Skip to content

GitLab

"src/targets/vscode:/vscode.git/clone" did not exist on "b027d6375239959a1ec3f824bd6c8176bd17fd23"
Unverified Commit 9c42a83a authored by rocking5566's avatar rocking5566 Committed by GitHub
Browse files

Merge branch 'develop' into normalization/splitK

parents f1238b6a e6cda9f8
Expand all Show whitespace changes
Inline Side-by-side
Showing with 2123 additions and 301 deletions
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
View file @ 9c42a83a
......@@ -3,6 +3,8 @@
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
using InDataType = BF16;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = F32;
......@@ -13,6 +15,46 @@ using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
View file @ 9c42a83a
......@@ -3,6 +3,8 @@
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
......@@ -12,6 +14,46 @@ using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ 9c42a83a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
......@@ -54,7 +14,18 @@ template <ck::index_t NDimSpatial>
bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
const ck::utils::conv::ConvParam& conv_param)
{
constexpr ck::index_t split_k = 2;
ck::index_t split_k;
// Set split_k = 2 for xdl op, split_k = 1 for dl
// Dl op doesn't support split_k > 1
// TODO: Add Dl op split_k > 1 support
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
{
split_k = 2;
}
else
{
split_k = 1;
}
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
......@@ -144,7 +115,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return false;
return true;
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
......
include/ck/ck.hpp
View file @ 9c42a83a
......@@ -168,6 +168,9 @@
// tuning parameter
#define CK_WORKAROUND_SWDEV_325164 0
// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
#define CK_WORKAROUND_SWDEV_383542 1
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ 9c42a83a
......@@ -111,6 +111,7 @@ __global__ void
ignore = p_b_grid;
ignore = p_b1_grid;
ignore = p_c_grid;
ignore = p_d0s_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = c0de_element_op;
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 9c42a83a
......@@ -586,6 +586,11 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
return false;
}
if(ck::get_device_name() != "gfx90a" && std::is_same<ADataType, double>::value)
{
return false;
}
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
arg.b_grid_desc_n_k_,
arg.ds_grid_desc_m_n_,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
View file @ 9c42a83a
......@@ -86,38 +86,30 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
static auto MakeAGridDescriptor_K0_M_K1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
#ifdef ENABLE_COLMAJOR
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
#endif
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
assert(K % K1 == 0);
const index_t K0 = K / K1;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
......@@ -125,81 +117,53 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
static auto MakeBGridDescriptor_K0_N_K1(index_t KRaw, index_t NRaw, index_t StrideB)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
assert(K % K1 == 0);
const index_t K0 = K / K1;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
template <typename ELayout_>
static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
{
const auto e_grid_desc_m_n = [&]() {
const auto e_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideE, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout_>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideE));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& Ms,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
View file @ 9c42a83a
......@@ -12,6 +12,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -78,38 +79,30 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
static auto MakeAGridDescriptor_K0_M_K1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
#ifdef ENABLE_COLMAJOR
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
#endif
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
assert(K % K1 == 0);
const index_t K0 = K / K1;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
......@@ -117,80 +110,52 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
static auto MakeBGridDescriptor_K0_N_K1(index_t KRaw, index_t NRaw, index_t StrideB)
{
assert(K % K1 == 0);
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
assert(K % K1 == 0);
const index_t K0 = K / K1;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
}
// Gridwise descriptor, mapping to whole given provblem.
......
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ 9c42a83a
......@@ -4,6 +4,7 @@
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
......@@ -280,43 +281,42 @@ struct AddHardswish
};
};
// C = A * B
// E = FastGelu(C + D)
struct AddFastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>;
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& e, const float& c, const float& d) const
{
static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
is_valid_param_type_v<D>);
const float x = c + d;
FastGelu{}.template operator()<float, float>(e, x);
}
const float y = GetFastGeLU(type_convert<float>(c) + type_convert<float>(d));
template <>
__host__ __device__ constexpr void
operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
{
const half_t x = c + d;
e = type_convert<E>(y);
ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
}
template <typename D>
__host__ __device__ constexpr void operator()(float& e, const float& c, const D& d) const
template <>
__host__ __device__ constexpr void
operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
{
static_assert(is_valid_param_type_v<D>);
const float x0_f = c + d;
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = GetFastGeLU(c + type_convert<float>(d));
e = type_convert<half_t>(x1_f);
}
};
......
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 9c42a83a
......@@ -16,7 +16,7 @@ namespace element_wise {
// Need to ensure compiler will fail if there is no matching candidate, instead of compiler
// siliently do implicit type conversion
//
// Method 1:
// Example:
//
// struct ExampleElementwiseOp
// {
......@@ -30,19 +30,6 @@ namespace element_wise {
// {
// }
// };
//
// Method 2:
//
// template <typename Y, typename X>
// struct ExampleElementwiseOp;
//
// template <>
// struct ExampleElementwiseOp<float, ck::bhalf_t>
// {
// __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
// {
// }
// };
struct AddReluAdd
{
......@@ -208,41 +195,74 @@ struct AddMultiply
}
};
// C = A * B
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
template <>
__host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
const float& c,
const float& d0,
const float& d1) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
const float x = c + d0 + d1;
FastGelu{}.template operator()<float, float>(e, x);
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<T, ck::int4_t>
#endif
;
template <>
__host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
{
const half_t x = c + d0 + d1;
template <typename E, typename C, typename D0, typename D1>
__host__ __device__ constexpr void
operator()(E& e, const C& c, const D0& d0, const D1& d1) const
ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
}
template <>
__host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
half_t& e, const float& c, const half_t& d0, const half_t& d1) const
{
static_assert(is_valid_param_type_v<E> && is_valid_param_type_v<C> &&
is_valid_param_type_v<D0> && is_valid_param_type_v<D1>);
const float x0_f = c + d0 + d1;
const float y =
GetFastGeLU(type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1));
float x1_f = 0;
e = type_convert<E>(y);
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<half_t>(x1_f);
}
template <>
__host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
bhalf_t& e, const float& c, const bhalf_t& d0, const bhalf_t& d1) const
{
const float x0_f = c + type_convert<float>(d0) + type_convert<float>(d1);
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<bhalf_t>(x1_f);
}
template <>
__host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t, int8_t>(
int8_t& e, const int32_t& c, const int8_t& d0, const int8_t& d1) const
{
const float x0_f =
type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1);
float x1_f = 0;
ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
x0_f);
e = type_convert<int8_t>(x1_f);
}
};
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 9c42a83a
......@@ -11,6 +11,10 @@ namespace ck {
namespace tensor_operation {
namespace element_wise {
#if CK_WORKAROUND_SWDEV_383542
extern "C" __device__ float __ocml_native_recip_f32(float);
#endif
struct PassThrough
{
template <typename Y, typename X>
......@@ -200,36 +204,83 @@ struct Relu
}
};
// Y = FastGelu(X)
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
// host code use higher accuracy "exp" and "div"
// gpu code use lower accuracy "__expf" and "rcp" function
struct FastGelu
{
// Fast GeLU
// https://paperswithcode.com/method/gelu
// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
__host__ __device__ static constexpr float GetFastGeLU(float x)
template <typename Y, typename X>
__host__ void operator()(Y& y, const X& x) const;
template <typename Y, typename X>
__device__ void operator()(Y& y, const X& x) const;
template <>
__host__ void operator()<float, float>(float& y, const float& x) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = exp(-u);
const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
return x * cdf;
y = x * cdf;
}
template <typename T>
static inline constexpr bool is_valid_param_type_v =
std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<T, ck::int4_t>
// device code, use lower precision "__expf" and "rcp"
template <>
__device__ void operator()<float, float>(float& y, const float& x) const
{
const float u = 2.f * x * (0.035677f * x * x + 0.797885f);
const float emu = __expf(-u);
#if !CK_WORKAROUND_SWDEV_383542
const float cdf = 0.5f + 0.5f * (2.f * __frcp_rn(1.f + emu) - 1.f);
#else
const float cdf = 0.5f + 0.5f * (2.f * __ocml_native_recip_f32(1.f + emu) - 1.f);
#endif
;
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const
y = x * cdf;
}
template <>
__host__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
{
static_assert(is_valid_param_type_v<Y> && is_valid_param_type_v<X>);
float y_f;
this->operator()<float, float>(y_f, type_convert<float>(x));
y = type_convert<half_t>(y_f);
}
template <>
__device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
{
float y_f;
this->operator()<float, float>(y_f, type_convert<float>(x));
y = type_convert<half_t>(y_f);
}
template <>
__host__ void operator()<half_t, float>(half_t& y, const float& x) const
{
float y_f;
this->operator()<float, float>(y_f, x);
y = type_convert<half_t>(y_f);
}
template <>
__device__ void operator()<half_t, float>(half_t& y, const float& x) const
{
float y_f;
this->operator()<float, float>(y_f, x);
const float tmp_y = GetFastGeLU(type_convert<float>(x));
y = type_convert<Y>(tmp_y);
y = type_convert<half_t>(y_f);
}
};
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View file @ 9c42a83a
......@@ -414,7 +414,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
FloatA,
FloatB,
FloatAcc,
......
script/cmake-ck-dev.sh
View file @ 9c42a83a
......@@ -10,8 +10,8 @@ cmake
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=OFF \
-D GPU_TARGETS="gfx90a" \
-D BUILD_DEV=ON \
-D GPU_TARGETS="gfx908;gfx90a" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
......
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