mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-24 06:44:36 +00:00
Unify Convolution FWD XDL 1D/2D implementation. (#93)
* Convolution ND
* Code unification across dimensions for generating tensor descriptors.
* Example
* Instances
* Move convnd f32 instance file to comply with repo structure.
* Conv 1D tensor layouts.
* Formatting and use ReferenceConv
* Reference ConvFwd supporting 1D and 2D convolution.
* Debug printing TensorLayout name.
* Conv fwd 1D instance f32
* Refactor conv ND example.
Needed to support various conv dimensio.
Needed to support various conv dimensions
* Rename conv nd example director to prevent conflicts.
* Refactor some common utility to single file.
Plus some tests.
* Refactor GetHostTensorDescriptor + UT.
* Add 1D test case.
* Test reference convolution 1d/2d
* Remove some leftovers.
* Fix convolution example error for 1D
* Refactor test check errors utility function.
* Test Conv2D Fwd XDL
* More UT for 1D case.
* Parameterize input & weight initializers.
* Rename example to prevent conflicts.
* Split convnd instance into separate files for 1d/2d
* Address review comments.
* Fix data type for flops/gbytes calculations.
* Assign example number 11.
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
[ROCm/composable_kernel commit: 756a761727]
This commit is contained in:
Adam Osewski
@@ -76,6 +76,11 @@ set(DEVICE_CONV2D_FWD_INSTANCE_SOURCE
|
||||
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp;
|
||||
)
|
||||
|
||||
# device_conv1d_fwd_instance
|
||||
set(DEVICE_CONV1D_FWD_INSTANCE_SOURCE
|
||||
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instance.cpp;
|
||||
)
|
||||
|
||||
# device_conv2d_fwd_bias_relu_instance
|
||||
set(DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE
|
||||
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp;
|
||||
@@ -96,16 +101,18 @@ add_library(device_gemm_bias_2d_instance SHARED ${DEVICE_GEMM_BIAS_2D_INSTANCE_S
|
||||
add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE})
|
||||
add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE})
|
||||
add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE})
|
||||
add_library(device_conv1d_fwd_instance SHARED ${DEVICE_CONV1D_FWD_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE})
|
||||
add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE})
|
||||
|
||||
target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_gemm_bias_2d_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_gemm_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_gemm_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_batched_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_conv1d_fwd_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_conv2d_fwd_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_conv2d_fwd_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
target_include_directories(device_conv2d_fwd_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
|
||||
@@ -116,6 +123,7 @@ target_compile_features(device_gemm_bias_2d_instance PUBLIC)
|
||||
target_compile_features(device_gemm_bias_relu_instance PUBLIC)
|
||||
target_compile_features(device_gemm_bias_relu_add_instance PUBLIC)
|
||||
target_compile_features(device_batched_gemm_instance PUBLIC)
|
||||
target_compile_features(device_conv1d_fwd_instance PUBLIC)
|
||||
target_compile_features(device_conv2d_fwd_instance PUBLIC)
|
||||
target_compile_features(device_conv2d_fwd_bias_relu_instance PUBLIC)
|
||||
target_compile_features(device_conv2d_fwd_bias_relu_add_instance PUBLIC)
|
||||
@@ -126,6 +134,7 @@ set_target_properties(device_gemm_bias_2d_instance PROPERTIES POSITION_INDEPENDE
|
||||
set_target_properties(device_gemm_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_gemm_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_batched_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_conv1d_fwd_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_conv2d_fwd_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_conv2d_fwd_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
set_target_properties(device_conv2d_fwd_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
@@ -136,7 +145,8 @@ install(TARGETS device_gemm_bias_2d_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_gemm_bias_relu_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_gemm_bias_relu_add_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_batched_gemm_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_atomic_add_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv1d_fwd_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib)
|
||||
install(TARGETS device_conv2d_fwd_bias_relu_atomic_add_instance LIBRARY DESTINATION lib)
|
||||
|
||||
198
device_operation/include/conv_utils.hpp
Normal file
198
device_operation/include/conv_utils.hpp
Normal file
@@ -0,0 +1,198 @@
|
||||
#ifndef CONV_UTILS_HPP
|
||||
#define CONV_UTILS_HPP
|
||||
|
||||
#include <cstdlib>
|
||||
#include <functional>
|
||||
#include <iterator>
|
||||
#include <numeric>
|
||||
#include <sstream>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
|
||||
#include "config.hpp"
|
||||
#include "host_tensor.hpp"
|
||||
#include "tensor_layout.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace conv_util {
|
||||
|
||||
/**
|
||||
* @brief Calculate number of FLOPs for Convolution
|
||||
*
|
||||
* @param[in] N Batch size.
|
||||
* @param[in] C Number of input channels.
|
||||
* @param[in] K Number of output channels.
|
||||
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
|
||||
* @param[in] output_spatial_lengths Convolution output spatial dimensions
|
||||
* lengths.
|
||||
*
|
||||
* @return The number of flops.
|
||||
*/
|
||||
std::size_t GetFlops(ck::index_t N,
|
||||
ck::index_t C,
|
||||
ck::index_t K,
|
||||
const std::vector<ck::index_t>& filter_spatial_lengths,
|
||||
const std::vector<ck::index_t>& output_spatial_lengths)
|
||||
{
|
||||
// 2 * N * K * <output spatial lengths product> * C * <filter spatial lengths product>
|
||||
return static_cast<std::size_t>(2) * N * K *
|
||||
std::accumulate(std::begin(output_spatial_lengths),
|
||||
std::end(output_spatial_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>()) *
|
||||
C *
|
||||
std::accumulate(std::begin(filter_spatial_lengths),
|
||||
std::end(filter_spatial_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>());
|
||||
}
|
||||
|
||||
/**
|
||||
* @brief Calculate number of bytes read/write by convolution algorithm.
|
||||
*
|
||||
* @param[in] N Batch size.
|
||||
* @param[in] C Number of input channels.
|
||||
* @param[in] K Number of output channels.
|
||||
* @param[in] input_spatial_lengths Input spatial dimensions lengths.
|
||||
* @param[in] filter_spatial_lengths Filter spatial dimensions lengths.
|
||||
* @param[in] output_spatial_lengths Output spatial dimensions lengths
|
||||
*
|
||||
* @tparam InDataType Input tensor data type.
|
||||
* @tparam WeiDataType Weights tensor data type.
|
||||
* @tparam OutDataType Output tensor data type.
|
||||
*
|
||||
* @return The number of used bytes.
|
||||
*/
|
||||
template <typename InDataType = float,
|
||||
typename WeiDataType = InDataType,
|
||||
typename OutDataType = InDataType>
|
||||
std::size_t GetBtype(ck::index_t N,
|
||||
ck::index_t C,
|
||||
ck::index_t K,
|
||||
const std::vector<ck::index_t>& input_spatial_lengths,
|
||||
const std::vector<ck::index_t>& filter_spatial_lengths,
|
||||
const std::vector<ck::index_t>& output_spatial_lengths)
|
||||
{
|
||||
// sizeof(InDataType) * (N * C * <input spatial lengths product>) +
|
||||
// sizeof(WeiDataType) * (K * C * <filter spatial lengths product>) +
|
||||
// sizeof(OutDataType) * (N * K * <output spatial lengths product>);
|
||||
return sizeof(InDataType) * (N * C *
|
||||
std::accumulate(std::begin(input_spatial_lengths),
|
||||
std::end(input_spatial_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>())) +
|
||||
sizeof(WeiDataType) * (K * C *
|
||||
std::accumulate(std::begin(filter_spatial_lengths),
|
||||
std::end(filter_spatial_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>())) +
|
||||
sizeof(OutDataType) * (N * K *
|
||||
std::accumulate(std::begin(output_spatial_lengths),
|
||||
std::end(output_spatial_lengths),
|
||||
static_cast<std::size_t>(1),
|
||||
std::multiplies<std::size_t>()));
|
||||
}
|
||||
|
||||
struct ConvParams
|
||||
{
|
||||
ConvParams()
|
||||
: num_dim_spatial(2),
|
||||
N(128),
|
||||
K(256),
|
||||
C(192),
|
||||
filter_spatial_lengths(2, 3),
|
||||
input_spatial_lengths(2, 71),
|
||||
conv_filter_strides(2, 2),
|
||||
conv_filter_dilations(2, 1),
|
||||
input_left_pads(2, 1),
|
||||
input_right_pads(2, 1)
|
||||
{
|
||||
}
|
||||
|
||||
ck::index_t num_dim_spatial;
|
||||
ck::index_t N;
|
||||
ck::index_t K;
|
||||
ck::index_t C;
|
||||
|
||||
std::vector<ck::index_t> filter_spatial_lengths;
|
||||
std::vector<ck::index_t> input_spatial_lengths;
|
||||
|
||||
std::vector<ck::index_t> conv_filter_strides;
|
||||
std::vector<ck::index_t> conv_filter_dilations;
|
||||
|
||||
std::vector<ck::index_t> input_left_pads;
|
||||
std::vector<ck::index_t> input_right_pads;
|
||||
|
||||
std::vector<ck::index_t> GetOutputSpatialLengths() const
|
||||
{
|
||||
std::vector<ck::index_t> out_spatial_len(num_dim_spatial, 0);
|
||||
for(ck::index_t i = 0; i < num_dim_spatial; ++i)
|
||||
{
|
||||
// XEff = (X - 1) * conv_dilation_w + 1;
|
||||
// Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
|
||||
const ck::index_t idx_eff =
|
||||
(filter_spatial_lengths[i] - 1) * conv_filter_dilations[i] + 1;
|
||||
out_spatial_len[i] =
|
||||
(input_spatial_lengths[i] + input_left_pads[i] + input_right_pads[i] - idx_eff) /
|
||||
conv_filter_strides[i] +
|
||||
1;
|
||||
}
|
||||
return out_spatial_len;
|
||||
}
|
||||
};
|
||||
|
||||
/**
|
||||
* @brief Gets the host tensor descriptor.
|
||||
*
|
||||
* @param[in] dims The tensor dimensions lengths. Always in NCHW format.
|
||||
* @param[in] layout The tensor data layout.
|
||||
*
|
||||
* @tparam TensorLayout Layout type.
|
||||
*
|
||||
* @return The host tensor descriptor object.
|
||||
*/
|
||||
template <typename TensorLayout>
|
||||
HostTensorDescriptor GetHostTensorDescriptor(const std::vector<std::size_t>& dims,
|
||||
const TensorLayout& layout)
|
||||
{
|
||||
std::size_t C = dims[1];
|
||||
// 1D
|
||||
if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NCW>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KCX>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NKW>::value)
|
||||
{
|
||||
|
||||
return HostTensorDescriptor(dims, std::vector<std::size_t>({C * dims[2], dims[2], 1}));
|
||||
}
|
||||
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NWC>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KXC>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NWK>::value)
|
||||
{
|
||||
return HostTensorDescriptor(dims, std::vector<std::size_t>({C * dims[2], 1, C}));
|
||||
}
|
||||
// 2D
|
||||
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NCHW>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KCYX>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NKHW>::value)
|
||||
{
|
||||
|
||||
return HostTensorDescriptor(
|
||||
dims, std::vector<std::size_t>{C * dims[2] * dims[3], dims[2] * dims[3], dims[3], 1});
|
||||
}
|
||||
else if constexpr(std::is_same<TensorLayout, ck::tensor_layout::convolution::NHWC>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::KYXC>::value ||
|
||||
std::is_same<TensorLayout, ck::tensor_layout::convolution::NHWK>::value)
|
||||
{
|
||||
return HostTensorDescriptor(
|
||||
dims, std::vector<std::size_t>{C * dims[2] * dims[3], 1, dims[3] * C, C});
|
||||
}
|
||||
|
||||
std::stringstream err_msg;
|
||||
err_msg << "Unsupported data layout provided: " << layout << "!";
|
||||
throw std::runtime_error(err_msg.str());
|
||||
}
|
||||
|
||||
} // namespace conv_util
|
||||
} // namespace ck
|
||||
|
||||
#endif
|
||||
@@ -215,7 +215,6 @@ struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_
|
||||
else if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
|
||||
{
|
||||
|
||||
static_assert(ConvForwardSpecialization == -1, "Not implemented!");
|
||||
}
|
||||
else
|
||||
|
||||
@@ -0,0 +1,865 @@
|
||||
#ifndef DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP
|
||||
#define DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP
|
||||
|
||||
#include <functional>
|
||||
#include <iostream>
|
||||
#include <iterator>
|
||||
#include <numeric>
|
||||
#include <sstream>
|
||||
|
||||
#include "device.hpp"
|
||||
#include "device_base.hpp"
|
||||
#include "device_conv_fwd.hpp"
|
||||
#include "convolution_forward_specialization.hpp"
|
||||
#include "common_header.hpp"
|
||||
#include "tensor_layout.hpp"
|
||||
#include "tensor_descriptor.hpp"
|
||||
#include "tensor_descriptor_helper.hpp"
|
||||
#include "gridwise_gemm_xdlops_v2r3.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
|
||||
//
|
||||
// @brief Device Convolution operation.
|
||||
//
|
||||
// Supports:
|
||||
// @li Inputs with up to 3 spatial dimentions
|
||||
// @li Input tensor in NHWC data format
|
||||
// @li Weight tensor in KYXC data format
|
||||
// @li Output tensor in NHWK data format
|
||||
//
|
||||
// 1D:
|
||||
// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
|
||||
// 2D:
|
||||
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
|
||||
// 3D:
|
||||
// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
|
||||
//
|
||||
template <typename InDataType,
|
||||
typename WeiDataType,
|
||||
typename OutDataType,
|
||||
typename AccDataType,
|
||||
typename InElementwiseOperation,
|
||||
typename WeiElementwiseOperation,
|
||||
typename OutElementwiseOperation,
|
||||
ConvolutionForwardSpecialization_t ConvForwardSpecialization,
|
||||
ck::index_t NumDimSpatial,
|
||||
ck::index_t BlockSize,
|
||||
ck::index_t MPerBlock,
|
||||
ck::index_t NPerBlock,
|
||||
ck::index_t K0PerBlock,
|
||||
ck::index_t K1,
|
||||
ck::index_t MPerXDL,
|
||||
ck::index_t NPerXDL,
|
||||
ck::index_t MXdlPerWave,
|
||||
ck::index_t NXdlPerWave,
|
||||
typename ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
typename ABlockTransferThreadClusterArrangeOrder,
|
||||
typename ABlockTransferSrcAccessOrder,
|
||||
ck::index_t ABlockTransferSrcVectorDim,
|
||||
ck::index_t ABlockTransferSrcScalarPerVector,
|
||||
ck::index_t ABlockTransferDstScalarPerVector_K1,
|
||||
bool ABlockLdsAddExtraM,
|
||||
typename BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
typename BBlockTransferThreadClusterArrangeOrder,
|
||||
typename BBlockTransferSrcAccessOrder,
|
||||
ck::index_t BBlockTransferSrcVectorDim,
|
||||
ck::index_t BBlockTransferSrcScalarPerVector,
|
||||
ck::index_t BBlockTransferDstScalarPerVector_K1,
|
||||
bool BBlockLdsAddExtraN,
|
||||
ck::index_t CThreadTransferSrcDstVectorDim,
|
||||
ck::index_t CThreadTransferDstScalarPerVector>
|
||||
struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
|
||||
: public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
|
||||
{
|
||||
using DeviceOp = DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
|
||||
|
||||
using ADataType = InDataType;
|
||||
using BDataType = WeiDataType;
|
||||
using CDataType = OutDataType;
|
||||
|
||||
// TODO make A/B datatype different
|
||||
using ABDataType = InDataType;
|
||||
|
||||
static constexpr index_t NDimSpatial = NumDimSpatial;
|
||||
|
||||
static constexpr auto I0 = Number<0>{};
|
||||
static constexpr auto I1 = Number<1>{};
|
||||
static constexpr auto I2 = Number<2>{};
|
||||
static constexpr auto I3 = Number<3>{};
|
||||
|
||||
static constexpr auto K1Number = Number<K1>{};
|
||||
static constexpr auto GemmK1Number = K1Number;
|
||||
|
||||
static auto GetWeightTensorDescriptor(ck::index_t gemm_n, ck::index_t gemm_k)
|
||||
{
|
||||
const ck::index_t gemm_k0 = gemm_k / GemmK1Number;
|
||||
const auto wei_k_yxc_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(gemm_n, gemm_k));
|
||||
|
||||
// wei_gemmk0_gemmn_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
wei_k_yxc_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_pass_through_transform(gemm_n)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
}
|
||||
|
||||
static auto
|
||||
GetOutputTensorDescriptor(ck::index_t gemm_m, ck::index_t gemm_n, ck::index_t gemm_m_pad)
|
||||
{
|
||||
const auto out_gemmmraw_gemmn_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_n));
|
||||
|
||||
// out_gemmm_gemmn_grid_desc
|
||||
return transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
|
||||
make_tuple(make_right_pad_transform(gemm_m, gemm_m_pad),
|
||||
make_pass_through_transform(gemm_n)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
}
|
||||
|
||||
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
|
||||
static auto GetInputTensorDescriptor(ck::index_t N,
|
||||
ck::index_t C,
|
||||
ck::index_t gemm_m,
|
||||
ck::index_t gemm_k,
|
||||
ck::index_t gemm_m_pad,
|
||||
const std::vector<ck::index_t>& input_spatial_lengths,
|
||||
const std::vector<ck::index_t>& filter_spatial_lengths,
|
||||
const std::vector<ck::index_t>& output_spatial_lengths,
|
||||
const std::vector<ck::index_t>& conv_filter_strides,
|
||||
const std::vector<ck::index_t>& conv_filter_dilations,
|
||||
const std::vector<ck::index_t>& input_left_pads,
|
||||
const std::vector<ck::index_t>& input_right_pads)
|
||||
{
|
||||
const ck::index_t gemm_k0 = gemm_k / GemmK1Number;
|
||||
const index_t Wi = input_spatial_lengths[0];
|
||||
const index_t Wo = output_spatial_lengths[0];
|
||||
const index_t ConvStrideW = conv_filter_strides[0];
|
||||
|
||||
if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
|
||||
{
|
||||
const auto in_gemmmraw_gemmk_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmmraw_gemmk_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
}
|
||||
else if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
|
||||
{
|
||||
const auto in_n_wi_c_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
|
||||
|
||||
const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_wi_c_grid_desc,
|
||||
make_tuple(make_pass_through_transform(N),
|
||||
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
|
||||
const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
|
||||
in_n_wo_c_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_merge_transform(make_tuple(N, Wo))),
|
||||
make_tuple(Sequence<2>{}, Sequence<0, 1>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmk0_gemmmraw_gemmk1_grid_desc,
|
||||
make_tuple(make_pass_through_transform(gemm_k0),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad),
|
||||
make_pass_through_transform(GemmK1Number)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
const index_t X = filter_spatial_lengths[0];
|
||||
const index_t ConvDilationW = conv_filter_dilations[0];
|
||||
const index_t InLeftPadW = input_left_pads[0];
|
||||
const index_t InRightPadW = input_right_pads[0];
|
||||
|
||||
const auto in_n_wi_c_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
|
||||
|
||||
const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_wi_c_grid_desc,
|
||||
make_tuple(make_pass_through_transform(N),
|
||||
make_pad_transform(Wi, InLeftPadW, InRightPadW),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
|
||||
const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_wip_c_grid_desc,
|
||||
make_tuple(
|
||||
make_pass_through_transform(N),
|
||||
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
|
||||
|
||||
const auto in_gemmk_gemmmraw_grid_desc =
|
||||
transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
|
||||
make_tuple(make_merge_transform(make_tuple(X, C)),
|
||||
make_merge_transform(make_tuple(N, Wo))),
|
||||
make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
|
||||
in_gemmk_gemmmraw_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_pass_through_transform(gemm_m)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmk0_gemmmraw_gemmk1_grid_desc,
|
||||
make_tuple(make_pass_through_transform(gemm_k0),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad),
|
||||
make_pass_through_transform(GemmK1Number)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
}
|
||||
}
|
||||
|
||||
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
|
||||
static auto GetInputTensorDescriptor(ck::index_t N,
|
||||
ck::index_t C,
|
||||
ck::index_t gemm_m,
|
||||
ck::index_t gemm_k,
|
||||
ck::index_t gemm_m_pad,
|
||||
const std::vector<ck::index_t>& input_spatial_lengths,
|
||||
const std::vector<ck::index_t>& filter_spatial_lengths,
|
||||
const std::vector<ck::index_t>& output_spatial_lengths,
|
||||
const std::vector<ck::index_t>& conv_filter_strides,
|
||||
const std::vector<ck::index_t>& conv_filter_dilations,
|
||||
const std::vector<ck::index_t>& input_left_pads,
|
||||
const std::vector<ck::index_t>& input_right_pads)
|
||||
{
|
||||
const ck::index_t gemm_k0 = gemm_k / GemmK1Number;
|
||||
const index_t Hi = input_spatial_lengths[0];
|
||||
const index_t Wi = input_spatial_lengths[1];
|
||||
|
||||
const index_t Ho = output_spatial_lengths[0];
|
||||
const index_t Wo = output_spatial_lengths[1];
|
||||
|
||||
const index_t ConvStrideH = conv_filter_strides[0];
|
||||
const index_t ConvStrideW = conv_filter_strides[1];
|
||||
|
||||
if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
|
||||
{
|
||||
const auto in_gemmmraw_gemmk_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(gemm_m, gemm_k));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmmraw_gemmk_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad)),
|
||||
make_tuple(Sequence<1>{}, Sequence<0>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
}
|
||||
else if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
|
||||
{
|
||||
const auto in_n_hi_wi_c_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
|
||||
|
||||
const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_hi_wi_c_grid_desc,
|
||||
make_tuple(make_pass_through_transform(N),
|
||||
make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
|
||||
make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
|
||||
|
||||
const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
|
||||
in_n_ho_wo_c_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_merge_transform(make_tuple(N, Ho, Wo))),
|
||||
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmk0_gemmmraw_gemmk1_grid_desc,
|
||||
make_tuple(make_pass_through_transform(gemm_k0),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad),
|
||||
make_pass_through_transform(GemmK1Number)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
}
|
||||
else
|
||||
{
|
||||
const index_t Y = filter_spatial_lengths[0];
|
||||
const index_t X = filter_spatial_lengths[1];
|
||||
|
||||
const index_t ConvDilationH = conv_filter_dilations[0];
|
||||
const index_t ConvDilationW = conv_filter_dilations[1];
|
||||
|
||||
const index_t InLeftPadH = input_left_pads[0];
|
||||
const index_t InLeftPadW = input_left_pads[1];
|
||||
|
||||
const index_t InRightPadH = input_right_pads[0];
|
||||
const index_t InRightPadW = input_right_pads[1];
|
||||
|
||||
const auto in_n_hi_wi_c_grid_desc =
|
||||
make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
|
||||
|
||||
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_hi_wi_c_grid_desc,
|
||||
make_tuple(make_pass_through_transform(N),
|
||||
make_pad_transform(Hi, InLeftPadH, InRightPadH),
|
||||
make_pad_transform(Wi, InLeftPadW, InRightPadW),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
|
||||
|
||||
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
|
||||
in_n_hip_wip_c_grid_desc,
|
||||
make_tuple(
|
||||
make_pass_through_transform(N),
|
||||
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
|
||||
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
|
||||
make_pass_through_transform(C)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
|
||||
|
||||
const auto in_gemmk_gemmmraw_grid_desc =
|
||||
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
|
||||
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
|
||||
make_merge_transform(make_tuple(N, Ho, Wo))),
|
||||
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}));
|
||||
|
||||
const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
|
||||
in_gemmk_gemmmraw_grid_desc,
|
||||
make_tuple(make_unmerge_transform(make_tuple(gemm_k0, GemmK1Number)),
|
||||
make_pass_through_transform(gemm_m)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}),
|
||||
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
|
||||
|
||||
// in_gemmk0_gemmm_gemmk1_grid_desc
|
||||
return transform_tensor_descriptor(
|
||||
in_gemmk0_gemmmraw_gemmk1_grid_desc,
|
||||
make_tuple(make_pass_through_transform(gemm_k0),
|
||||
make_right_pad_transform(gemm_m, gemm_m_pad),
|
||||
make_pass_through_transform(GemmK1Number)),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
|
||||
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
|
||||
}
|
||||
}
|
||||
|
||||
static index_t GetGemmMRaw(ck::index_t N,
|
||||
const std::vector<ck::index_t>& output_spatial_lengths)
|
||||
{
|
||||
return N * std::accumulate(std::begin(output_spatial_lengths),
|
||||
std::end(output_spatial_lengths),
|
||||
1,
|
||||
std::multiplies<ck::index_t>());
|
||||
}
|
||||
|
||||
static index_t GetGemmK(ck::index_t C, const std::vector<ck::index_t>& filter_spatial_lengths)
|
||||
{
|
||||
return C * std::accumulate(std::begin(filter_spatial_lengths),
|
||||
std::end(filter_spatial_lengths),
|
||||
1,
|
||||
std::multiplies<ck::index_t>());
|
||||
}
|
||||
|
||||
static auto
|
||||
MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t C,
|
||||
std::vector<ck::index_t> input_spatial_lengths,
|
||||
std::vector<ck::index_t> filter_spatial_lengths,
|
||||
std::vector<ck::index_t> output_spatial_lengths,
|
||||
std::vector<ck::index_t> conv_filter_strides,
|
||||
std::vector<ck::index_t> conv_filter_dilations,
|
||||
std::vector<ck::index_t> input_left_pads,
|
||||
std::vector<ck::index_t> input_right_pads)
|
||||
{
|
||||
using namespace ck;
|
||||
|
||||
const index_t GemmMRaw = GetGemmMRaw(N, output_spatial_lengths);
|
||||
const index_t GemmN = K;
|
||||
const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
|
||||
|
||||
const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
|
||||
|
||||
assert(GemmK % GemmK1Number == 0);
|
||||
|
||||
// C = A^T*B
|
||||
// A:
|
||||
const auto in_gemmk0_gemmm_gemmk1_grid_desc =
|
||||
GetInputTensorDescriptor<NumDimSpatial>(N,
|
||||
C,
|
||||
GemmMRaw,
|
||||
GemmK,
|
||||
GemmMPad,
|
||||
input_spatial_lengths,
|
||||
filter_spatial_lengths,
|
||||
output_spatial_lengths,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
// B:
|
||||
const auto wei_gemmk0_gemmn_gemmk1_grid_desc = GetWeightTensorDescriptor(GemmN, GemmK);
|
||||
// C:
|
||||
const auto out_gemmm_gemmn_grid_desc = GetOutputTensorDescriptor(GemmMRaw, GemmN, GemmMPad);
|
||||
|
||||
return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
|
||||
wei_gemmk0_gemmn_gemmk1_grid_desc,
|
||||
out_gemmm_gemmn_grid_desc);
|
||||
}
|
||||
|
||||
template <ck::index_t NDim, typename std::enable_if<NDim == 1, bool>::type = false>
|
||||
static auto GetABCGridDesc()
|
||||
{
|
||||
return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
|
||||
1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1});
|
||||
}
|
||||
|
||||
template <ck::index_t NDim, typename std::enable_if<NDim == 2, bool>::type = false>
|
||||
static auto GetABCGridDesc()
|
||||
{
|
||||
return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
|
||||
1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1});
|
||||
}
|
||||
|
||||
using ABCGridDescs = decltype(GetABCGridDesc<NumDimSpatial>());
|
||||
|
||||
using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
|
||||
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
|
||||
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
|
||||
|
||||
// GridwiseGemm
|
||||
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
|
||||
BlockSize,
|
||||
ABDataType, // TODO: distinguish A/B datatype
|
||||
AccDataType,
|
||||
CDataType,
|
||||
InMemoryDataOperationEnum_t::Set,
|
||||
AGridDesc_K0_M_K1,
|
||||
BGridDesc_K0_N_K1,
|
||||
CGridDesc_M_N,
|
||||
InElementwiseOperation,
|
||||
WeiElementwiseOperation,
|
||||
OutElementwiseOperation,
|
||||
MPerBlock,
|
||||
NPerBlock,
|
||||
K0PerBlock,
|
||||
MPerXDL,
|
||||
NPerXDL,
|
||||
K1,
|
||||
MXdlPerWave,
|
||||
NXdlPerWave,
|
||||
ABlockTransferThreadClusterLengths_K0_M_K1,
|
||||
Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
|
||||
Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
|
||||
2, // ABlockTransferSrcVectorDim,
|
||||
ABlockTransferSrcScalarPerVector,
|
||||
ABlockTransferDstScalarPerVector_K1,
|
||||
false, // AThreadTransferSrcResetCoordinateAfterRun,
|
||||
ABlockLdsAddExtraM,
|
||||
BBlockTransferThreadClusterLengths_K0_N_K1,
|
||||
Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
|
||||
Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
|
||||
2, // BBlockTransferSrcVectorDim,
|
||||
BBlockTransferSrcScalarPerVector,
|
||||
BBlockTransferDstScalarPerVector_K1,
|
||||
false, // BThreadTransferSrcResetCoordinateAfterRun,
|
||||
BBlockLdsAddExtraN,
|
||||
Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
|
||||
7, // CThreadTransferSrcDstVectorDim,
|
||||
CThreadTransferDstScalarPerVector>;
|
||||
|
||||
// Argument
|
||||
struct Argument : public BaseArgument
|
||||
{
|
||||
Argument(const InDataType* p_in_grid,
|
||||
const WeiDataType* p_wei_grid,
|
||||
OutDataType* p_out_grid,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t C,
|
||||
std::vector<ck::index_t> input_spatial_lengths,
|
||||
std::vector<ck::index_t> filter_spatial_lengths,
|
||||
std::vector<ck::index_t> output_spatial_lengths,
|
||||
std::vector<ck::index_t> conv_filter_strides,
|
||||
std::vector<ck::index_t> conv_filter_dilations,
|
||||
std::vector<ck::index_t> input_left_pads,
|
||||
std::vector<ck::index_t> input_right_pads,
|
||||
ck::index_t M01,
|
||||
ck::index_t N01,
|
||||
InElementwiseOperation in_element_op,
|
||||
WeiElementwiseOperation wei_element_op,
|
||||
OutElementwiseOperation out_element_op)
|
||||
: p_a_grid_{p_in_grid},
|
||||
p_b_grid_{p_wei_grid},
|
||||
p_c_grid_{p_out_grid},
|
||||
a_grid_desc_k0_m_k1_{},
|
||||
b_grid_desc_k0_n_k1_{},
|
||||
c_grid_desc_m_n_{},
|
||||
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
|
||||
block_2_ctile_map_{},
|
||||
M01_{M01},
|
||||
N01_{N01},
|
||||
in_element_op_{in_element_op},
|
||||
wei_element_op_{wei_element_op},
|
||||
out_element_op_{out_element_op},
|
||||
Conv_N_{N},
|
||||
Conv_K_{K},
|
||||
Conv_C_{C},
|
||||
filter_spatial_lengths_{filter_spatial_lengths},
|
||||
conv_filter_strides_{conv_filter_strides},
|
||||
input_left_pads_{input_left_pads},
|
||||
input_right_pads_{input_right_pads}
|
||||
{
|
||||
const auto descs =
|
||||
DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
|
||||
K,
|
||||
C,
|
||||
input_spatial_lengths,
|
||||
filter_spatial_lengths,
|
||||
output_spatial_lengths,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads);
|
||||
|
||||
a_grid_desc_k0_m_k1_ = descs[I0];
|
||||
b_grid_desc_k0_n_k1_ = descs[I1];
|
||||
c_grid_desc_m_n_ = descs[I2];
|
||||
|
||||
if(GridwiseGemm::CheckValidity(
|
||||
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
|
||||
{
|
||||
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
|
||||
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
|
||||
|
||||
block_2_ctile_map_ =
|
||||
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
|
||||
}
|
||||
}
|
||||
|
||||
// private:
|
||||
const ADataType* p_a_grid_;
|
||||
const BDataType* p_b_grid_;
|
||||
CDataType* p_c_grid_;
|
||||
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
|
||||
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
|
||||
CGridDesc_M_N c_grid_desc_m_n_;
|
||||
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
|
||||
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
|
||||
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
|
||||
index_t M01_;
|
||||
index_t N01_;
|
||||
InElementwiseOperation in_element_op_;
|
||||
WeiElementwiseOperation wei_element_op_;
|
||||
OutElementwiseOperation out_element_op_;
|
||||
// for checking IsSupportedArgument()
|
||||
index_t Conv_N_;
|
||||
index_t Conv_K_;
|
||||
index_t Conv_C_;
|
||||
std::vector<index_t> filter_spatial_lengths_;
|
||||
std::vector<index_t> conv_filter_strides_;
|
||||
std::vector<index_t> input_left_pads_;
|
||||
std::vector<index_t> input_right_pads_;
|
||||
};
|
||||
|
||||
// Invoker
|
||||
struct Invoker : public BaseInvoker
|
||||
{
|
||||
using Argument = DeviceOp::Argument;
|
||||
|
||||
float Run(const Argument& arg, int nrepeat = 1)
|
||||
{
|
||||
{
|
||||
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
|
||||
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
|
||||
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
|
||||
|
||||
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
|
||||
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
|
||||
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
|
||||
|
||||
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
|
||||
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
|
||||
}
|
||||
|
||||
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
|
||||
arg.b_grid_desc_k0_n_k1_,
|
||||
arg.c_grid_desc_m_n_,
|
||||
arg.M01_,
|
||||
arg.N01_))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
|
||||
}
|
||||
|
||||
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
|
||||
|
||||
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
|
||||
|
||||
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
|
||||
|
||||
float ave_time = 0;
|
||||
|
||||
if(has_main_k0_block_loop)
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdlops_v2r3<
|
||||
GridwiseGemm,
|
||||
ADataType, // TODO: distiguish A/B datatype
|
||||
CDataType,
|
||||
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
|
||||
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
|
||||
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
|
||||
InElementwiseOperation,
|
||||
WeiElementwiseOperation,
|
||||
OutElementwiseOperation,
|
||||
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
|
||||
true>;
|
||||
|
||||
ave_time = launch_and_time_kernel(kernel,
|
||||
nrepeat,
|
||||
dim3(grid_size),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
arg.p_a_grid_,
|
||||
arg.p_b_grid_,
|
||||
arg.p_c_grid_,
|
||||
arg.a_grid_desc_k0_m_k1_,
|
||||
arg.b_grid_desc_k0_n_k1_,
|
||||
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
|
||||
arg.in_element_op_,
|
||||
arg.wei_element_op_,
|
||||
arg.out_element_op_,
|
||||
arg.block_2_ctile_map_);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto kernel = kernel_gemm_xdlops_v2r3<
|
||||
GridwiseGemm,
|
||||
ADataType, // TODO: distiguish A/B datatype
|
||||
CDataType,
|
||||
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
|
||||
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
|
||||
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
|
||||
InElementwiseOperation,
|
||||
WeiElementwiseOperation,
|
||||
OutElementwiseOperation,
|
||||
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
|
||||
false>;
|
||||
|
||||
ave_time = launch_and_time_kernel(kernel,
|
||||
nrepeat,
|
||||
dim3(grid_size),
|
||||
dim3(BlockSize),
|
||||
0,
|
||||
arg.p_a_grid_,
|
||||
arg.p_b_grid_,
|
||||
arg.p_c_grid_,
|
||||
arg.a_grid_desc_k0_m_k1_,
|
||||
arg.b_grid_desc_k0_n_k1_,
|
||||
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
|
||||
arg.in_element_op_,
|
||||
arg.wei_element_op_,
|
||||
arg.out_element_op_,
|
||||
arg.block_2_ctile_map_);
|
||||
}
|
||||
|
||||
return ave_time;
|
||||
}
|
||||
|
||||
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
|
||||
{
|
||||
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
|
||||
}
|
||||
};
|
||||
|
||||
static constexpr bool IsValidCompilationParameter()
|
||||
{
|
||||
// TODO: properly implement this check
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool IsSupportedArgument(const Argument& arg)
|
||||
{
|
||||
if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
|
||||
{
|
||||
// check if it's 1x1, stride=1 conv
|
||||
for(ck::index_t i = 0; i < NumDimSpatial; ++i)
|
||||
{
|
||||
if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
|
||||
arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
else if constexpr(ConvForwardSpecialization ==
|
||||
ConvolutionForwardSpecialization_t::Filter1x1Pad0)
|
||||
{
|
||||
// check if it's 1x1 conv
|
||||
for(ck::index_t i = 0; i < NumDimSpatial; ++i)
|
||||
{
|
||||
if(!(arg.filter_spatial_lengths_[i] == 1 && arg.input_left_pads_[i] == 0 &&
|
||||
arg.input_right_pads_[i] == 0))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// vector load A/B matrix from global memory
|
||||
if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
|
||||
arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
|
||||
arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// vector store C matrix into global memory
|
||||
if(!(arg.Conv_K_ % CThreadTransferDstScalarPerVector == 0))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
// Gridwise GEMM size
|
||||
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
|
||||
arg.b_grid_desc_k0_n_k1_,
|
||||
arg.c_grid_desc_m_n_,
|
||||
arg.M01_,
|
||||
arg.N01_);
|
||||
}
|
||||
|
||||
bool IsSupportedArgument(const BaseArgument* p_arg) override
|
||||
{
|
||||
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
|
||||
}
|
||||
|
||||
static auto MakeArgument(const InDataType* p_in_grid,
|
||||
const WeiDataType* p_wei_grid,
|
||||
OutDataType* p_out_grid,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t C,
|
||||
std::vector<ck::index_t> input_spatial_lengths,
|
||||
std::vector<ck::index_t> filter_spatial_lengths,
|
||||
std::vector<ck::index_t> output_spatial_lengths,
|
||||
std::vector<ck::index_t> conv_filter_strides,
|
||||
std::vector<ck::index_t> conv_filter_dilations,
|
||||
std::vector<ck::index_t> input_left_pads,
|
||||
std::vector<ck::index_t> input_right_pads,
|
||||
InElementwiseOperation in_element_op,
|
||||
WeiElementwiseOperation wei_element_op,
|
||||
OutElementwiseOperation out_element_op)
|
||||
{
|
||||
return Argument{p_in_grid,
|
||||
p_wei_grid,
|
||||
p_out_grid,
|
||||
N,
|
||||
K,
|
||||
C,
|
||||
input_spatial_lengths,
|
||||
filter_spatial_lengths,
|
||||
output_spatial_lengths,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
1,
|
||||
1,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op};
|
||||
}
|
||||
|
||||
static auto MakeInvoker() { return Invoker{}; }
|
||||
|
||||
std::unique_ptr<BaseArgument>
|
||||
MakeArgumentPointer(const void* p_in_grid,
|
||||
const void* p_wei_grid,
|
||||
void* p_out_grid,
|
||||
ck::index_t N,
|
||||
ck::index_t K,
|
||||
ck::index_t C,
|
||||
std::vector<ck::index_t> input_spatial_lengths,
|
||||
std::vector<ck::index_t> filter_spatial_lengths,
|
||||
std::vector<ck::index_t> output_spatial_lengths,
|
||||
std::vector<ck::index_t> conv_filter_strides,
|
||||
std::vector<ck::index_t> conv_filter_dilations,
|
||||
std::vector<ck::index_t> input_left_pads,
|
||||
std::vector<ck::index_t> input_right_pads,
|
||||
InElementwiseOperation in_element_op,
|
||||
WeiElementwiseOperation wei_element_op,
|
||||
OutElementwiseOperation out_element_op) override
|
||||
{
|
||||
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
|
||||
static_cast<const WeiDataType*>(p_wei_grid),
|
||||
static_cast<OutDataType*>(p_out_grid),
|
||||
N,
|
||||
K,
|
||||
C,
|
||||
input_spatial_lengths,
|
||||
filter_spatial_lengths,
|
||||
output_spatial_lengths,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
1,
|
||||
1,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op);
|
||||
}
|
||||
|
||||
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
|
||||
{
|
||||
return std::make_unique<Invoker>(Invoker{});
|
||||
}
|
||||
|
||||
std::string GetTypeString() const override
|
||||
{
|
||||
auto str = std::stringstream();
|
||||
|
||||
// clang-format off
|
||||
str << "DeviceConv" << std::to_string(NumDimSpatial)
|
||||
<< "DFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
|
||||
<< "<"
|
||||
<< BlockSize << ", "
|
||||
<< MPerBlock << ", "
|
||||
<< NPerBlock << ", "
|
||||
<< K0PerBlock
|
||||
<< ">";
|
||||
// clang-format on
|
||||
|
||||
return str.str();
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
#endif
|
||||
@@ -12,37 +12,77 @@ namespace gemm {
|
||||
|
||||
struct RowMajor : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "RowMajor";
|
||||
};
|
||||
|
||||
struct ColumnMajor : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "ColumnMajor";
|
||||
};
|
||||
} // namespace gemm
|
||||
|
||||
namespace convolution {
|
||||
|
||||
// 1D Conv
|
||||
struct NWC : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NWC";
|
||||
};
|
||||
|
||||
struct KXC : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "KXC";
|
||||
};
|
||||
|
||||
struct NWK : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NWK";
|
||||
};
|
||||
|
||||
struct NCW : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NCW";
|
||||
};
|
||||
|
||||
struct KCX : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "KCX";
|
||||
};
|
||||
|
||||
struct NKW : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NKW";
|
||||
};
|
||||
|
||||
// 2D Conv
|
||||
struct NHWC : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NHWC";
|
||||
};
|
||||
|
||||
struct KYXC : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "KYXC";
|
||||
};
|
||||
|
||||
struct NHWK : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NHWK";
|
||||
};
|
||||
|
||||
struct NCHW : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NCHW";
|
||||
};
|
||||
|
||||
struct KCYX : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "KCYX";
|
||||
};
|
||||
|
||||
struct NKHW : public BaseTensorLayout
|
||||
{
|
||||
static constexpr const char* name = "NKHW";
|
||||
};
|
||||
|
||||
struct NDHWC : public BaseTensorLayout
|
||||
@@ -59,6 +99,15 @@ struct NDHWK : public BaseTensorLayout
|
||||
|
||||
} // namespace convolution
|
||||
|
||||
template <
|
||||
typename Layout,
|
||||
typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>
|
||||
std::ostream& operator<<(std::ostream& os, const Layout&)
|
||||
{
|
||||
os << Layout::name;
|
||||
return os;
|
||||
}
|
||||
|
||||
} // namespace tensor_layout
|
||||
} // namespace ck
|
||||
#endif
|
||||
|
||||
@@ -0,0 +1,112 @@
|
||||
#include <stdlib.h>
|
||||
#include "config.hpp"
|
||||
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace device_conv1d_fwd_instance {
|
||||
|
||||
using F32 = float;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization_t::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization_t::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0;
|
||||
|
||||
//------------------------------------------------------------------------------
|
||||
// Conv1D
|
||||
//------------------------------------------------------------------------------
|
||||
|
||||
// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
|
||||
using device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 1, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_conv1d_fwd_xdl_nwc_kxc_nwk_1x1_p0_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 1, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_conv1d_fwd_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 1, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances(
|
||||
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances, device_conv1d_fwd_xdl_nwc_kxc_nwk_f32_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_conv1d_fwd_xdl_nwc_kxc_nwk_1x1_p0_f32_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_conv1d_fwd_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances{});
|
||||
}
|
||||
|
||||
} // namespace device_conv1d_fwd_instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,6 +1,6 @@
|
||||
#include <stdlib.h>
|
||||
#include "config.hpp"
|
||||
#include "device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
|
||||
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "device_operation_instance.hpp"
|
||||
|
||||
@@ -28,67 +28,67 @@ static constexpr auto ConvFwd1x1S1P0 =
|
||||
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
|
||||
using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, 2, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, 2, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
//################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| NumDim| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
|
||||
//################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization|Spatial| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
|
||||
//################################################################| | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
|
||||
//################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>,
|
||||
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, 2, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 7, 1>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
|
||||
Reference in New Issue
Block a user