Skip to content

GitLab

Unverified Commit aaf4defa authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #9 from ROCmSoftwarePlatform/mergedown-from-public 

Merge down from public repo
parents e0c2a70f e51806cc
Hide whitespace changes
Inline Side-by-side
Showing with 417 additions and 126 deletions
example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc 0 → 100644
View file @ aaf4defa
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNDFwdInstance>
bool run_grouped_conv_fwd(bool do_verification,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& requant_scale_g_k_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_desc);
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "requant_scale: " << requant_scale.mDesc << std::endl;
std::cout << "out: " << out_host.mDesc << std::endl;
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
requant_scale.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
requant_scale_device_buf.ToDevice(requant_scale.mData.data());
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(requant_scale_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
copy(requant_scale_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
// do Conv
auto conv = DeviceConvNDFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
{requant_scale_device_buf.GetDeviceBuffer()},
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
{d0_g_n_k_wos_lengths},
{d0_g_n_k_wos_strides},
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< conv.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
Tensor<AccDataType> c_host(out_g_n_k_wos_desc);
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
InDataType,
WeiDataType,
AccDataType,
InElementOp,
WeiElementOp,
PassThrough>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei,
c_host,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
in_element_op,
wei_element_op,
PassThrough{});
ref_invoker.Run(ref_argument);
// TODO: implement elementwise operation for host
out_host.ForEach([&](auto&, auto idx) {
out_element_op(out_host(idx), c_host(idx), requant_scale(idx));
});
out_device_buf.FromDevice(out_device.mData.data());
pass &=
ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
}
return (pass ? 0 : 1);
}
int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_element_op)
{
bool do_verification = true;
bool time_kernel = true;
const ck::index_t ndim_spatial = 2;
ck::utils::conv::ConvParam conv_param{
ndim_spatial, // n_dim
1, // group
4, // batch
64, // output channels
192, // input chanels
{3, 3}, // weight HW
{71, 71}, // x HW
{2, 2}, // strides
{1, 1}, // dilations
{1, 1}, // left_pads
{1, 1} // right_pads
};
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
const auto requant_scale_g_k_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
0, // n
1, // k
0, // ho
0 // wo
});
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
InLayout,
WeiLayout,
RequantScaleLayout,
OutLayout>;
return run_grouped_conv_fwd<ndim_spatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
deviceOp>(do_verification,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
requant_scale_g_k_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
example/44_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp → example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
View file @ aaf4defa
// 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/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.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/utility/literals.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
16>;
#pragma once
template <ck::index_t NDimSpatial,
typename InDataType,
......@@ -221,10 +139,10 @@ bool run_grouped_conv_fwd(bool do_verification,
return (pass ? 0 : 1);
}
int main()
int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element_op)
{
bool do_verification = true;
bool time_kernel = true;
bool time_kernel = false;
const ck::index_t ndim_spatial = 2;
ck::utils::conv::ConvParam conv_param{
......@@ -232,7 +150,7 @@ int main()
1, // group
4, // batch
64, // output channels
32, // input chanels
192, // input chanels
{3, 3}, // weight HW
{71, 71}, // x HW
{2, 2}, // strides
......@@ -243,7 +161,6 @@ int main()
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
......
example/44_conv2d_fwd_quantization/CMakeLists.txt deleted 100644 → 0
View file @ e0c2a70f
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
include/ck/ck.hpp
View file @ aaf4defa
......@@ -27,14 +27,6 @@
#define CK_WAVELET_MIN_BLOCK_PER_CU 2
#endif
// check GPU target
#ifdef __HIP_DEVICE_COMPILE__
#if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx940__))
#error Not supported target
#endif
#endif
// buffer resource
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
......@@ -43,8 +35,8 @@
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx1030__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
#elif defined(__gfx1100__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
#endif
// FMA instruction
......@@ -76,7 +68,7 @@
// WMMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_WMMA
#elif defined(__gfx1100__) // for GPU code
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
#define CK_USE_AMD_WMMA
#endif
......@@ -172,6 +164,11 @@
// tuning parameter
#define CK_WORKAROUND_SWDEV_325164 0
// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
#define CK_WORKAROUND_SWDEV_383542 1
// workaround: compiler issue on gfx908
#define CK_WORKAROUND_SWDEV_388832 1
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ aaf4defa
......@@ -622,11 +622,16 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
}
};
// Blockwise gemm supporting
// 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4
// 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS
// source buffer
// 3. configurable k index starting position and step size after each FMA/XDL instruction
/**
* @brief Blockwise gemm
*
* Supports
* 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4
* 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS
* source buffer
* 3. configurable k index starting position and step size after each FMA/XDL instruction
*/
template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
......
include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
View file @ aaf4defa
......@@ -12,6 +12,16 @@
namespace ck {
/**
* @brief Blockwise softmax
*
* @tparam BlockSize Block size
* @tparam AccDataType Accumulator data type
* @tparam ThreadMap_M_K Thread id to m_k
* @tparam ThreadClusterDesc_M_K Threadwise cluster descriptor
* @tparam ThreadSliceDesc_M_K Threadwise slices descriptor
* @tparam IgnoreNaN Flag to ignore NaN, false by default
*/
template <index_t BlockSize,
typename AccDataType,
typename ThreadMap_M_K, // thread_id to m_k
......
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
View file @ aaf4defa
......@@ -11,10 +11,15 @@
namespace ck {
// this version does following things to avoid scratch memory issue
// 1. Use StaticallyIndexedArray instead of C array for thread buffer
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
/**
* @brief Blockwise data transfer
*
* This version does following things to avoid scratch memory issue
* 1. Use StaticallyIndexedArray instead of C array for thread buffer
* 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
* 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
*
*/
template <typename ThreadGroup,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -134,7 +134,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx90a__) || defined(__gfx908__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -314,9 +315,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
const auto in_gemmm_gemmk_desc =
matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
const auto M = in_gemmm_gemmk_desc.GetLength(I0);
const auto K = in_gemmm_gemmk_desc.GetLength(I1);
const auto M = in_gemmm_gemmk_desc.GetLength(I0);
const auto K = in_gemmm_gemmk_desc.GetLength(I1);
const auto AK0 = K / K1;
return transform_tensor_descriptor(
......@@ -709,7 +709,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
namespace ctc = tensor_layout::convolution;
// check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908"))
{
return false;
}
......@@ -834,6 +835,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
{
return false;
}
// check Gridwise GEMM
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);
......@@ -946,7 +948,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
<< K1
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -824,7 +824,19 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1 << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< CBlockTransferScalarPerVector_NWaveNPerXdl
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
View file @ aaf4defa
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ aaf4defa
......@@ -866,7 +866,16 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
<< "B0Spec" << getTensorSpecializationString(BSpec) << ", "
<< "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
<< "CSpec" << getTensorSpecializationString(CSpec) << ", "
<< getMaskingSpecializationString(MaskingSpec) << ">";
<< getMaskingSpecializationString(MaskingSpec) << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle
<< ">";
// clang-format on
return str.str();
......
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
View file @ aaf4defa
......@@ -770,7 +770,8 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
View file @ aaf4defa
......@@ -789,6 +789,20 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
return true;
}
// check if DsLayout is supported
template <typename RefLayout, typename DsLayout, const index_t NumDTensor>
static bool CheckDLayout()
{
static bool valid = true;
// iterate over DLayout tuple
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
// if RefLayout and DLayout are same, keep valid true, otherwise false
valid = valid && is_same_v<RefLayout, DLayout>;
});
return valid;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
......@@ -797,6 +811,26 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
return false;
}
// Check supported layouts
// A0 - Row
// B0 - Col
// D0s - Rows
// B1 - Row or Col
// D1s - Rows
// E1 - Row
if(!(is_same_v<tensor_layout::gemm::RowMajor, A0Layout> &&
is_same_v<tensor_layout::gemm::ColumnMajor, B0Layout> &&
CheckDLayout<tensor_layout::gemm::RowMajor, D0sLayout, NumD0Tensor>() &&
(is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ||
is_same_v<tensor_layout::gemm::ColumnMajor,
B1Layout>)&&CheckDLayout<tensor_layout::gemm::RowMajor,
D1sLayout,
NumD1Tensor>() &&
is_same_v<tensor_layout::gemm::RowMajor, E1Layout>))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a0_grid_desc_m_k_,
arg.b0_grid_desc_n_k_,
arg.b1_grid_desc_n_k_,
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ aaf4defa
......@@ -588,6 +588,11 @@ 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_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -777,7 +777,19 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1 << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< CBlockTransferScalarPerVector_NWaveNPerXdl
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -822,7 +822,14 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< K0PerBlock << ", "
<< K1 << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -956,7 +956,19 @@ struct
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1 << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< CBlockTransferScalarPerVector_NWaveNPerXdl
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -913,7 +913,19 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1 << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< CBlockTransferScalarPerVector_NWaveNPerXdl
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -880,7 +880,17 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
<< K1 << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1 << ", "
<< CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", "
<< CBlockTransferScalarPerVector_NWaveNPerXdl
<< ">";
// clang-format on
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ aaf4defa
......@@ -720,7 +720,16 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< getConvForwardSpecializationString(ConvForwardSpecialization)
<< getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ABlockTransferSrcScalarPerVector << ", "
<< ABlockTransferDstScalarPerVector_K1 << ", "
<< BBlockTransferSrcScalarPerVector << ", "
<< BBlockTransferDstScalarPerVector_K1
<< ">";
// clang-format on
......
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