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
// Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
// clang-format on
}
TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, AdhocTest)
{
this->lengths_ = std::vector<std::vector<int>>{
{49, 49, 64, 64, 24},
{64, 49, 64, 64, 24},
{1020, 1020, 64, 128, 24},
{576, 576, 64, 64, 24},
};
this->bench_ = true;
this->Run();
}
test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
View file @ ceeca94a
......@@ -4,7 +4,10 @@
#include <iostream>
#include <vector>
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
#include "profiler/include/profile_batched_gemm_softmax_gemm_impl.hpp"
using ck::tensor_operation::device::GemmSpecialization;
template <ck::index_t N>
using I = ck::Number<N>;
......@@ -66,3 +69,121 @@ struct TestBatchedGemmSoftmaxGemm : public ::testing::Test
}
}
};
template <GemmSpecialization GemmSpec>
struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CLayout = Row;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = float;
using CShuffleDataType = float;
using CDataType = F16;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = PassThrough;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
using DeviceGemmGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
128, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
4, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CShuffleBlockTransferScalarPerVector_NPerBlock
bool IsSupported(int M, int N, int K, int O)
{
auto gemm = DeviceGemmGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
static_cast<B0DataType*>(nullptr),
static_cast<B1DataType*>(nullptr),
static_cast<CDataType*>(nullptr),
M,
N,
K,
O,
0, // BatchCount
0, // StrideA
0, // StrideB0
0, // StrideB1
0, // StrideC
0, // BatchStrideA
0, // BatchStrideB0
0, // BatchStrideB1
0, // BatchStrideC
PassThrough{}, // a_element_op
PassThrough{}, // b0_element_op
PassThrough{}, // acc0_element_op
PassThrough{}, // b1_element_op
PassThrough{}); // c_element_op
return gemm.IsSupportedArgument(argument);
}
};
test/softmax/test_softmax_util.hpp
View file @ ceeca94a
......@@ -9,7 +9,8 @@
#include "ck/ck.hpp"
#include "ck/utility/number.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -51,19 +52,23 @@ class TestSoftmax : public ::testing::Test
using ReferenceInstance =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
using DeviceInstance = tensor_operation::device::DeviceSoftmax<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceInstance = tensor_operation::device::DeviceSoftmaxImpl<InDataType,
AccDataType,
OutDataType,
PassThrough,
PassThrough,
Rank,
NumReduceDim,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize>;
TestSoftmax() : ref_instance_invoker_(ReferenceInstance{}.MakeInvoker()) {}
......@@ -97,7 +102,9 @@ class TestSoftmax : public ::testing::Test
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
out_dev.GetDeviceBuffer(),
PassThrough{},
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment