Skip to content

GitLab

Commit cd100621 authored by Chao Liu's avatar Chao Liu
Browse files

refactor

parent d8c89b68
Hide whitespace changes
Inline Side-by-side
Showing with 135 additions and 129 deletions
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
View file @ cd100621
......@@ -487,71 +487,74 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(
in_left_pads,
in_right_pads);
float ave_time = launch_kernel_dynamic_gemm_v1<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerThread,
GemmNPerThread,
GemmKPerThread,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmABlockTransferThreadSliceLengths_GemmK_GemmM,
GemmABlockTransferThreadClusterLengths_GemmK_GemmM,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_GemmM,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK_GemmN,
GemmBBlockTransferThreadClusterLengths_GemmK_GemmN,
Sequence<0, 1>,
Sequence<0, 1>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmN,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
3,
GemmCThreadTransferDstScalarPerVector_GemmN1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (float)calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
for(index_t i = 0; i < 5; ++i)
{
float ave_time = launch_kernel_dynamic_gemm_v1<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerThread,
GemmNPerThread,
GemmKPerThread,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmABlockTransferThreadSliceLengths_GemmK_GemmM,
GemmABlockTransferThreadClusterLengths_GemmK_GemmM,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_GemmM,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK_GemmN,
GemmBBlockTransferThreadClusterLengths_GemmK_GemmN,
Sequence<0, 1>,
Sequence<0, 1>,
1,
GemmBBlockTransferSrcScalarPerVector_GemmN,
GemmBBlockTransferDstScalarPerVector_GemmN,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
3,
GemmCThreadTransferDstScalarPerVector_GemmN1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_c_y_x_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (float)calculate_convolution_flops(
in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());
......
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp
View file @ cd100621
......@@ -393,70 +393,73 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(
in_left_pads,
in_right_pads);
float ave_time = launch_kernel_dynamic_gemm_v1<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerThread,
GemmNPerThread,
GemmKPerThread,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmABlockTransferThreadSliceLengths_GemmK_GemmM,
GemmABlockTransferThreadClusterLengths_GemmK_GemmM,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_GemmM,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK_GemmN,
GemmBBlockTransferThreadClusterLengths_GemmK_GemmN,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmBBlockTransferSrcScalarPerVector_GemmK,
GemmBBlockTransferDstScalarPerVector_GemmN,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
1,
GemmCThreadTransferDstScalarPerVector_GemmM1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (float)(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
for(index_t i = 0; i < 5; ++i)
{
float ave_time = launch_kernel_dynamic_gemm_v1<
BlockSize,
typename vector_type<TInWei, InWeiVectorSize>::type,
TAcc,
TOut,
InMemoryDataOperation::Set,
decltype(descs[I0]),
decltype(descs[I1]),
decltype(descs[I2]),
decltype(descs[I3]),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerThread,
GemmNPerThread,
GemmKPerThread,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmABlockTransferThreadSliceLengths_GemmK_GemmM,
GemmABlockTransferThreadClusterLengths_GemmK_GemmM,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmABlockTransferSrcScalarPerVector_GemmK,
GemmABlockTransferDstScalarPerVector_GemmM,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK_GemmN,
GemmBBlockTransferThreadClusterLengths_GemmK_GemmN,
Sequence<1, 0>,
Sequence<1, 0>,
0,
GemmBBlockTransferSrcScalarPerVector_GemmK,
GemmBBlockTransferDstScalarPerVector_GemmN,
false, // don't move back src coordinate after threadwise copy, which will be fused with
// MoveSrcSliceWindow() to save addr computation
Sequence<2, 3, 0, 1>,
1,
GemmCThreadTransferDstScalarPerVector_GemmM1,
decltype(descs[I4]),
decltype(descs[I5]),
decltype(descs[I6]),
decltype(descs[I7]),
decltype(descs[I8])>(static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<typename vector_type<TInWei, InWeiVectorSize>::type*>(
in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
descs[I0],
descs[I1],
descs[I2],
descs[I3],
descs[I4],
descs[I5],
descs[I6],
descs[I7],
descs[I8],
nrepeat);
float perf = (float)(std::size_t(2) * N * K * Ho * Wo * C * Y * X) /
(std::size_t(1000) * 1000 * 1000) / ave_time;
std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
}
// copy result back to host
out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data());
......
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