ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 17:26:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

enable gfx940

Browse Source
This commit is contained in:
illsilin
2023-01-30 21:23:23 -08:00
parent 7494c1c611
commit dc58fa9a89
44 changed files with 66 additions and 117 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
.gitignore vendored
View File

@@ -1,49 +0,0 @@
# Compiled Object files
*.slo
*.lo
*.o
*.obj
# Precompiled Headers
*.gch
*.pch
*.ipch
# Compiled Dynamic libraries
*.so
*.dylib
*.dll
# Fortran module files
*.mod
# Compiled Static libraries
*.lai
*.la
*.a
*.lib
# Executables
*.exe
*.out
*.app
# vim tags
tags
.tags
.*.swp
# Editors
.vscode
# build-in-source directory
build*
# emacs temporary/backup files
.\#*
\#*\#
*~
# GDB temporary files
.gdb_history
install.dir*

14
include/ck/ck.hpp
View File

@@ -30,7 +30,7 @@
// check GPU target
#ifdef __HIP_DEVICE_COMPILE__
#if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__))
defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx940__))
#error Not supported target
#endif
#endif
@@ -39,7 +39,7 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) // for GPU code
defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx1030__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
@@ -52,7 +52,7 @@
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
#define CK_USE_AMD_V_MAC_F32
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx1030__) // for GPU code
defined(__gfx1030__) || defined(__gfx940__) // for GPU code
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
@@ -61,11 +61,11 @@
// MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA
#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_USE_AMD_MFMA
#endif
#if defined(__gfx90a__)
#if (defined(__gfx90a__) || defined(__gfx940__))
#define CK_USE_AMD_MFMA_BF16_1K_OP
#endif
@@ -88,13 +88,13 @@
// buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
#endif
#if defined(__gfx90a__) // for GPU code
#if (defined(__gfx90a__) || defined(__gfx940__)) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
#else
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0

4
include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp
View File

@@ -47,7 +47,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -416,7 +416,7 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View File

@@ -43,7 +43,7 @@ __global__ void
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
@@ -678,7 +678,7 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
View File

@@ -56,7 +56,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
@@ -938,7 +938,7 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View File

@@ -56,7 +56,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
@@ -839,7 +839,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
View File

@@ -74,7 +74,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

4
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
View File

@@ -60,7 +60,7 @@ __global__ void
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -588,7 +588,7 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
View File

@@ -83,7 +83,7 @@ __global__ void
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -579,7 +579,7 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -68,7 +68,7 @@ __global__ void
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -790,7 +790,7 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
View File

@@ -59,7 +59,7 @@ __global__ void
const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

4
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View File

@@ -62,7 +62,7 @@ __global__ void
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -636,7 +636,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
arg.Print();
#endif
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
View File

@@ -62,7 +62,7 @@ __global__ void
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -612,7 +612,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
View File

@@ -75,7 +75,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

4
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View File

@@ -52,7 +52,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -581,7 +581,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
View File

@@ -55,7 +55,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

4
include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_e_permute_xdl.hpp
View File

@@ -51,7 +51,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -456,7 +456,7 @@ struct DeviceGemmBiasEPermute_Xdl : public DeviceGemmBiasCPermute<AElementwiseOp
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View File

@@ -63,7 +63,7 @@ __global__ void
const Block2ETileMap block_2_etile_map,
index_t NRaw)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
GridwiseGemmWelford::template Run<HasMainKBlockLoop>(
@@ -851,7 +851,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
View File

@@ -60,7 +60,7 @@ __global__ void
const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -554,7 +554,7 @@ struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -51,7 +51,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -490,7 +490,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
View File

@@ -432,7 +432,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
return false;
}
}
else if(ck::get_device_name() == "gfx90a")
else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

2
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
View File

@@ -574,7 +574,7 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
View File

@@ -648,7 +648,7 @@ struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

4
include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
View File

@@ -37,7 +37,7 @@ __global__ void
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
@@ -703,7 +703,7 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940"))
{
return false;
}

2
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View File

@@ -130,7 +130,7 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);

2
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp
View File

@@ -78,7 +78,7 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

4
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View File

@@ -155,7 +155,7 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -810,7 +810,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
return false;
}
}
else if(get_device_name() == "gfx90a")
else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

4
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
View File

@@ -135,7 +135,7 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -684,7 +684,7 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
return false;
}
}
else if(get_device_name() == "gfx90a")
else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

2
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
View File

@@ -38,7 +38,7 @@ __global__ void
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
View File

@@ -66,7 +66,7 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View File

@@ -54,7 +54,7 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
View File

@@ -44,7 +44,7 @@ __global__ void
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
View File

@@ -57,7 +57,7 @@ __global__ void
const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
// TODO ANT: separate into MMA + Epilogue

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
View File

@@ -165,7 +165,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
View File

@@ -44,7 +44,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View File

@@ -43,7 +43,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
View File

@@ -42,7 +42,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View File

@@ -44,7 +44,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
View File

@@ -46,7 +46,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
View File

@@ -49,7 +49,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(

2
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
View File

@@ -53,7 +53,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(

2
include/ck/utility/amd_xdlops.hpp
View File

@@ -306,7 +306,7 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
template <class FloatC>
__device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
{
#ifdef __gfx90a__
#if defined(__gfx90a__) || defined(__gfx940__)
reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
#else

3
script/cmake-ck-dev.sh
View File

@@ -11,9 +11,8 @@ cmake
-D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \
-D GPU_TARGETS="gfx908;gfx90a" \
-D GPU_TARGETS="gfx90a" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
#-D AMDGPU_TARGETS=gfx90a;gfx908

3
script/cmake-ck-release.sh
View File

@@ -11,9 +11,8 @@ cmake
-D CMAKE_CXX_FLAGS="-O3" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=OFF \
-D GPU_TARGETS="gfx908;gfx90a" \
-D GPU_TARGETS="gfx908;gfx90a;gfx940" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
#-D AMDGPU_TARGETS=gfx90a;gfx908