ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-18 12:00:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

[gfx110x] support Navi3x architectures. (#628)

Browse Source 
* enable building on Nav31

* fix syntax

* replace GPU_TARGETS with offload-arch

* add gfx1102 rachitecture

* fix typo

* update changelog

[ROCm/composable_kernel commit: 0ccecc7c31]
This commit is contained in:
Illia Silin
2023-03-09 05:56:40 -08:00
committed by GitHub
parent 0d23b0d1c9
commit 9ce65cae0e
13 changed files with 27 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
View File

@@ -770,7 +770,8 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{

3
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
View File

@@ -476,7 +476,8 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{

3
include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
View File

@@ -404,7 +404,8 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{

3
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
View File

@@ -579,7 +579,8 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
namespace ctc = tensor_layout::convolution;
// check device
if(get_device_name() == "gfx1100")
if(get_device_name() == "gfx1100" || get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{

11
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View File

@@ -54,7 +54,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -147,7 +148,8 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__))
// printf("entry kernel launch");
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
@@ -242,7 +244,8 @@ __global__ void
const CDEElementwiseOperation cde_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__))
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -271,7 +274,7 @@ __global__ void
ignore = b_element_op;
ignore = cde_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx1100__))
#endif // end of if (defined(__gfx1100__ ))
}
template < // DataType Family

3
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View File

@@ -49,7 +49,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,