ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-02 20:51:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Optimize fmha fwd decode & prefill for gfx950 (#2641)" (#2670)

Browse Source 
This reverts commit b7322a521a.
This commit is contained in:
asleepzzz
2025-08-12 20:27:10 +08:00
committed by GitHub
parent b7322a521a
commit 5b39de4bb6
31 changed files with 639 additions and 3545 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
include/ck_tile/core/numeric/bfloat16.hpp
View File

@@ -6,9 +6,6 @@
#include "ck_tile/core/numeric/half.hpp"
#include "ck_tile/core/numeric/integral_constant.hpp"
#include "ck_tile/core/numeric/numeric.hpp"
#if CK_TILE_USE_LLVM_BUILTIN_BF16
#include <hip/hip_bfloat16.h>
#endif
#include <stdint.h>
#pragma once
@@ -105,11 +102,7 @@ struct native_t<bfloat16_t>
using bf16_t = bfloat16_t;
using bf16_raw_t = typename bf16_t::raw_type;
#else
#if CK_TILE_USE_LLVM_BUILTIN_BF16
using bfloat16_t = __bf16;
#else
using bfloat16_t = ushort;
#endif
using bf16_t = bfloat16_t;
using bf16_raw_t = uint16_t;
#endif
@@ -287,11 +280,7 @@ template <bf16_rounding_mode rounding =
static_cast<bf16_rounding_mode>(CK_TILE_FLOAT_TO_BFLOAT16_DEFAULT)>
CK_TILE_HOST_DEVICE constexpr bfloat16_t float_to_bf16(float f, constant<rounding> = {})
{
#if defined(__gfx950__)
return static_cast<bfloat16_t>(f);
#else
return bit_cast<bfloat16_t>(float_to_bf16_raw(f, constant<rounding>{}));
#endif
}
template <bf16_rounding_mode rounding =

2
include/ck_tile/core/numeric/pk_fp4.hpp
View File

@@ -21,7 +21,7 @@ namespace ck_tile {
using fp32_t = float;
using fp32x2_t = float __attribute__((ext_vector_type(2)));
using fp16x2_t = _Float16 __attribute__((ext_vector_type(2)));
using bf16x2_t = bfloat16_t __attribute__((ext_vector_type(2)));
using bf16x2_t = bf16_raw_t __attribute__((ext_vector_type(2)));
CK_TILE_HOST_DEVICE constexpr uint8_t float_to_e2m1(float x, float scale = 1.f);

2
include/ck_tile/core/numeric/pk_int4.hpp
View File

@@ -99,7 +99,7 @@ struct numeric_traits<pk_int4_t>
using fp32x2_t = float __attribute__((ext_vector_type(2)));
using fp16x2_t = _Float16 __attribute__((ext_vector_type(2)));
using bf16x2_t = bfloat16_t __attribute__((ext_vector_type(2)));
using bf16x2_t = bf16_raw_t __attribute__((ext_vector_type(2)));
CK_TILE_HOST_DEVICE fp32x2_t pk_int4_t_to_fp32x2_t(const pk_int4_t& x)
{

12
include/ck_tile/core/numeric/vector_type.hpp
View File

@@ -131,12 +131,12 @@ using fp16x64_t = _Float16 __attribute__((ext_vector_type(64)));
// bf16
// using bf16_t = ...
using bf16x2_t = bfloat16_t __attribute__((ext_vector_type(2)));
using bf16x4_t = bfloat16_t __attribute__((ext_vector_type(4)));
using bf16x8_t = bfloat16_t __attribute__((ext_vector_type(8)));
using bf16x16_t = bfloat16_t __attribute__((ext_vector_type(16)));
using bf16x32_t = bfloat16_t __attribute__((ext_vector_type(32)));
using bf16x64_t = bfloat16_t __attribute__((ext_vector_type(64)));
using bf16x2_t = bf16_raw_t __attribute__((ext_vector_type(2)));
using bf16x4_t = bf16_raw_t __attribute__((ext_vector_type(4)));
using bf16x8_t = bf16_raw_t __attribute__((ext_vector_type(8)));
using bf16x16_t = bf16_raw_t __attribute__((ext_vector_type(16)));
using bf16x32_t = bf16_raw_t __attribute__((ext_vector_type(32)));
using bf16x64_t = bf16_raw_t __attribute__((ext_vector_type(64)));
// i32
// using int32_t = ...