Skip to content

GitLab

Commit cf18e9a3 authored by letaoqin's avatar letaoqin
Browse files

add instance for dl multiple d

parent 0601203a
Expand all Show whitespace changes
Inline Side-by-side
Showing with 181 additions and 292 deletions
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ cf18e9a3
...@@ -620,7 +620,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK ...@@ -620,7 +620,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
throw std::runtime_error( throw std::runtime_error(
"wrong! DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK has invalid setting"); "wrong! DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK has invalid setting");
} }
std::cout << stream_config.log_level_ << std::endl;
const index_t grid_size = const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0), GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0),
arg.e_grid_desc_m_n_.GetLength(I1)) * arg.e_grid_desc_m_n_.GetLength(I1)) *
...@@ -728,7 +728,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK ...@@ -728,7 +728,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0)) if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
{ {
std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
<< " ConvStride = " << ConvStride << " LeftPad = " << LeftPad << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
<< " RightPad = " << RightPad << std::endl; << " RightPad = " << RightPad << std::endl;
return false; return false;
...@@ -747,7 +747,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK ...@@ -747,7 +747,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
if(!(X == 1 && LeftPad == 0 && RightPad == 0)) if(!(X == 1 && LeftPad == 0 && RightPad == 0))
{ {
std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
<< " LeftPad = " << LeftPad << " RightPad = " << RightPad << " LeftPad = " << LeftPad << " RightPad = " << RightPad
<< std::endl; << std::endl;
return false; return false;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
View file @ cf18e9a3
...@@ -251,8 +251,8 @@ struct GridwiseGemmDlMultipleD_km_kn_mn ...@@ -251,8 +251,8 @@ struct GridwiseGemmDlMultipleD_km_kn_mn
DsGridPointer p_ds_grid, DsGridPointer p_ds_grid,
FloatC* __restrict__ p_c_grid, FloatC* __restrict__ p_c_grid,
FloatAB* __restrict__ p_shared_block, FloatAB* __restrict__ p_shared_block,
const AElementwiseOperation& , const AElementwiseOperation&,
const BElementwiseOperation& , const BElementwiseOperation&,
const CDEElementwiseOperation& cde_element_op, const CDEElementwiseOperation& cde_element_op,
const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1, const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1, const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ cf18e9a3
...@@ -131,6 +131,47 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances( ...@@ -131,6 +131,47 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// grouped conv2d forward, NHWGC/GKYXC/NHWGK // grouped conv2d forward, NHWGC/GKYXC/NHWGK
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances( void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
...@@ -273,11 +314,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -273,11 +314,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<WeiDataType, ck::bhalf_t> &&
...@@ -289,6 +332,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -289,6 +332,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
} }
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> && else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp deleted 100644 → 0
View file @ 0601203a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwd<2,
GNHWC,
GKYXC,
GNHWK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwd<2,
GNHWC,
GKYXC,
GNHWK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwd<2,
GNHWC,
GKYXC,
GNHWK,
int8_t,
int8_t,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename OutDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwd<
NumDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceGroupedConvFwd<NumDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profile_grouped_conv_fwd_impl.hpp
View file @ cf18e9a3
...@@ -14,9 +14,6 @@ ...@@ -14,9 +14,6 @@
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
...@@ -201,8 +198,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification, ...@@ -201,8 +198,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
} }
}; };
// xdl
{
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial, using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
InLayout, InLayout,
WeiLayout, WeiLayout,
...@@ -220,12 +215,12 @@ bool profile_grouped_conv_fwd_impl(int do_verification, ...@@ -220,12 +215,12 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl; std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
for(auto& op_ptr : op_ptrs) for(auto& op_ptr : op_ptrs)
{ {
auto argument_ptr = op_ptr->MakeArgumentPointer( auto argument_ptr =
in_device_buf.GetDeviceBuffer(), op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(), wei_device_buf.GetDeviceBuffer(),
std::array<const void*, 0>{}, std::array<const void*, 0>{},
out_device_buf.GetDeviceBuffer(), out_device_buf.GetDeviceBuffer(),
...@@ -247,47 +242,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification, ...@@ -247,47 +242,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
run_impl(op_ptr, argument_ptr); run_impl(op_ptr, argument_ptr);
} }
}
// dl
{
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwd<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "dl found " << op_ptrs.size() << " instances" << std::endl;
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
run_impl(op_ptr, argument_ptr);
}
}
std::cout << "Best configuration parameters:" std::cout << "Best configuration parameters:"
<< "\nname: " << best_op_name << "\navg_time: " << best_avg_time << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment