Skip to content

GitLab

Commit 5683ea4e authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents f0831350 dddc2115
Hide whitespace changes
Inline Side-by-side
Showing with 340 additions and 339 deletions
example/40_conv2d_fwd_quantization/CMakeLists.txt
View file @ 5683ea4e
......@@ -10,21 +10,19 @@ foreach(gpu IN LISTS GPU_TARGETS)
set(target 1)
endif()
endforeach()
# Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
# Conv perchannel quantization
add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
# Conv + bias + relu perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
# Conv + bias + relu perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
# Conv + bias + tanh perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
if(DL_KERNELS)
# Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
# Conv perchannel quantization
add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
# Conv + bias + relu perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
# Conv + bias + relu perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
# Conv + bias + tanh perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
endif()
endif()
\ No newline at end of file
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @ 5683ea4e
......@@ -3,9 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4)
......@@ -14,5 +20,7 @@ foreach(gpu IN LISTS GPU_TARGETS)
endforeach()
if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
endif()
endif()
example/42_groupnorm/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
endif()
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
endif()
example/44_elementwise_permute/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
endif()
example/46_gemm_add_multiply/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
if(DL_KERNELS)
add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
endif()
add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
endif()
example/48_pool3d_fwd/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
endif()
example/49_maxpool2d_bwd/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
endif()
example/50_put_element/CMakeLists.txt
View file @ 5683ea4e
add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
endif()
include/ck/ck.hpp
View file @ 5683ea4e
......@@ -198,7 +198,7 @@
#define CK_WORKAROUND_SWDEV_388832 1
// workaround: Grouped Conv2d_bwd_data fails for already implemented instance
#define CK_WORKAROUND_SWDEV_3318619 0
#define CK_WORKAROUND_GITHUB_ISSUE_824 1
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
View file @ 5683ea4e
......@@ -27,15 +27,12 @@ struct DeviceGroupedConvBwdWeight : public BaseOperator
MakeArgumentPointer(const void* p_in,
void* p_wei,
const void* p_out,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp
View file @ 5683ea4e
......@@ -784,15 +784,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
Argument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& /*input_strides*/,
const std::array<ck::index_t, NDimSpatial + 3>& /*output_strides*/,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& /*a_g_n_c_wis_strides*/,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& /*b_g_k_c_xs_strides*/,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& /*e_g_n_k_wos_strides*/,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
......@@ -812,27 +809,38 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
a_element_op_{out_element_op},
b_element_op_{wei_element_op},
c_element_op_{in_element_op},
Conv_G_{G},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
input_spatial_lengths_{input_spatial_lengths},
filter_spatial_lengths_{filter_spatial_lengths},
output_spatial_lengths_{output_spatial_lengths},
Conv_G_{a_g_n_c_wis_lengths[0]},
Conv_N_{a_g_n_c_wis_lengths[1]},
Conv_K_{b_g_k_c_xs_lengths[1]},
Conv_C_{a_g_n_c_wis_lengths[2]},
input_spatial_lengths_{},
filter_spatial_lengths_{},
output_spatial_lengths_{},
conv_filter_strides_{conv_filter_strides},
conv_filter_dilations_{conv_filter_dilations},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads},
k_batch_{split_k}
{
constexpr index_t spatial_offset = 3;
std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
end(a_g_n_c_wis_lengths),
begin(input_spatial_lengths_));
std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
end(b_g_k_c_xs_lengths),
begin(filter_spatial_lengths_));
std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
end(e_g_n_k_wos_lengths),
begin(output_spatial_lengths_));
const auto descs =
DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
Conv_N_,
Conv_K_,
Conv_C_,
input_spatial_lengths_,
filter_spatial_lengths_,
output_spatial_lengths_,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -856,21 +864,21 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
// A/B/C Batch Stride
compute_ptr_offset_of_batch_.BatchStrideA_ =
N * K *
std::accumulate(begin(output_spatial_lengths),
end(output_spatial_lengths),
Conv_N_ * Conv_K_ *
std::accumulate(begin(output_spatial_lengths_),
end(output_spatial_lengths_),
index_t{1},
std::multiplies<>{});
compute_ptr_offset_of_batch_.BatchStrideB_ =
N * C *
std::accumulate(begin(input_spatial_lengths),
end(input_spatial_lengths),
Conv_N_ * Conv_C_ *
std::accumulate(begin(input_spatial_lengths_),
end(input_spatial_lengths_),
index_t{1},
std::multiplies<>{});
compute_ptr_offset_of_batch_.BatchStrideC_ =
K * C *
std::accumulate(begin(filter_spatial_lengths),
end(filter_spatial_lengths),
Conv_K_ * Conv_C_ *
std::accumulate(begin(filter_spatial_lengths_),
end(filter_spatial_lengths_),
index_t{1},
std::multiplies<>{});
}
......@@ -904,9 +912,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
const index_t Conv_K_;
const index_t Conv_C_;
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths_;
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
......@@ -1110,39 +1118,34 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k)
static auto
MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
a_g_n_c_wis_lengths, // input
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths, // weight
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths, // output
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -1159,15 +1162,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
MakeArgumentPointer(const void* p_in_grid,
void* p_wei_grid,
const void* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
......@@ -1180,15 +1180,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<WeiDataType*>(p_wei_grid),
static_cast<const OutDataType*>(p_out_grid),
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
a_g_n_c_wis_lengths, // input
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths, // weight
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths, // output
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
View file @ 5683ea4e
......@@ -245,21 +245,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const ck::index_t K,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
{
if constexpr(is_GNHWK_GKYXC_GNHWC)
{
return make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
}
else if constexpr(is_NHWGK_GKYXC_NHWGC)
{
const index_t WoStride = output_strides[4];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
make_tuple(WoStride, KStride));
}
else
{
throw std::runtime_error("wrong! unsupported layout: " + OutLayout::name());
}
const index_t WoStride = output_strides[4];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
make_tuple(WoStride, KStride));
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
......@@ -270,42 +259,36 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
{
if constexpr(is_GNHWK_GKYXC_GNHWC)
{
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
}
else
{
return make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
}
}
else if constexpr(is_NHWGK_GKYXC_NHWGC)
const index_t NStride = input_strides[1];
const index_t HiStride = input_strides[3];
const index_t WiStride = input_strides[4];
const auto CStride = input_strides[2];
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
const index_t NStride = input_strides[1];
const index_t HiStride = input_strides[3];
const index_t WiStride = input_strides[4];
const auto CStride = input_strides[2];
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, C),
make_tuple(WiStride, CStride));
}
else
{
return make_naive_tensor_descriptor(
make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
}
return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, C),
make_tuple(WiStride, CStride));
}
else
{
throw std::runtime_error("wrong! unsupported layout: " + InLayout::name());
return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
make_tuple(NStride, HiStride, WiStride, CStride));
}
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
constexpr static auto
make_wei_grid_desc(const ck::index_t K,
const ck::index_t Y,
const ck::index_t X,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
{
const auto CStride = Number<1>{};
const auto KStride = weights_strides[1];
return make_naive_tensor_descriptor(make_tuple(K, Y * X * C), make_tuple(KStride, CStride));
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
constexpr static auto
make_out_grid_desc(const ck::index_t N,
......@@ -315,21 +298,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const ck::index_t K,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
{
if constexpr(is_GNDHWK_GKZYXC_GNDHWC)
{
return make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
}
else if constexpr(is_NDHWGK_GKZYXC_NDHWGC)
{
const index_t WoStride = output_strides[5];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
make_tuple(WoStride, KStride));
}
else
{
throw std::runtime_error("wrong! unsupported layout: " + OutLayout::name());
}
const index_t WoStride = output_strides[5];
const auto KStride = Number<1>{};
return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
make_tuple(WoStride, KStride));
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
......@@ -341,44 +313,40 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
{
if constexpr(is_GNDHWK_GKZYXC_GNDHWC)
{
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
}
else
{
return make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
}
}
else if constexpr(is_NDHWGK_GKZYXC_NDHWGC)
const index_t NStride = input_strides[1];
const index_t DiStride = input_strides[3];
const index_t HiStride = input_strides[4];
const index_t WiStride = input_strides[5];
const auto CStride = input_strides[2];
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
const index_t NStride = input_strides[1];
const index_t DiStride = input_strides[3];
const index_t HiStride = input_strides[4];
const index_t WiStride = input_strides[5];
const auto CStride = input_strides[2];
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, C),
make_tuple(WiStride, CStride));
}
else
{
return make_naive_tensor_descriptor(
make_tuple(N, Di, Hi, Wi, C),
make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
}
return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, C),
make_tuple(WiStride, CStride));
}
else
{
throw std::runtime_error("wrong! unsupported layout: " + InLayout::name());
return make_naive_tensor_descriptor(
make_tuple(N, Di, Hi, Wi, C),
make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
}
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
constexpr static auto
make_wei_grid_desc(const ck::index_t K,
const ck::index_t Z,
const ck::index_t Y,
const ck::index_t X,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
{
const auto CStride = Number<1>{};
const auto KStride = weights_strides[1];
return make_naive_tensor_descriptor(make_tuple(K, Z * Y * X * C),
make_tuple(KStride, CStride));
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
const ck::index_t N,
......@@ -388,6 +356,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& /* input_strides */,
const std::array<ck::index_t, NDimSpatial + 3>& /* weights_strides */,
const std::array<ck::index_t, NDimSpatial + 3>& /* output_strides */,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
......@@ -542,6 +511,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
......@@ -584,6 +554,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const auto out_grid_desc = make_out_grid_desc<NDim>(N, Ho, Wo, K, output_strides);
const auto in_grid_desc = make_in_grid_desc<NDim>(N, Hi, Wi, C, input_strides);
const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Y, X, C, weights_strides);
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
......@@ -618,13 +589,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
wei_grid_desc);
}
else
{
......@@ -684,13 +651,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
wei_grid_desc);
}
}
......@@ -703,6 +666,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
......@@ -752,6 +716,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const auto out_grid_desc = make_out_grid_desc<NDim>(N, Do, Ho, Wo, K, output_strides);
const auto in_grid_desc = make_in_grid_desc<NDim>(N, Di, Hi, Wi, C, input_strides);
const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Z, Y, X, C, weights_strides);
if constexpr(ConvBackwardWeightSpecialization ==
ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
......@@ -786,13 +751,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
wei_grid_desc);
}
else
{
......@@ -861,13 +822,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
wei_grid_desc);
}
} // function end
......@@ -887,6 +844,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
lengths,
strides,
strides,
strides,
params,
params,
params,
......@@ -910,6 +868,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
lengths,
strides,
strides,
strides,
params,
params,
params,
......@@ -933,6 +892,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
lengths,
strides,
strides,
strides,
params,
params,
params,
......@@ -1051,15 +1011,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
Argument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
......@@ -1084,27 +1041,40 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
a_element_op_{out_element_op},
b_element_op_{in_element_op},
c_element_op_{wei_element_op},
Conv_G_{G},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
output_spatial_lengths_{output_spatial_lengths},
filter_spatial_lengths_{filter_spatial_lengths},
Conv_G_{a_g_n_c_wis_lengths[0]},
Conv_N_{a_g_n_c_wis_lengths[1]},
Conv_K_{b_g_k_c_xs_lengths[1]},
Conv_C_{a_g_n_c_wis_lengths[2]},
input_spatial_lengths_{},
filter_spatial_lengths_{},
output_spatial_lengths_{},
conv_filter_strides_{conv_filter_strides},
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads},
k_batch_{split_k}
{
constexpr index_t spatial_offset = 3;
std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
end(a_g_n_c_wis_lengths),
begin(input_spatial_lengths_));
std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
end(b_g_k_c_xs_lengths),
begin(filter_spatial_lengths_));
std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
end(e_g_n_k_wos_lengths),
begin(output_spatial_lengths_));
const auto descs =
DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
Conv_N_,
Conv_K_,
Conv_C_,
input_spatial_lengths_,
filter_spatial_lengths_,
output_spatial_lengths_,
a_g_n_c_wis_strides,
b_g_k_c_xs_strides,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -1119,12 +1089,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
// A/B/C Batch Stride
compute_ptr_offset_of_batch_.BatchStrideA_ = output_strides[0];
compute_ptr_offset_of_batch_.BatchStrideB_ = input_strides[0];
compute_ptr_offset_of_batch_.BatchStrideA_ = e_g_n_k_wos_strides[0];
compute_ptr_offset_of_batch_.BatchStrideB_ = a_g_n_c_wis_strides[0];
compute_ptr_offset_of_batch_.BatchStrideC_ =
K * C *
std::accumulate(begin(filter_spatial_lengths),
end(filter_spatial_lengths),
Conv_K_ * Conv_C_ *
std::accumulate(begin(filter_spatial_lengths_),
end(filter_spatial_lengths_),
index_t{1},
std::multiplies<>{});
......@@ -1163,8 +1133,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
const index_t Conv_N_;
const index_t Conv_K_;
const index_t Conv_C_;
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
......@@ -1339,39 +1310,34 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const ck::index_t split_k)
static auto
MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid,
const OutDataType* p_out_grid,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const ck::index_t split_k)
{
return Argument{p_in_grid,
p_wei_grid,
p_out_grid,
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
a_g_n_c_wis_lengths, // input
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths, // weight
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths, // output
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -1390,15 +1356,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
MakeArgumentPointer(const void* p_in_grid,
void* p_wei_grid,
const void* p_out_grid,
const ck::index_t G,
const ck::index_t N,
const ck::index_t K,
const ck::index_t C,
const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
......@@ -1411,15 +1374,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<WeiDataType*>(p_wei_grid),
static_cast<const OutDataType*>(p_out_grid),
G,
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
input_strides,
output_strides,
a_g_n_c_wis_lengths, // input
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths, // weight
b_g_k_c_xs_strides,
e_g_n_k_wos_lengths, // output
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
View file @ 5683ea4e
......@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef __bf16__
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
......@@ -36,7 +36,8 @@ void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef __fp16__
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -56,7 +57,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef __fp32__
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
......@@ -76,7 +78,8 @@ void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef __int8__
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
Row,
......@@ -120,7 +123,7 @@ void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
template <typename ALayout,
typename BLayout,
typename CLayout,
......@@ -151,7 +154,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef __fp32__
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<CDataType, float>)
{
......@@ -176,8 +179,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
#endif
#ifdef __fp16__
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -200,8 +205,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
is_same_v<CDataType, bhalf_t>)
#endif
#ifdef __bf16__
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -224,8 +231,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, int8_t>)
#endif
#ifdef __int8__
if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, int8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -248,7 +257,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
}
}
#endif
return op_ptrs;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
View file @ 5683ea4e
......@@ -14,7 +14,7 @@
using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
#ifdef __fp16__
namespace ck {
namespace tensor_operation {
namespace device {
......@@ -137,3 +137,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
View file @ 5683ea4e
......@@ -13,7 +13,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef __fp16__
namespace ck {
namespace tensor_operation {
namespace device {
......@@ -91,3 +91,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
View file @ 5683ea4e
......@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef __fp16__
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
......@@ -58,7 +58,8 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances);
#endif
#ifdef __bf16__
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
......@@ -100,7 +101,7 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances);
#endif
template <typename ADataType,
typename B0DataType,
typename B1DataType,
......@@ -147,7 +148,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef __fp16__
if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
Acc0BiasDataType::Size() == 1 &&
......@@ -164,6 +165,8 @@ struct DeviceOperationInstanceFactory<
op_ptrs);
}
}
#endif
#ifdef __bf16__
else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
Acc0BiasDataType::Size() == 1 &&
......@@ -180,6 +183,7 @@ struct DeviceOperationInstanceFactory<
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
View file @ 5683ea4e
......@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef __fp16__
void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
Col,
......@@ -111,3 +111,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
View file @ 5683ea4e
......@@ -19,7 +19,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef __fp16__
void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
Row,
......@@ -123,7 +123,8 @@ void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instan
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef __int8__
void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
Row,
......@@ -227,7 +228,7 @@ void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
template <typename ALayout,
typename BLayout,
typename ELayout,
......@@ -262,7 +263,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef __fp16__
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
......@@ -295,6 +296,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
op_ptrs);
}
}
#endif
#ifdef __int8__
else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<EDataType, int8_t>)
{
......@@ -327,7 +330,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
View file @ 5683ea4e
......@@ -11,7 +11,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef __fp16__
namespace ck {
namespace tensor_operation {
namespace device {
......@@ -119,3 +119,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
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