Skip to content

GitLab

"tests/git@developer.sourcefind.cn:renzhc/diffusers_dcu.git" did not exist on "ccdd96ca5231a59617661d60d5dd822cd97281fc"
Unverified Commit aa0b9798 authored by Daming Feng's avatar Daming Feng Committed by GitHub
Browse files

Add compute type check for convolution instances (#1015) 



* add compute type check for fp16 in forward convolution instances

* Add compute type check for default compute types

---------
Co-authored-by: default avatarBartlomiej Kocot <barkocot@amd.com>
parent b0568b72
Hide whitespace changes
Inline Side-by-side
Showing with 108 additions and 57 deletions
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
View file @ aa0b9798
...@@ -496,7 +496,8 @@ struct DeviceOperationInstanceFactory< ...@@ -496,7 +496,8 @@ struct DeviceOperationInstanceFactory<
{ {
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, F16> &&
is_same_v<ComputeTypeB, F16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_instances( add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_instances(
...@@ -507,14 +508,16 @@ struct DeviceOperationInstanceFactory< ...@@ -507,14 +508,16 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16> && is_same_v<ComputeTypeA, BF16> &&
is_same_v<ComputeTypeB, BF16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances( add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
op_ptrs); op_ptrs);
...@@ -522,7 +525,9 @@ struct DeviceOperationInstanceFactory< ...@@ -522,7 +525,9 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_1x1s1p0_instances( add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_1x1s1p0_instances(
...@@ -535,7 +540,8 @@ struct DeviceOperationInstanceFactory< ...@@ -535,7 +540,8 @@ struct DeviceOperationInstanceFactory<
{ {
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, F16> &&
is_same_v<ComputeTypeB, F16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_instances( add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_instances(
...@@ -546,14 +552,16 @@ struct DeviceOperationInstanceFactory< ...@@ -546,14 +552,16 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16> && is_same_v<ComputeTypeA, BF16> &&
is_same_v<ComputeTypeB, BF16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances( add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
op_ptrs); op_ptrs);
...@@ -561,7 +569,9 @@ struct DeviceOperationInstanceFactory< ...@@ -561,7 +569,9 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_1x1s1p0_instances( add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_1x1s1p0_instances(
...@@ -578,7 +588,8 @@ struct DeviceOperationInstanceFactory< ...@@ -578,7 +588,8 @@ struct DeviceOperationInstanceFactory<
{ {
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, F16> &&
is_same_v<ComputeTypeB, F16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
op_ptrs); op_ptrs);
...@@ -590,7 +601,8 @@ struct DeviceOperationInstanceFactory< ...@@ -590,7 +601,8 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
op_ptrs); op_ptrs);
...@@ -598,7 +610,8 @@ struct DeviceOperationInstanceFactory< ...@@ -598,7 +610,8 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16> && is_same_v<ComputeTypeA, BF16> &&
is_same_v<ComputeTypeB, BF16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
op_ptrs); op_ptrs);
...@@ -606,7 +619,9 @@ struct DeviceOperationInstanceFactory< ...@@ -606,7 +619,9 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_instances( add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_instances(
op_ptrs); op_ptrs);
...@@ -642,7 +657,8 @@ struct DeviceOperationInstanceFactory< ...@@ -642,7 +657,8 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances( add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs); op_ptrs);
...@@ -650,7 +666,8 @@ struct DeviceOperationInstanceFactory< ...@@ -650,7 +666,8 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16> && is_same_v<ComputeTypeA, BF16> &&
is_same_v<ComputeTypeB, BF16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances( add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
op_ptrs); op_ptrs);
...@@ -658,7 +675,9 @@ struct DeviceOperationInstanceFactory< ...@@ -658,7 +675,9 @@ struct DeviceOperationInstanceFactory<
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_instances( add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_instances(
op_ptrs); op_ptrs);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ aa0b9798
...@@ -618,7 +618,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -618,7 +618,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
...@@ -628,7 +629,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -628,7 +629,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
...@@ -638,7 +640,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -638,7 +640,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
...@@ -655,21 +659,25 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -655,21 +659,25 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef DL_KERNELS #ifdef DL_KERNELS
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances( add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
...@@ -685,7 +693,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -685,7 +693,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances( add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances(
...@@ -697,7 +706,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -697,7 +706,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances( add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances(
...@@ -709,7 +719,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -709,7 +719,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
...@@ -725,7 +737,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -725,7 +737,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances( add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances(
...@@ -737,7 +750,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -737,7 +750,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances( add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances(
...@@ -749,7 +763,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -749,7 +763,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
...@@ -768,7 +784,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -768,7 +784,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances( add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances(
...@@ -780,7 +797,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -780,7 +797,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances( add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances(
...@@ -796,7 +814,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -796,7 +814,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
...@@ -808,7 +828,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -808,7 +828,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instances( add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(
op_ptrs); op_ptrs);
...@@ -822,7 +844,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -822,7 +844,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeTypeA, float> &&
is_same_v<ComputeTypeB, float>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances( add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
...@@ -851,7 +874,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -851,7 +874,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
...@@ -863,7 +888,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -863,7 +888,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> &&
is_same_v<ComputeTypeA, int8_t> &&
is_same_v<ComputeTypeB, int8_t>)
{ {
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances( add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(
op_ptrs); op_ptrs);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ aa0b9798
...@@ -928,28 +928,29 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -928,28 +928,29 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>) is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, bhalf_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> && is_same_v<ComputeType, int8_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs);
} }
...@@ -961,7 +962,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -961,7 +962,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
} }
...@@ -969,7 +970,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -969,7 +970,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS)) #if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
} }
...@@ -977,7 +978,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -977,7 +978,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
...@@ -989,7 +990,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -989,7 +990,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS)) #if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
} }
...@@ -997,7 +998,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -997,7 +998,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>) is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, bhalf_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
} }
...@@ -1005,7 +1007,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1005,7 +1007,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> && is_same_v<ComputeType, int8_t>)
{ {
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instances(op_ptrs);
...@@ -1021,7 +1023,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1021,7 +1023,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
} }
...@@ -1029,7 +1031,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1029,7 +1031,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS)) #if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
} }
...@@ -1037,7 +1039,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1037,7 +1039,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
} }
...@@ -1045,7 +1047,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1045,7 +1047,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS)) #if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
} }
...@@ -1053,14 +1055,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1053,14 +1055,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>) is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, bhalf_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> && is_same_v<ComputeType, int8_t>)
{ {
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instances(op_ptrs);
...@@ -1075,14 +1078,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1075,14 +1078,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_FP16 #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
...@@ -1095,14 +1098,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1095,14 +1098,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>) is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, bhalf_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> && is_same_v<ComputeType, int8_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(op_ptrs); add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(op_ptrs);
...@@ -1119,7 +1123,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1119,7 +1123,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
#ifdef CK_ENABLE_FP32 #ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float> && is_same_v<ComputeType, float>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
} }
...@@ -1148,14 +1152,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -1148,14 +1152,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
#ifdef CK_ENABLE_BF16 #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> && if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>) is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, bhalf_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
} }
#endif #endif
#ifdef CK_ENABLE_INT8 #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t> && is_same_v<ComputeType, int8_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(op_ptrs); add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(op_ptrs);
......
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