ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-01 04:07:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' of github.com:ROCm/composable_kernel into ck_moe_bs_splitk_pr

Browse Source
This commit is contained in:
yadaish
2025-12-16 04:40:20 +00:00
parent f3162c067e 1e6bbed1fb
commit d37c19e367
65 changed files with 3693 additions and 1067 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
example/26_contraction/run_contraction_bilinear_example.inc
View File

@@ -233,7 +233,20 @@ int run_contraction_bilinear_example(int argc, char* argv[])
}
}
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
if(ck::is_gfx11_supported())
{
return ck::utils::check_err(e_ms_ns_device_result,
e_ms_ns_host_result,
"Error: Incorrect results!",
1e-4,
1e-4)
? 0
: 1;
}
else
{
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
}
return 0;

15
example/26_contraction/run_contraction_scale_example.inc
View File

@@ -216,7 +216,20 @@ int run_contraction_scale_example(int argc, char* argv[])
}
}
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
if(ck::is_gfx11_supported())
{
return ck::utils::check_err(e_ms_ns_device_result,
e_ms_ns_host_result,
"Error: Incorrect results!",
1e-4,
1e-4)
? 0
: 1;
}
else
{
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
}
return 0;

58
example/ck_tile/03_gemm/gemm_utils.hpp
View File

@@ -12,40 +12,6 @@
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/utility/json_dump.hpp"
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
constexpr bool is_8bit_float =
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
if constexpr(M_Warp_Tile == 32)
return is_8bit_float ? 64 : 16;
else
return is_8bit_float ? 128 : 32;
#else
if constexpr(M_Warp_Tile == 32)
return 16;
else
return 32;
#endif
}
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile_flatmm()
{
#if defined(CK_GFX950_SUPPORT)
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 64;
else
return sizeof(PrecType) == 2 ? 32 : 128;
#else
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 32;
else
return sizeof(PrecType) == 2 ? 32 : 64;
#endif
}
struct GemmConfigBase
{
static constexpr bool kPadM = false;
@@ -122,7 +88,8 @@ struct GemmConfigComputeV3 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
@@ -141,7 +108,8 @@ struct GemmConfigComputeV3_1 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
@@ -160,7 +128,8 @@ struct GemmConfigComputeV3_2 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
@@ -204,7 +173,8 @@ struct GemmConfigComputeV4 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V4;
@@ -223,7 +193,8 @@ struct GemmConfigComputeV4_1 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V4;
@@ -242,7 +213,8 @@ struct GemmConfigComputeV5 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V5;
@@ -282,7 +254,8 @@ struct GemmConfigPreshuffleDecode : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr int kBlockPerCu = 1;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
@@ -306,7 +279,8 @@ struct GemmConfigPreshufflePrefill : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr int kBlockPerCu = 2;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;

43
example/ck_tile/03_gemm/universal_gemm_invoker.hpp
View File

@@ -63,25 +63,30 @@ struct UniversalInvoker
const auto Run = [&](const auto memory_operation_) {
constexpr auto memory_operation = memory_operation_.value;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType,
BDataType,
DsDataType,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation,
GemmConfig::NumWaveGroups>>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
ADataType,
BDataType,
DsDataType,
AccDataType,
CDataType,
DsLayout,
ELayout,
CDEElementWise,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
GemmConfig::N_Warp,
GemmConfig::M_Warp_Tile,
GemmConfig::N_Warp_Tile,
GemmConfig::K_Warp_Tile,
UniversalGemmProblem::TransposeC,
memory_operation,
GemmConfig::NumWaveGroups,
false, /*FixedVectorSize_*/
1, /*VectorSizeC_*/
false, /*TiledMMAPermuteN_*/
1, /*BlockedXDLN_PerWarp_*/
GemmConfig::DoubleSmemBuffer /*DoubleSmemBuffer*/>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);

49
example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
View File

@@ -11,40 +11,6 @@
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/utility/json_dump.hpp"
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
constexpr bool is_8bit_float =
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
if constexpr(M_Warp_Tile == 32)
return is_8bit_float ? 64 : 16;
else
return is_8bit_float ? 128 : 32;
#else
if constexpr(M_Warp_Tile == 32)
return 16;
else
return 32;
#endif
}
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile_flatmm()
{
#if defined(CK_GFX950_SUPPORT)
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 64;
else
return sizeof(PrecType) == 2 ? 32 : 128;
#else
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 32;
else
return sizeof(PrecType) == 2 ? 32 : 64;
#endif
}
template <typename DataType>
struct GemmTypeConfig;
@@ -111,7 +77,8 @@ struct GemmConfigComputeV3_2 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = false;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V3;
@@ -134,7 +101,8 @@ struct GemmConfigComputeV4 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V4;
@@ -157,7 +125,8 @@ struct GemmConfigComputeV4_V2 : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr bool DoubleSmemBuffer = true;
static constexpr ck_tile::GemmPipeline Pipeline = ck_tile::GemmPipeline::COMPUTE_V4;
@@ -178,7 +147,8 @@ struct GemmConfigPreshuffleDecode : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr bool kPadK = true;
@@ -203,7 +173,8 @@ struct GemmConfigPreshufflePrefill : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile_flatmm<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr int kBlockPerCu = 2;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;

18
example/ck_tile/17_grouped_gemm/grouped_gemm_multi_d.hpp
View File

@@ -11,24 +11,6 @@
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/utility/json_dump.hpp"
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
constexpr bool is_8bit_float =
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
if constexpr(M_Warp_Tile == 32)
return is_8bit_float ? 64 : 16;
else
return is_8bit_float ? 128 : 32;
#else
if constexpr(M_Warp_Tile == 32)
return 16;
else
return 32;
#endif
}
struct GemmConfigBase
{
static constexpr bool kPadM = false;

39
example/ck_tile/17_grouped_gemm/quant_grouped_gemm.hpp
View File

@@ -10,40 +10,6 @@
#include "ck_tile/ops/gemm.hpp"
#include "ck_tile/ops/elementwise/unary_element_wise_operation.hpp"
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
constexpr bool is_8bit_float =
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
if constexpr(M_Warp_Tile == 32)
return is_8bit_float ? 64 : 16;
else
return is_8bit_float ? 128 : 32;
#else
if constexpr(M_Warp_Tile == 32)
return 16;
else
return 32;
#endif
}
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_from_preshuffled_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 64;
else
return sizeof(PrecType) == 2 ? 32 : 128;
#else
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 32;
else
return sizeof(PrecType) == 2 ? 32 : 64;
#endif
}
template <typename DataType>
struct GemmTypeConfig;
@@ -100,7 +66,8 @@ struct GemmConfigComputeV3_2 : public GemmConfigBase<Persistent>
static constexpr ck_tile::index_t M_Warp_Tile = 32;
static constexpr ck_tile::index_t N_Warp_Tile = 32;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
};
template <typename PrecType, bool Persistent>
@@ -117,7 +84,7 @@ struct GemmConfigPreshuffleB_Bquant_prefill : public GemmConfigBase<Persistent>
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr bool PreshuffleB = true;
static constexpr bool DoubleSmemBuffer = true;

48
example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
View File

@@ -24,39 +24,6 @@ inline size_t hash_multiple_strings(const std::vector<std::string>& inputs)
return combined_hash;
}
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
constexpr bool is_8bit_float =
std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
if constexpr(M_Warp_Tile == 32)
return is_8bit_float ? 64 : 16;
else
return is_8bit_float ? 128 : 32;
#else
if constexpr(M_Warp_Tile == 32)
return 16;
else
return 32;
#endif
}
template <typename PrecType, ck_tile::index_t M_Warp_Tile>
constexpr ck_tile::index_t get_k_from_preshuffled_warp_tile()
{
#if defined(CK_GFX950_SUPPORT)
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 64;
else
return sizeof(PrecType) == 2 ? 32 : 128;
#else
if constexpr(M_Warp_Tile == 32)
return sizeof(PrecType) == 2 ? 16 : 32;
else
return sizeof(PrecType) == 2 ? 32 : 64;
#endif
}
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
@@ -124,7 +91,8 @@ struct GemmConfigQuantDecode : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
};
template <typename PrecType>
@@ -140,7 +108,8 @@ struct GemmConfigRowColQuant : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
};
template <typename PrecType>
@@ -157,7 +126,7 @@ struct GemmConfigPreshuffleQuantDecode : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr bool PreshuffleQuant = true;
};
@@ -176,7 +145,7 @@ struct GemmConfigPreshuffleB_BQuant_Decode : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr bool PreshuffleB = true;
static constexpr bool DoubleSmemBuffer = true;
@@ -206,7 +175,7 @@ struct GemmConfigPreshuffleB_BQuant_Prefill : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile =
get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile, true>();
static constexpr bool PreshuffleB = true;
static constexpr bool DoubleSmemBuffer = true;
@@ -236,7 +205,8 @@ struct GemmConfigQuantPrefill : public GemmConfigBase
static constexpr ck_tile::index_t M_Warp_Tile = 16;
static constexpr ck_tile::index_t N_Warp_Tile = 16;
static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
static constexpr ck_tile::index_t K_Warp_Tile =
ck_tile::get_k_warp_tile<PrecType, M_Warp_Tile>();
};
template <typename PrecType>