Skip to content

GitLab

Unverified Commit cff6de7e authored by Bartlomiej Wroblewski's avatar Bartlomiej Wroblewski Committed by GitHub
Browse files

Merge branch 'develop' into bwroblew/test_remove_sleep

parents 61123e03 6ef034f6
Hide whitespace changes
Inline Side-by-side
Showing with 487 additions and 141 deletions
include/ck/utility/type_convert.hpp
View file @ cff6de7e
......@@ -95,9 +95,113 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
return type_convert<bhalf_t>(x_fp32);
}
// convert fp32 to fp8
// Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float max_fp8 = 240.0f;
x = x > max_fp8 ? max_fp8 : (x < -max_fp8 ? -max_fp8 : x);
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
rng);
#endif
}
// convert fp16 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp16 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_rne(X x);
// convert fp32 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_t f8_convert_rne<f8_t, float>(float x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
float max_fp8 = 240.0f;
......@@ -124,6 +228,80 @@ inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
#endif
}
// convert fp16 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_t f8_convert_rne<f8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_rne<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, float>(float x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_bf8_f32(val.fval, val.fval, ival, false); // false -> WORD0
val.i32val = ival;
return val.i8val[0];
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp16 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_t f8_convert_rne<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_rne<bf8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to fp8
template <>
inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
{
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<f8_t>(x);
#else
return f8_convert_rne<f8_t>(x);
#endif
}
// convert fp8 to fp32
template <>
inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
......@@ -174,17 +352,10 @@ inline __host__ __device__ half2_t type_convert<half2_t, float2_t>(float2_t x)
template <>
inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return type_convert<f8_t>(type_convert<float>(x));
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<f8_t>(x);
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
return f8_convert_nre<f8_t>(x);
#endif
}
......@@ -205,26 +376,10 @@ inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, float>(float x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_pk_bf8_f32(val.fval, val.fval, ival, false); // false -> WORD0
val.i32val = ival;
return val.i8val[0];
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<bf8_t>(x);
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
return f8_convert_rne<bf8_t>(x);
#endif
}
......@@ -248,17 +403,10 @@ inline __host__ __device__ float type_convert<float, bf8_t>(bf8_t x)
template <>
inline __host__ __device__ bf8_t type_convert<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return type_convert<bf8_t>(type_convert<float>(x));
#if defined CK_USE_SR_F8_CONVERSION
return f8_convert_sr<bf8_t>(x);
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::standard;
constexpr uint32_t rng = 0;
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
return f8_convert_rne<bf8_t>(x);
#endif
}
......@@ -331,104 +479,4 @@ inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(h
return bf16_convert_rtn<bhalf_t>(x_fp32);
}
// Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y f8_convert_sr(X x);
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_fp8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(x,
rng);
#endif
}
// convert fp16 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
{
constexpr int seed = 42;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
union
{
float fval;
uint32_t i32val;
uint8_t i8val[4]; // not endian independent
} val;
val.fval = x;
uint32_t ival = 0;
ival = __builtin_amdgcn_cvt_sr_bf8_f32(val.fval, rng, ival, 0); // 0 pos
val.i32val = ival;
return val.i8val[0]; // little endian
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
return utils::
cast_to_f8<float, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
// convert fp16 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
{
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
// convert to float and use native converion
return f8_convert_sr<f8_t>(type_convert<float>(x));
#else
constexpr bool negative_zero_nan = true;
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 42;
// as thread id is not available on host, use 0 for prn generation
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
#endif
}
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ cff6de7e
......@@ -227,6 +227,10 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
......@@ -289,6 +293,26 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef CK_ENABLE_FP64
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
......@@ -382,6 +406,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
......@@ -391,6 +417,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -400,6 +428,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
......@@ -409,6 +439,8 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
op_ptrs);
}
}
#ifdef CK_ENABLE_FP16
......@@ -439,6 +471,8 @@ struct DeviceOperationInstanceFactory<
#endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......
library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
View file @ cff6de7e
......@@ -13,6 +13,10 @@ list(APPEND GEMM_INSTANCES
device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp
device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp
device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp
device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp
device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
......@@ -41,6 +45,7 @@ list(APPEND GEMM_INSTANCES
device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ cff6de7e
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
using device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
// ##################################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp 0 → 100644
View file @ cff6de7e
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances = std::tuple<
// clang-format off
// ##################################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp 0 → 100644
View file @ cff6de7e
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances = std::tuple<
// clang-format off
// ##################################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp 0 → 100644
View file @ cff6de7e
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances = std::tuple<
// clang-format off
// ##################################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp 0 → 100644
View file @ cff6de7e
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances = std::tuple<
// clang-format off
// ##################################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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