Skip to content

GitLab

"...composable_kernel_onnx.git" did not exist on "5c7cec11159d22636dd4c1119e7e430d156a8df7"
Unverified Commit a2ad6d35 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

refactor dynamic xdlops iGemm (#13) 

* xdlops refactor

* fixed commnt

* clean xdlops_gemm

* add make c into xldops-gemm

* change mfma_info

* refactor xdlops, hide c desc

* clean

* clean

* clean

* apply hacks changes to v4r4r4_nhwc

* rename hacks and use single stage adapter

* enable fp16 mfma
parent ba6f79a7
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 690 additions and 1227 deletions
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp
View file @ a2ad6d35
...@@ -48,10 +48,10 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -48,10 +48,10 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths); const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
#if 1 #if 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16 // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256; constexpr index_t GemmMPerBlock = 128;
constexpr index_t GemmNPerBlock = 128; constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4; constexpr index_t GemmKPerBlock = 4;
...@@ -59,10 +59,10 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -59,10 +59,10 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
constexpr index_t GemmNPerWave = 32; constexpr index_t GemmNPerWave = 32;
constexpr index_t GemmK1 = 8; constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 4; constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 2; constexpr index_t NRepeat = 2;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>; using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 2, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>; using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8; constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
...@@ -106,22 +106,22 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw( ...@@ -106,22 +106,22 @@ void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw(
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{})); Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_step_hacks = constexpr auto out_m0_m1_m2_n_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}, Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}, Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}), Sequence<0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{}, make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}, Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}, Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{})); Sequence<0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{}; Sequence<0, 0, 0, 0, 0>{};
......
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp deleted 100644 → 0
View file @ ba6f79a7
#include <unistd.h>
#include "device.hpp"
#include "host_tensor.hpp"
#include "transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp"
#include "driver_gemm_xdlops_v2r2.hpp"
template <typename TInWei,
typename TAcc,
typename TOut,
typename InLengths,
typename WeiLengths,
typename OutLengths,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk(
const InLengths& in_n_hi_wi_c_lengths,
const WeiLengths& wei_k_y_x_c_lengths,
const OutLengths& out_n_ho_wo_k_lengths,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
const Tensor<TInWei>& in_n_hi_wi_c,
const Tensor<TInWei>& wei_k_y_x_c,
Tensor<TOut>& out_n_ho_wo_k,
ck::index_t nrepeat)
{
using namespace ck;
std::cout << __func__ << std::endl;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
DeviceMem in_n_hi_wi_c_device_buf(sizeof(TInWei) * in_n_hi_wi_c.mDesc.GetElementSpace());
DeviceMem wei_k_y_x_c_device_buf(sizeof(TInWei) * wei_k_y_x_c.mDesc.GetElementSpace());
DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
#if 1
// [M, N, K0, K1] = [256, 128, 4, 4] for fp32
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 4;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 4>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 4>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 4;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#elif 1
// [M, N, K0, K1] = [256, 128, 4, 8] for fp16
constexpr index_t BlockSize = 256;
constexpr index_t GemmMPerBlock = 256;
constexpr index_t GemmNPerBlock = 128;
constexpr index_t GemmKPerBlock = 4;
constexpr index_t GemmMPerWave = 64;
constexpr index_t GemmNPerWave = 64;
constexpr index_t GemmK1 = 8;
constexpr index_t MRepeat = 2;
constexpr index_t NRepeat = 1;
using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8>;
using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 8;
using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 2, 8>;
using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmK1 = 8;
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
#endif
const auto descs =
transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(wei_k_y_x_c_desc,
in_n_hi_wi_c_desc,
out_n_ho_wo_k_desc,
conv_strides,
conv_dilations,
in_left_pads,
in_right_pads,
Number<GemmK1>{});
const auto wei_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
const auto in_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
const auto out_gemmm_gemmn_grid_desc = descs[I2];
// HACK: hacks that control index calculation when iterating over A, B, C matrix
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}),
make_tuple(
Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}, Sequence<0, 0, 0, 0, 0>{}));
constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{}));
constexpr auto out_m0_m1_m2_n_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 1, 0, 0>{}),
make_tuple(Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 0, 0, 0>{},
Sequence<0, 0, 2, 0, 0>{}));
constexpr auto wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0>{};
constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
for(index_t i = 0; i < 5; ++i)
{
float ave_time = driver_gemm_xdlops_v2r2<
BlockSize,
TInWei,
TAcc,
TOut,
InMemoryDataOperationEnum_t::Set,
decltype(wei_gemmk0_gemmm_gemmk1_grid_desc),
decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
decltype(out_gemmm_gemmn_grid_desc),
GemmMPerBlock,
GemmNPerBlock,
GemmKPerBlock,
GemmMPerWave,
GemmNPerWave,
MRepeat,
NRepeat,
GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmABlockTransferSrcScalarPerVector_GemmK1,
GemmABlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
Sequence<1, 0, 2>,
Sequence<1, 0, 2>,
2,
GemmBBlockTransferSrcScalarPerVector_GemmK1,
GemmBBlockTransferDstScalarPerVector_GemmK1,
false, // don't move back src coordinate after threadwise copy
Sequence<2, 3, 0, 1>,
2,
GemmCThreadTransferDstScalarPerVector,
decltype(wei_gemmk0_gemmm_gemmk1_grid_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
decltype(out_m0_m1_m2_n_grid_step_hacks),
decltype(wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks)>(
static_cast<TInWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
wei_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmk0_gemmn_gemmk1_grid_desc,
out_gemmm_gemmn_grid_desc,
wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_step_hacks,
out_m0_m1_m2_n_grid_step_hacks,
wei_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
nrepeat);
{
const auto N = out_n_ho_wo_k_lengths[I0];
const auto K = out_n_ho_wo_k_lengths[I3];
const auto C = wei_k_y_x_c_lengths[I3];
const auto Ho = out_n_ho_wo_k_lengths[I1];
const auto Wo = out_n_ho_wo_k_lengths[I2];
const auto Y = wei_k_y_x_c_lengths[I1];
const auto X = wei_k_y_x_c_lengths[I2];
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());
}
host/driver_offline/include/device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp
View file @ a2ad6d35
...@@ -250,22 +250,22 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk( ...@@ -250,22 +250,22 @@ void device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1 Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
constexpr auto out_m0_m1_m2_n_grid_step_hacks = constexpr auto out_m0_m1_m2_n_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: MRepeat make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: MRepeat
Sequence<0, 0, 0, 0, 0>{}, // 1+: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: NRepeat
Sequence<0, 0, 0, 0, 0>{}, // 2+: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: MWaves
Sequence<0, 0, 0, 0, 0>{}, // 3+: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: NWaves
Sequence<0, 0, 0, 0, 0>{}, // 4+: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M0
Sequence<0, 0, 0, 0, 0>{}, // 5+: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M1
Sequence<0, 0, 0, 0, 0>{}, // 6+: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M2
Sequence<0, 0, 0, 0, 0>{}), // 7+: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: MRepeat make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: MRepeat
Sequence<0, 0, 0, 0, 0>{}, // 1-: NRepeat Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: NRepeat
Sequence<0, 0, 0, 0, 0>{}, // 2-: MWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: MWaves
Sequence<0, 0, 0, 0, 0>{}, // 3-: NWaves Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: NWaves
Sequence<0, 0, 0, 0, 0>{}, // 4-: M0 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M0
Sequence<0, 0, 0, 0, 0>{}, // 5-: M1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M1
Sequence<0, 0, 0, 0, 0>{}, // 6-: M2 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M2
Sequence<0, 0, 0, 0, 0>{})); // 7-: N1 Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N1
constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks = constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{}; Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0>{};
......
host/driver_offline/include/driver_gemm_xdlops_v2r3.hpp
View file @ a2ad6d35
...@@ -129,9 +129,10 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid, ...@@ -129,9 +129,10 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting"); "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
} }
const auto c_m0_m1_m2_n_grid_desc = GridwiseGemm::MakeCM0M1M2NGridDescriptor(c_m_n_grid_desc); const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
using CM0M1M2NGridDesc = decltype(c_m0_m1_m2_n_grid_desc); using CM0N0M1N1M2M3M4N2GridDesc = decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
const auto c_block_cluster_adaptor = GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc); const auto c_block_cluster_adaptor = GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc);
...@@ -144,7 +145,7 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid, ...@@ -144,7 +145,7 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
FloatC, FloatC,
remove_reference_t<AK0MK1GridDesc>, remove_reference_t<AK0MK1GridDesc>,
remove_reference_t<BK0NK1GridDesc>, remove_reference_t<BK0NK1GridDesc>,
remove_reference_t<CM0M1M2NGridDesc>, remove_reference_t<CM0N0M1N1M2M3M4N2GridDesc>,
remove_reference_t<CBlockClusterAdaptor>>; remove_reference_t<CBlockClusterAdaptor>>;
#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE #if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
...@@ -158,18 +159,18 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid, ...@@ -158,18 +159,18 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
p_c_grid, p_c_grid,
a_k0_m_k1_grid_desc, a_k0_m_k1_grid_desc,
b_k0_n_k1_grid_desc, b_k0_n_k1_grid_desc,
c_m0_m1_m2_n_grid_desc, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_block_cluster_adaptor); c_block_cluster_adaptor);
#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER #elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
DeviceMem a_k0_m_k1_grid_desc_dev_buf(sizeof(AK0MK1GridDesc)); DeviceMem a_k0_m_k1_grid_desc_dev_buf(sizeof(AK0MK1GridDesc));
DeviceMem b_k0_n_k1_grid_desc_dev_buf(sizeof(BK0NK1GridDesc)); DeviceMem b_k0_n_k1_grid_desc_dev_buf(sizeof(BK0NK1GridDesc));
DeviceMem c_m0_m1_m2_n_grid_desc_dev_buf(sizeof(CM0M1M2NGridDesc)); DeviceMem c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf(sizeof(CM0N0M1N1M2M3M4N2GridDesc));
DeviceMem c_block_cluster_adaptor_dev_buf(sizeof(CBlockClusterAdaptor)); DeviceMem c_block_cluster_adaptor_dev_buf(sizeof(CBlockClusterAdaptor));
a_k0_m_k1_grid_desc_dev_buf.ToDevice(&a_k0_m_k1_grid_desc); a_k0_m_k1_grid_desc_dev_buf.ToDevice(&a_k0_m_k1_grid_desc);
b_k0_n_k1_grid_desc_dev_buf.ToDevice(&b_k0_n_k1_grid_desc); b_k0_n_k1_grid_desc_dev_buf.ToDevice(&b_k0_n_k1_grid_desc);
c_m0_m1_m2_n_grid_desc_dev_buf.ToDevice(&c_m0_m1_m2_n_grid_desc); c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.ToDevice(&c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
c_block_cluster_adaptor_dev_buf.ToDevice(&c_block_cluster_adaptor); c_block_cluster_adaptor_dev_buf.ToDevice(&c_block_cluster_adaptor);
float ave_time = launch_and_time_kernel( float ave_time = launch_and_time_kernel(
...@@ -183,7 +184,8 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid, ...@@ -183,7 +184,8 @@ __host__ float driver_gemm_xdlops_v2r3(const FloatAB* p_a_grid,
p_c_grid, p_c_grid,
cast_pointer_to_constant_address_space(a_k0_m_k1_grid_desc_dev_buf.GetDeviceBuffer()), cast_pointer_to_constant_address_space(a_k0_m_k1_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(b_k0_n_k1_grid_desc_dev_buf.GetDeviceBuffer()), cast_pointer_to_constant_address_space(b_k0_n_k1_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(c_m0_m1_m2_n_grid_desc_dev_buf.GetDeviceBuffer()), cast_pointer_to_constant_address_space(
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.GetDeviceBuffer()),
cast_pointer_to_constant_address_space(c_block_cluster_adaptor_dev_buf.GetDeviceBuffer())); cast_pointer_to_constant_address_space(c_block_cluster_adaptor_dev_buf.GetDeviceBuffer()));
#endif #endif
return ave_time; return ave_time;
......
host/driver_offline/src/conv_fwd_driver_offline.cpp
View file @ a2ad6d35
...@@ -24,8 +24,8 @@ ...@@ -24,8 +24,8 @@
#define USE_CONV_FWD_V4R4R2_NHWC 1 #define USE_CONV_FWD_V4R4R2_NHWC 1
#define USE_CONV_FWD_V6R1_NCHW 0 #define USE_CONV_FWD_V6R1_NCHW 0
#define USE_CONV_FWD_V5R1_NCHW 0 #define USE_CONV_FWD_V5R1_NCHW 0
#define USE_CONV_FWD_V4R4R2_XDL_NCHW 0 #define USE_CONV_FWD_V4R4R2_XDL_NCHW 1
#define USE_CONV_FWD_V4R4R4_XDL_NHWC 0 #define USE_CONV_FWD_V4R4R4_XDL_NHWC 1
enum ConvForwardAlgo enum ConvForwardAlgo
{ {
......
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