Skip to content

GitLab

Unverified Commit 7e9a9d32 authored by rocking5566's avatar rocking5566 Committed by GitHub
Browse files

[Bf16 & int8] [example & ckprofiler] (#100) 



* Add int8 of mk_nk_mn to the ckProfiler

* Add example of int8 gemm

* Fix typo, use ushort instead of half_t for bfloat16

* replace ushortXXX_t to bhalfXXX_t

* rename ushort to bhalf_t

* Add bf16 example

* Add bf16 gemm to ckProfiler

* Fix alignment

* Fix typo

* Add unit test for gemm_xdl int8

* Add gemm_xdl fp32 unit test

* Add gemm_xdl bf16 unit test

* fix build

* fix build issue due to merge conflict

* Fix build

* Fix build error
Co-authored-by: default avatarrocking <chunylai@amd.com>
Co-authored-by: default avatarChao Liu <chao.liu2@amd.com>
parent 0c79af12
Expand all Show whitespace changes
Inline Side-by-side
Showing with 817 additions and 141 deletions
composable_kernel/include/tensor_operation/element_wise_operation.hpp
View file @ 7e9a9d32
...@@ -14,7 +14,7 @@ struct PassThrough ...@@ -14,7 +14,7 @@ struct PassThrough
__host__ __device__ void operator()(half_t& y, const half_t& x) const { y = x; } __host__ __device__ void operator()(half_t& y, const half_t& x) const { y = x; }
__host__ __device__ void operator()(ushort& y, const ushort& x) const { y = x; } __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const { y = x; }
__host__ __device__ void operator()(int32_t& y, const int32_t& x) const { y = x; } __host__ __device__ void operator()(int32_t& y, const int32_t& x) const { y = x; }
......
composable_kernel/include/tensor_operation/xdlops_gemm.hpp
View file @ 7e9a9d32
...@@ -474,7 +474,7 @@ struct MfmaSelector ...@@ -474,7 +474,7 @@ struct MfmaSelector
} }
template <> template <>
static constexpr auto GetMfma<ushort, 32, 32>() static constexpr auto GetMfma<bhalf_t, 32, 32>()
{ {
#if defined(CK_AMD_GPU_GFX90A) #if defined(CK_AMD_GPU_GFX90A)
return MfmaInstr::mfma_f32_32x32x8bf16_1k; return MfmaInstr::mfma_f32_32x32x8bf16_1k;
...@@ -484,7 +484,7 @@ struct MfmaSelector ...@@ -484,7 +484,7 @@ struct MfmaSelector
} }
template <> template <>
static constexpr auto GetMfma<ushort, 16, 16>() static constexpr auto GetMfma<bhalf_t, 16, 16>()
{ {
#if defined(CK_AMD_GPU_GFX90A) #if defined(CK_AMD_GPU_GFX90A)
return MfmaInstr::mfma_f32_16x16x16bf16_1k; return MfmaInstr::mfma_f32_16x16x16bf16_1k;
...@@ -662,8 +662,8 @@ struct XdlopsGemm ...@@ -662,8 +662,8 @@ struct XdlopsGemm
__device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const __device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
{ {
static_assert(is_same<base_type, float>::value || is_same<base_type, half_t>::value || static_assert(is_same<base_type, float>::value || is_same<base_type, half_t>::value ||
is_same<base_type, ushort>::value || is_same<base_type, int8_t>::value, is_same<base_type, bhalf_t>::value || is_same<base_type, int8_t>::value,
"base base_type must be float, half, ushort, and int8_t!"); "base base_type must be float, half, bfloat16, and int8_t!");
static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) { static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
mfma_instr.template run<MPerXdlops, NPerXdlops>(p_a_wave[k], p_b_wave[k], p_c_thread); mfma_instr.template run<MPerXdlops, NPerXdlops>(p_a_wave[k], p_b_wave[k], p_c_thread);
......
composable_kernel/include/utility/amd_buffer_addressing.hpp
View file @ 7e9a9d32
...@@ -51,19 +51,19 @@ llvm_amdgcn_raw_buffer_load_i8x4(int32x4_t srsrc, ...@@ -51,19 +51,19 @@ llvm_amdgcn_raw_buffer_load_i8x4(int32x4_t srsrc,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8"); index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8");
// buffer load i16 // buffer load i16
__device__ ushort __device__ bhalf_t
llvm_amdgcn_raw_buffer_load_i16(int32x4_t srsrc, llvm_amdgcn_raw_buffer_load_i16(int32x4_t srsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16"); index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16");
__device__ ushort2_t __device__ bhalf2_t
llvm_amdgcn_raw_buffer_load_i16x2(int32x4_t srsrc, llvm_amdgcn_raw_buffer_load_i16x2(int32x4_t srsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16"); index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16");
__device__ ushort4_t __device__ bhalf4_t
llvm_amdgcn_raw_buffer_load_i16x4(int32x4_t srsrc, llvm_amdgcn_raw_buffer_load_i16x4(int32x4_t srsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
...@@ -149,21 +149,21 @@ llvm_amdgcn_raw_buffer_store_i8x4(int8x4_t vdata, ...@@ -149,21 +149,21 @@ llvm_amdgcn_raw_buffer_store_i8x4(int8x4_t vdata,
// buffer store i16 // buffer store i16
__device__ void __device__ void
llvm_amdgcn_raw_buffer_store_i16(ushort vdata, llvm_amdgcn_raw_buffer_store_i16(bhalf_t vdata,
int32x4_t rsrc, int32x4_t rsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16"); index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16");
__device__ void __device__ void
llvm_amdgcn_raw_buffer_store_i16x2(ushort2_t vdata, llvm_amdgcn_raw_buffer_store_i16x2(bhalf2_t vdata,
int32x4_t rsrc, int32x4_t rsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16"); index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16");
__device__ void __device__ void
llvm_amdgcn_raw_buffer_store_i16x4(ushort4_t vdata, llvm_amdgcn_raw_buffer_store_i16x4(bhalf4_t vdata,
int32x4_t rsrc, int32x4_t rsrc,
index_t voffset, index_t voffset,
index_t soffset, index_t soffset,
...@@ -266,7 +266,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w ...@@ -266,7 +266,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
(is_same<T, double>::value && (N == 1 || N == 2 || N == 4)) || (is_same<T, double>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, ushort>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)), (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
"wrong! not implemented"); "wrong! not implemented");
...@@ -365,7 +365,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w ...@@ -365,7 +365,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
return bit_cast<half8_t>(tmp); return bit_cast<half8_t>(tmp);
} }
} }
else if constexpr(is_same<T, ushort>::value) else if constexpr(is_same<T, bhalf_t>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
...@@ -387,7 +387,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w ...@@ -387,7 +387,7 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4( int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
return bit_cast<ushort8_t>(tmp); return bit_cast<bhalf8_t>(tmp);
} }
} }
else if constexpr(is_same<T, int32_t>::value) else if constexpr(is_same<T, int32_t>::value)
...@@ -522,7 +522,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -522,7 +522,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
(is_same<T, double>::value && (N == 1 || N == 2)) || (is_same<T, double>::value && (N == 1 || N == 2)) ||
(is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) || (is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, ushort>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) || (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)), (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
"wrong! not implemented"); "wrong! not implemented");
...@@ -625,7 +625,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -625,7 +625,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
#endif #endif
} }
} }
else if constexpr(is_same<T, ushort>::value) else if constexpr(is_same<T, bhalf_t>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
...@@ -653,18 +653,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -653,18 +653,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 8) else if constexpr(N == 8)
{ {
vector_type<half_t, 8> tmp{src_thread_data}; vector_type<bhalf_t, 8> tmp{src_thread_data};
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<0>{}], llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<0>{}],
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<1>{}], llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<1>{}],
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset + 4 * sizeof(half_t), dst_wave_addr_offset + 4 * sizeof(bhalf_t),
0); 0);
} }
} }
......
composable_kernel/include/utility/amd_xdlops.hpp
View file @ 7e9a9d32
...@@ -207,7 +207,7 @@ template <> ...@@ -207,7 +207,7 @@ template <>
struct intrin_mfma_f32_32x32x8bf16_1k<32, 32> struct intrin_mfma_f32_32x32x8bf16_1k<32, 32>
{ {
template <class FloatC> template <class FloatC>
__device__ static void Run(const ushort4_t& reg_a, const ushort4_t& reg_b, FloatC& reg_c) __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
{ {
reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k( reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(
reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0); reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
...@@ -221,7 +221,7 @@ template <> ...@@ -221,7 +221,7 @@ template <>
struct intrin_mfma_f32_16x16x16bf16_1k<16, 16> struct intrin_mfma_f32_16x16x16bf16_1k<16, 16>
{ {
template <class FloatC> template <class FloatC>
__device__ static void Run(const ushort4_t& reg_a, const ushort4_t& reg_b, FloatC& reg_c) __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
{ {
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k( reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0); reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
...@@ -235,7 +235,7 @@ template <> ...@@ -235,7 +235,7 @@ template <>
struct intrin_mfma_f32_32x32x4bf16<32, 32> struct intrin_mfma_f32_32x32x4bf16<32, 32>
{ {
template <class FloatC> template <class FloatC>
__device__ static void Run(const ushort2_t& reg_a, const ushort2_t& reg_b, FloatC& reg_c) __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
{ {
reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16( reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16(
reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0); reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
...@@ -249,7 +249,7 @@ template <> ...@@ -249,7 +249,7 @@ template <>
struct intrin_mfma_f32_16x16x8bf16<16, 16> struct intrin_mfma_f32_16x16x8bf16<16, 16>
{ {
template <class FloatC> template <class FloatC>
__device__ static void Run(const ushort2_t& reg_a, const ushort2_t& reg_b, FloatC& reg_c) __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
{ {
reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16( reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16(
reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0); reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
......
composable_kernel/include/utility/data_type.hpp
View file @ 7e9a9d32
...@@ -5,6 +5,7 @@ ...@@ -5,6 +5,7 @@
namespace ck { namespace ck {
using bhalf_t = ushort;
using half_t = _Float16; using half_t = _Float16;
// vector_type // vector_type
...@@ -107,9 +108,9 @@ struct scalar_type<half_t> ...@@ -107,9 +108,9 @@ struct scalar_type<half_t>
}; };
template <> template <>
struct scalar_type<ushort> struct scalar_type<bhalf_t>
{ {
using type = ushort; using type = bhalf_t;
static constexpr index_t vector_size = 1; static constexpr index_t vector_size = 1;
}; };
...@@ -904,12 +905,12 @@ using half32_t = typename vector_type<half_t, 32>::type; ...@@ -904,12 +905,12 @@ using half32_t = typename vector_type<half_t, 32>::type;
using half64_t = typename vector_type<half_t, 64>::type; using half64_t = typename vector_type<half_t, 64>::type;
// bfp16 // bfp16
using ushort2_t = typename vector_type<ushort, 2>::type; using bhalf2_t = typename vector_type<bhalf_t, 2>::type;
using ushort4_t = typename vector_type<ushort, 4>::type; using bhalf4_t = typename vector_type<bhalf_t, 4>::type;
using ushort8_t = typename vector_type<ushort, 8>::type; using bhalf8_t = typename vector_type<bhalf_t, 8>::type;
using ushort16_t = typename vector_type<ushort, 16>::type; using bhalf16_t = typename vector_type<bhalf_t, 16>::type;
using ushort32_t = typename vector_type<ushort, 32>::type; using bhalf32_t = typename vector_type<bhalf_t, 32>::type;
using ushort64_t = typename vector_type<ushort, 64>::type; using bhalf64_t = typename vector_type<bhalf_t, 64>::type;
// i32 // i32
using int32x2_t = typename vector_type<int32_t, 2>::type; using int32x2_t = typename vector_type<int32_t, 2>::type;
...@@ -936,7 +937,7 @@ __host__ __device__ Y type_convert(X x) ...@@ -936,7 +937,7 @@ __host__ __device__ Y type_convert(X x)
// convert bfp16 to fp32 // convert bfp16 to fp32
template <> template <>
inline __host__ __device__ float type_convert(ushort x) inline __host__ __device__ float type_convert(bhalf_t x)
{ {
union union
{ {
...@@ -949,7 +950,7 @@ inline __host__ __device__ float type_convert(ushort x) ...@@ -949,7 +950,7 @@ inline __host__ __device__ float type_convert(ushort x)
// convert fp32 to bfp16 // convert fp32 to bfp16
template <> template <>
inline __host__ __device__ ushort type_convert(float x) inline __host__ __device__ bhalf_t type_convert(float x)
{ {
union union
{ {
......
composable_kernel/include/utility/type.hpp
View file @ 7e9a9d32
#ifndef CK_TYPE_HPP #ifndef CK_TYPE_HPP
#define CK_TYPE_HPP #define CK_TYPE_HPP
#include "config.hpp"
#include "integral_constant.hpp" #include "integral_constant.hpp"
#include "enable_if.hpp" #include "enable_if.hpp"
......
device_operation/CMakeLists.txt
View file @ 7e9a9d32
...@@ -22,6 +22,8 @@ set(DEVICE_GEMM_INSTANCE_SOURCE ...@@ -22,6 +22,8 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp;
......
device_operation/src/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp 0 → 100644
View file @ 7e9a9d32
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using BF16 = ck::bhalf_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;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
// clang-format off
//#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instance.cpp 0 → 100644
View file @ 7e9a9d32
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_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;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances = std::tuple<
// clang-format off
//#####################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
example/1_gemm_xdl/gemm_xdl_bf16.cpp 0 → 100644
View file @ 7e9a9d32
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = BF16;
using CDataType = BF16;
using AccDataType = F32;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
PassThrough, // AElementwiseOperation
PassThrough, // BElementwiseOperation
PassThrough, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, PassThrough, PassThrough, PassThrough>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{};
auto c_element_op = PassThrough{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_f32_m_k, b_f32_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_f32_result);
}
return 0;
}
example/1_gemm_xdl/gemm_xdl_int8.cpp 0 → 100644
View file @ 7e9a9d32
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
PassThrough, // AElementwiseOperation
PassThrough, // BElementwiseOperation
PassThrough, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = PassThrough{};
auto b_element_op = PassThrough{};
auto c_element_op = PassThrough{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result);
}
return 0;
}
example/CMakeLists.txt
View file @ 7e9a9d32
...@@ -14,6 +14,8 @@ include_directories(BEFORE ...@@ -14,6 +14,8 @@ include_directories(BEFORE
) )
set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp) set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
set(GEMM_XDL_INT8_SOURCE 1_gemm_xdl/gemm_xdl_int8.cpp)
set(GEMM_XDL_BF16_SOURCE 1_gemm_xdl/gemm_xdl_bf16.cpp)
set(GEMM_XDL_BIAS_RELU_SOURCE 2_gemm_xdl_bias_relu/gemm_xdl_bias_relu.cpp) set(GEMM_XDL_BIAS_RELU_SOURCE 2_gemm_xdl_bias_relu/gemm_xdl_bias_relu.cpp)
set(GEMM_XDL_BIAS_RELU_ADD_SOURCE 3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp) set(GEMM_XDL_BIAS_RELU_ADD_SOURCE 3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp)
set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp) set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
...@@ -27,6 +29,8 @@ set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp) ...@@ -27,6 +29,8 @@ set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp)
set(CONV2D_BWD_DATA_XDL_SOURCE 12_conv2d_bwd_data_xdl/conv2d_bwd_data_xdl.cpp) set(CONV2D_BWD_DATA_XDL_SOURCE 12_conv2d_bwd_data_xdl/conv2d_bwd_data_xdl.cpp)
add_executable(gemm_xdl ${GEMM_XDL_SOURCE}) add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
add_executable(gemm_xdl_int8 ${GEMM_XDL_INT8_SOURCE})
add_executable(gemm_xdl_bf16 ${GEMM_XDL_BF16_SOURCE})
add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE}) add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE}) add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE}) add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
...@@ -40,6 +44,8 @@ add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE}) ...@@ -40,6 +44,8 @@ add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
add_executable(conv2d_bwd_data_xdl ${CONV2D_BWD_DATA_XDL_SOURCE}) add_executable(conv2d_bwd_data_xdl ${CONV2D_BWD_DATA_XDL_SOURCE})
target_link_libraries(gemm_xdl PRIVATE host_tensor) target_link_libraries(gemm_xdl PRIVATE host_tensor)
target_link_libraries(gemm_xdl_int8 PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bf16 PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor) target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor) target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor) target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)
......
host/driver_offline/src/conv_fwd_driver_offline.cpp
View file @ 7e9a9d32
...@@ -77,7 +77,7 @@ void host_convolution_forward(const Tensor<TIn>& in, ...@@ -77,7 +77,7 @@ void host_convolution_forward(const Tensor<TIn>& in,
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 && if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3]) wi < in.mDesc.GetLengths()[3])
{ {
if constexpr(is_same<TIn, ushort>::value) if constexpr(is_same<TIn, bhalf_t>::value)
{ {
v += ck::type_convert<float>(in(n, c, hi, wi)) * v += ck::type_convert<float>(in(n, c, hi, wi)) *
ck::type_convert<float>(wei(k, c, y, x)); ck::type_convert<float>(wei(k, c, y, x));
...@@ -92,9 +92,9 @@ void host_convolution_forward(const Tensor<TIn>& in, ...@@ -92,9 +92,9 @@ void host_convolution_forward(const Tensor<TIn>& in,
} }
} }
if constexpr(is_same<TOut, ushort>::value) if constexpr(is_same<TOut, bhalf_t>::value)
{ {
out(n, k, ho, wo) = ck::type_convert<ushort>(static_cast<float>(v)); out(n, k, ho, wo) = ck::type_convert<bhalf_t>(static_cast<float>(v));
} }
else else
{ {
...@@ -115,7 +115,7 @@ void host_convolution_forward(const Tensor<TIn>& in, ...@@ -115,7 +115,7 @@ void host_convolution_forward(const Tensor<TIn>& in,
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 && if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2]) wi < in.mDesc.GetLengths()[2])
{ {
if constexpr(is_same<TIn, ushort>::value) if constexpr(is_same<TIn, bhalf_t>::value)
{ {
v += ck::type_convert<float>(in(n, hi, wi, c)) * v += ck::type_convert<float>(in(n, hi, wi, c)) *
ck::type_convert<float>(wei(k, y, x, c)); ck::type_convert<float>(wei(k, y, x, c));
...@@ -129,9 +129,9 @@ void host_convolution_forward(const Tensor<TIn>& in, ...@@ -129,9 +129,9 @@ void host_convolution_forward(const Tensor<TIn>& in,
} }
} }
} }
if constexpr(is_same<TOut, ushort>::value) if constexpr(is_same<TOut, bhalf_t>::value)
{ {
out(n, ho, wo, k) = ck::type_convert<ushort>(static_cast<float>(v)); out(n, ho, wo, k) = ck::type_convert<bhalf_t>(static_cast<float>(v));
} }
else else
{ {
...@@ -259,9 +259,9 @@ int main(int argc, char* argv[]) ...@@ -259,9 +259,9 @@ int main(int argc, char* argv[])
using acc_data_t = float; using acc_data_t = float;
using out_data_t = half_t; using out_data_t = half_t;
#elif 0 #elif 0
using in_data_t = ushort; using in_data_t = bhalf_t;
using acc_data_t = float; using acc_data_t = float;
using out_data_t = ushort; using out_data_t = bhalf_t;
#elif 1 #elif 1
using in_data_t = int8_t; using in_data_t = int8_t;
using acc_data_t = int32_t; using acc_data_t = int32_t;
......
host/host_tensor/CMakeLists.txt
View file @ 7e9a9d32
include_directories(BEFORE include_directories(BEFORE
${PROJECT_SOURCE_DIR}/composable_kernel/include
${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
include include
) )
......
host/host_tensor/include/host_tensor.hpp
View file @ 7e9a9d32
...@@ -8,6 +8,7 @@ ...@@ -8,6 +8,7 @@
#include <utility> #include <utility>
#include <cassert> #include <cassert>
#include <iostream> #include <iostream>
#include "data_type.hpp"
template <typename Range> template <typename Range>
std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim) std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
...@@ -311,7 +312,9 @@ HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> s ...@@ -311,7 +312,9 @@ HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> s
void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout); void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os = std::cout);
float bf16_to_f32_(ushort src_val); float bf16_to_f32_(ck::bhalf_t src_val);
void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst);
template <typename T> template <typename T>
void check_error(const Tensor<T>& ref, const Tensor<T>& result) void check_error(const Tensor<T>& ref, const Tensor<T>& result)
...@@ -320,7 +323,7 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result) ...@@ -320,7 +323,7 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result)
float max_diff = -1; float max_diff = -1;
float ref_value = 0, result_value = 0; float ref_value = 0, result_value = 0;
if constexpr(std::is_same<ushort, T>::value) if constexpr(std::is_same<ck::bhalf_t, T>::value)
{ {
for(int i = 0; i < ref.mData.size(); ++i) for(int i = 0; i < ref.mData.size(); ++i)
{ {
......
host/host_tensor/include/host_tensor_generator.hpp
View file @ 7e9a9d32
...@@ -3,7 +3,6 @@ ...@@ -3,7 +3,6 @@
#include <cmath> #include <cmath>
#include "config.hpp" #include "config.hpp"
#include "data_type.hpp"
template <typename T> template <typename T>
struct GeneratorTensor_0 struct GeneratorTensor_0
...@@ -28,14 +27,14 @@ struct GeneratorTensor_1 ...@@ -28,14 +27,14 @@ struct GeneratorTensor_1
}; };
template <> template <>
struct GeneratorTensor_1<ushort> struct GeneratorTensor_1<ck::bhalf_t>
{ {
float value = 1.0; float value = 1.0;
template <typename... Is> template <typename... Is>
ushort operator()(Is...) ck::bhalf_t operator()(Is...)
{ {
return ck::type_convert<ushort>(value); return ck::type_convert<ck::bhalf_t>(value);
} }
}; };
...@@ -65,16 +64,16 @@ struct GeneratorTensor_2 ...@@ -65,16 +64,16 @@ struct GeneratorTensor_2
}; };
template <> template <>
struct GeneratorTensor_2<ushort> struct GeneratorTensor_2<ck::bhalf_t>
{ {
int min_value = 0; int min_value = 0;
int max_value = 1; int max_value = 1;
template <typename... Is> template <typename... Is>
ushort operator()(Is...) ck::bhalf_t operator()(Is...)
{ {
float tmp = (std::rand() % (max_value - min_value)) + min_value; float tmp = (std::rand() % (max_value - min_value)) + min_value;
return ck::type_convert<ushort>(tmp); return ck::type_convert<ck::bhalf_t>(tmp);
} }
}; };
...@@ -107,19 +106,19 @@ struct GeneratorTensor_3 ...@@ -107,19 +106,19 @@ struct GeneratorTensor_3
}; };
template <> template <>
struct GeneratorTensor_3<ushort> struct GeneratorTensor_3<ck::bhalf_t>
{ {
float min_value = 0; float min_value = 0;
float max_value = 1; float max_value = 1;
template <typename... Is> template <typename... Is>
ushort operator()(Is...) ck::bhalf_t operator()(Is...)
{ {
float tmp = float(std::rand()) / float(RAND_MAX); float tmp = float(std::rand()) / float(RAND_MAX);
float fp32_tmp = min_value + tmp * (max_value - min_value); float fp32_tmp = min_value + tmp * (max_value - min_value);
return ck::type_convert<ushort>(fp32_tmp); return ck::type_convert<ck::bhalf_t>(fp32_tmp);
} }
}; };
......
host/host_tensor/src/host_tensor.cpp
View file @ 7e9a9d32
#include <cassert> #include <cassert>
#include "host_tensor.hpp" #include "host_tensor.hpp"
void HostTensorDescriptor::CalculateStrides() void HostTensorDescriptor::CalculateStrides()
...@@ -65,7 +64,7 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream ...@@ -65,7 +64,7 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream
os << "}" << std::endl; os << "}" << std::endl;
} }
float bf16_to_f32_(ushort src_val) float bf16_to_f32_(ck::bhalf_t src_val)
{ {
union union
{ {
...@@ -74,3 +73,9 @@ float bf16_to_f32_(ushort src_val) ...@@ -74,3 +73,9 @@ float bf16_to_f32_(ushort src_val)
} u = {uint32_t(src_val) << 16}; } u = {uint32_t(src_val) << 16};
return u.fp32; return u.fp32;
} }
void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst)
{
for(int i = 0; i < src.mData.size(); ++i)
dst.mData[i] = bf16_to_f32_(src.mData[i]);
}
profiler/include/profile_conv_fwd_impl.hpp
View file @ 7e9a9d32
...@@ -174,9 +174,9 @@ void profile_conv_fwd_impl(int do_verification, ...@@ -174,9 +174,9 @@ void profile_conv_fwd_impl(int do_verification,
ck::tensor_operation::device::device_conv2d_fwd_instance:: ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs); add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
} }
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ushort> && else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ushort> && ck::is_same_v<ck::remove_cv_t<WeiDataType>, bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ushort>) ck::is_same_v<ck::remove_cv_t<OutDataType>, bhalf_t>)
{ {
ck::tensor_operation::device::device_conv2d_fwd_instance:: ck::tensor_operation::device::device_conv2d_fwd_instance::
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
......
profiler/include/profile_gemm_impl.hpp
View file @ 7e9a9d32
...@@ -26,11 +26,17 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNo ...@@ -26,11 +26,17 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
...@@ -91,12 +97,11 @@ void profile_gemm_impl(int do_verification, ...@@ -91,12 +97,11 @@ void profile_gemm_impl(int do_verification,
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method) switch(init_method)
...@@ -122,19 +127,10 @@ void profile_gemm_impl(int do_verification, ...@@ -122,19 +127,10 @@ void profile_gemm_impl(int do_verification,
const auto b_element_op = BElementOp{}; const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{}; const auto c_element_op = CElementOp{};
if(do_verification) // if(do_verification)
{ // {
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
auto ref_gemm = ReferenceGemmInstance{}; // }
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace()); DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace()); DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
...@@ -290,6 +286,29 @@ void profile_gemm_impl(int do_verification, ...@@ -290,6 +286,29 @@ void profile_gemm_impl(int do_verification,
} }
} }
} }
else if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
is_same<CDataType, int8_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0) if(gemm_ptrs.size() <= 0)
{ {
...@@ -351,14 +370,79 @@ void profile_gemm_impl(int do_verification, ...@@ -351,14 +370,79 @@ void profile_gemm_impl(int do_verification,
{ {
c_device_buf.FromDevice(c_m_n_device_result.mData.data()); c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
is_same<BDataType, ck::bhalf_t>::value &&
is_same<CDataType, ck::bhalf_t>::value)
{
Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<float> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<float> c_m_n_device_f32_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
bf16_to_f32_(a_m_k, a_f32_m_k);
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_f32_m_k,
b_f32_k_n,
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_f32_result);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
else
{
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
check_error(c_m_n_host_result, c_m_n_device_result); check_error(c_m_n_host_result, c_m_n_device_result);
if(do_log)
{
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
if(do_log) if(do_log)
{ {
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl; LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl; LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",") LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl; << std::endl;
} }
......
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