Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "dbfd7bd0276dadefa2563441009da05b5804bd79"
Commit e06b9871 authored by carlushuang's avatar carlushuang
Browse files

add more tile, fix a bug in 6x16 kernel

parent 5771a040
Show whitespace changes
Inline Side-by-side
Showing with 16 additions and 9 deletions
include/ck/tensor_operation/cpu/thread/threadwise_gemm_avx2.hpp
View file @ e06b9871
......@@ -1290,7 +1290,7 @@ struct ThreadwiseGemmAvx2_MxN_6x16_Dispatch
index_t im = mr - 1;
index_t in = (nr >> 3) - 1;
assert(im >= 0 && im <= 5 && in >= 0 && in <= 1);
return dispatch_table[mr][nr](param);
return dispatch_table[im][in](param);
}
};
......
library/src/tensor_operation_instance/cpu/conv2d_fwd/device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_instance.cpp
View file @ e06b9871
......@@ -48,22 +48,29 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
// clang-format off
#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(a_elem_op, b_elem_op, c_elem_op, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf) \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>
DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>
// clang-format on
using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_instances = std::tuple<
// clang-format off
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 256, 120, 64, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 512, 144, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 256, 128, 64, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 256, 128, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 128, 256, 128, 6, 16, true, true, false),
// DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 512, 144, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 512, 240, 128, 4, 24, true, true, false),
// DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 768, 192, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 768, 288, 128, 4, 24, true, true, false)>;
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 512, 256, 128, 6, 16, true, true, false),
// DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 768, 288, 128, 4, 24, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 768, 320, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 896, 352, 128, 6, 16, true, true, false),
DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true, true, false)>;
// clang-format on
void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
......
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