Skip to content

GitLab

Commit c31867bd authored by Chao Liu's avatar Chao Liu
Browse files

add non-padded

parent 5f0c56d0
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1040 additions and 15 deletions
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp
View file @ c31867bd
......@@ -39,7 +39,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc
wei_kcyx_device_buf.ToDevice(wei_kcyx.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 1
#if 0
// run-time variables
const auto in_n_c_hi_wi_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(to_multi_index(InDesc::GetLengths()));
......@@ -67,7 +67,110 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc
const auto in_right_pads = sequence_to_tuple_of_number(InRightPads{});
#endif
#if 0
#if 1
// cdata = 16, BlockSize = 64, 16x64x4
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 2;
constexpr index_t GemmNPerThread = 2;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t ThreadGemmDataPerReadM = 2;
constexpr index_t ThreadGemmDataPerReadN = 2;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<4, 1>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 1;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 2;
#elif 0
// cdata = 16, BlockSize = 64, 16x64x4
// GemmBBlockCopySrcDataPerRead_GemmN = 4
// GemmCThreadCopyDstDataPerWrite_GemmN1 = 2
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 2;
constexpr index_t GemmNPerThread = 2;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t ThreadGemmDataPerReadM = 2;
constexpr index_t ThreadGemmDataPerReadN = 2;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<1, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<4, 16>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 2;
#elif 0
// cdata = 32, BlockSize = 64, 16x128x4
// GemmBBlockCopySrcDataPerRead_GemmN = 4
// GemmCThreadCopyDstDataPerWrite_GemmN1 = 4
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 2;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t ThreadGemmDataPerReadM = 2;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<2, 32>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmN1 = 4;
#elif 0
// cdata = 64, BlockSize = 128, 32x256x8
constexpr index_t BlockSize = 128;
......@@ -75,14 +178,14 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t ThreadGemmDataPerReadM = 4;
constexpr index_t ThreadGemmDataPerReadN = 4;
......@@ -259,7 +362,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(InDesc
constexpr auto conv_driver =
#if 1
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nchw_kcyx_nkhw_pad
#elif 1
#elif 0
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nchw_kcyx_nkhw_no_pad
#elif 1
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nchw_kcyx_nkhw_1x1
......
driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp
View file @ c31867bd
......@@ -48,7 +48,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(InDesc
constexpr auto Y = WeiDesc::GetLengths()[I2];
constexpr auto X = WeiDesc::GetLengths()[I3];
#if 1
#if 0
// run-time variables
constexpr auto in_n_hi_wi_c_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(make_multi_index(N, Hi, Wi, C));
......@@ -110,6 +110,204 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(InDesc
out_nhwk_device_buf.ToDevice(out_nhwk.mData.data());
#if 0
// cdata = 16, BlockSize = 64, 16x64x4
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 64;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 2;
constexpr index_t GemmNPerThread = 2;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t ThreadGemmDataPerReadM = 2;
constexpr index_t ThreadGemmDataPerReadN = 2;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<4, 1>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 2;
#elif 0
// cdata = 32, BlockSize = 64, 16x128x4
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 2;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 8;
constexpr index_t ThreadGemmDataPerReadM = 2;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 2;
#elif 1
// cdata = 64, BlockSize = 64, 16x256x2
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 2;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 1;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t ThreadGemmDataPerReadM = 4;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<2, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 2;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 4;
#elif 0
// cdata = 64, BlockSize = 64, 16x256x4
constexpr index_t BlockSize = 64;
constexpr index_t GemmMPerBlock = 16;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 1;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t ThreadGemmDataPerReadM = 4;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 16>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<4, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 64>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 4;
#elif 0
// cdata = 64, BlockSize = 128, 32x256x4
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 32;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t ThreadGemmDataPerReadM = 4;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<1, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 32>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 1;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<4, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 128>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 4;
#elif 0
// cdata = 64, BlockSize = 128, 32x256x8
constexpr index_t BlockSize = 128;
constexpr index_t GemmMPerBlock = 32;
constexpr index_t GemmNPerBlock = 256;
constexpr index_t GemmKPerBlock = 8;
constexpr index_t GemmMPerThread = 4;
constexpr index_t GemmNPerThread = 4;
constexpr index_t GemmKPerThread = 1;
constexpr index_t GemmMLevel0Cluster = 2;
constexpr index_t GemmNLevel0Cluster = 2;
constexpr index_t GemmMLevel1Cluster = 2;
constexpr index_t GemmNLevel1Cluster = 16;
constexpr index_t ThreadGemmDataPerReadM = 4;
constexpr index_t ThreadGemmDataPerReadN = 4;
using GemmABlockTransferThreadSliceLengths_GemmK_GemmM = Sequence<2, 1>;
using GemmABlockTransferThreadClusterLengths_GemmK_GemmM = Sequence<4, 32>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK = 2;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmM = 1;
using GemmBBlockTransferThreadSliceLengths_GemmK_GemmN = Sequence<8, 2>;
using GemmBBlockTransferThreadClusterLengths_GemmK_GemmN = Sequence<1, 128>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmN = 1;
constexpr index_t GemmCThreadTransferDstScalarPerVector_GemmM1 = 4;
#elif 0
// cdata = 64, BlockSize = 256, 128x128x8
constexpr index_t BlockSize = 256;
......@@ -174,7 +372,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(InDesc
constexpr auto conv_driver =
#if 1
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nhwc_kyxc_nhwk_pad
#elif 1
#elif 0
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nhwc_kyxc_nhwk_no_pad
#elif 1
DriverDynamicConvolutionForwardImplicitGemm_v4r4_nhwc_kyxc_nhwk_1x1
......
driver/src/conv_driver.cpp
View file @ c31867bd
......@@ -22,10 +22,80 @@ int main(int argc, char* argv[])
#if 0
constexpr index_t N = 1;
constexpr index_t C = 32;
constexpr index_t C = 4;
constexpr index_t HI = 1080;
constexpr index_t WI = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 4;
constexpr index_t HI = 540;
constexpr index_t WI = 960;
constexpr index_t K = 32;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 4;
constexpr index_t HI = 270;
constexpr index_t WI = 480;
constexpr index_t K = 16;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 1
constexpr index_t N = 1;
constexpr index_t C = 2;
constexpr index_t HI = 1080;
constexpr index_t WI = 1920;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 0
constexpr index_t N = 1;
constexpr index_t C = 4;
constexpr index_t HI = 540;
constexpr index_t WI = 960;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 1
constexpr index_t N = 1;
constexpr index_t C = 4;
constexpr index_t HI = 270;
constexpr index_t WI = 480;
constexpr index_t K = 16;
constexpr index_t Y = 3;
constexpr index_t X = 3;
......@@ -616,7 +686,7 @@ int main(int argc, char* argv[])
LeftPads{},
RightPads{},
nrepeat);
#elif 1
#elif 0
device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
......
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