Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add NEON and SVE implementations for Float16 conversions #3424

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add NEON and SVE implementations for Float16 conversions #3424

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion defs.bzl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
# All rights reserved.
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
Expand Down Expand Up @@ -143,7 +144,11 @@ def get_fbgemm_inline_avx512_srcs(msvc = False, buck = False):
return asm_srcs if not msvc else intrinsics_srcs

def get_fbgemm_inline_sve_srcs(msvc = False, buck = False):
intrinsics_srcs = ["src/FbgemmFP16UKernelsSve128.cc"]
intrinsics_srcs = [
"src/FbgemmFP16UKernelsSve128.cc",
"src/FbgemmFloat16ConvertNeon.cc",
"src/FbgemmFloat16ConvertSve.cc"
]

#FP16 kernels contain inline assembly and inline assembly syntax for MSVC is different.
asm_srcs = ["src/FbgemmFP16UKernelsSve128.cc"]
Expand Down
35 changes: 35 additions & 0 deletions include/fbgemm/FbgemmConvert.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
Expand Down Expand Up @@ -135,6 +136,26 @@ FBGEMM_API void FloatToFloat16_avx512(
size_t size,
bool do_clip = false);

/**
* @brief SVE implementation to convert fp32 numbers to fp16 numbers.
*
*/
FBGEMM_API void FloatToFloat16_sve(
const float* src,
float16* dst,
size_t size,
bool do_clip = false);

/**
* @brief NEON implementation to convert fp32 numbers to fp16 numbers.
*
*/
FBGEMM_API void FloatToFloat16_neon(
const float* src,
float16* dst,
size_t size,
bool do_clip = false);

/**
* @brief AVX2 implementation to convert fp16 numbers to fp32 numbers.
*
Expand All @@ -149,6 +170,20 @@ Float16ToFloat_avx2(const float16* src, float* dst, size_t size);
FBGEMM_API void
Float16ToFloat_avx512(const float16* src, float* dst, size_t size);

/**
* @brief SVE implementation to convert fp16 numbers to fp32 numbers.
*
*/
FBGEMM_API void
Float16ToFloat_sve(const float16* src, float* dst, size_t size);

/**
* @brief NEON implementation to convert fp16 numbers to fp32 numbers.
*
*/
FBGEMM_API void
Float16ToFloat_neon(const float16* src, float* dst, size_t size);

/**
* @brief Transform all entries in a matrix from fp32 to float16 and back to
* fp32.
Expand Down
15 changes: 15 additions & 0 deletions src/FbgemmFloat16Convert.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
Expand Down Expand Up @@ -39,10 +40,17 @@ void FloatToFloat16_simd(
bool do_clip) {
// Run time CPU detection
if (cpuinfo_initialize()) {
#ifdef __aarch64__
if (fbgemmHasArmSveSupport()) {
FloatToFloat16_sve(src, dst, size, do_clip);
} else if (fbgemmHasArmNeonSupport()) {
FloatToFloat16_neon(src, dst, size, do_clip);
#else
if (fbgemmHasAvx512Support()) {
FloatToFloat16_avx512(src, dst, size, do_clip);
} else if (fbgemmHasAvx2Support()) {
FloatToFloat16_avx2(src, dst, size, do_clip);
#endif
} else {
FloatToFloat16_ref(src, dst, size, do_clip);
return;
Expand All @@ -55,10 +63,17 @@ void FloatToFloat16_simd(
void Float16ToFloat_simd(const float16* src, float* dst, size_t size) {
// Run time CPU detection
if (cpuinfo_initialize()) {
#ifdef __aarch64__
if (fbgemmHasArmSveSupport()) {
Float16ToFloat_sve(src, dst, size);
} else if (fbgemmHasArmNeonSupport()) {
Float16ToFloat_neon(src, dst, size);
#else
if (fbgemmHasAvx512Support()) {
Float16ToFloat_avx512(src, dst, size);
} else if (fbgemmHasAvx2Support()) {
Float16ToFloat_avx2(src, dst, size);
#endif
} else {
Float16ToFloat_ref(src, dst, size);
return;
Expand Down
122 changes: 122 additions & 0 deletions src/FbgemmFloat16ConvertNeon.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,122 @@
/*
* SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliate <open-source-office@arm.com>
* SPDX-License-Identifier: BSD-3-Clause
*/

#include <arm_neon.h>
#define FBGEMM_EXPORTS
#include "fbgemm/FbgemmConvert.h"

namespace fbgemm {

void FloatToFloat16_neon(
const float* src,
float16* dst,
size_t size,
bool do_clip) {
if (do_clip) {
constexpr float FP16_MAX = 65504.f;
auto vpos = vdupq_n_f32(FP16_MAX);
auto vneg = vdupq_n_f32(-FP16_MAX);
size_t i = 0;
for (; i + 16 < size; i += 16) {
auto f32_vec1 = vld1q_f32(src + i);
auto f32_vec2 = vld1q_f32(src + i + 4);
auto f32_vec3 = vld1q_f32(src + i + 8);
auto f32_vec4 = vld1q_f32(src + i + 12);
f32_vec1 = vmaxq_f32(vminq_f32(f32_vec1, vpos), vneg);
f32_vec2 = vmaxq_f32(vminq_f32(f32_vec2, vpos), vneg);
f32_vec3 = vmaxq_f32(vminq_f32(f32_vec3, vpos), vneg);
f32_vec4 = vmaxq_f32(vminq_f32(f32_vec4, vpos), vneg);
auto f16_vec1 = vcvt_f16_f32(f32_vec1);
auto f16_vec2 = vcvt_f16_f32(f32_vec2);
auto f16_vec3 = vcvt_f16_f32(f32_vec3);
auto f16_vec4 = vcvt_f16_f32(f32_vec4);
vst1_f16((__fp16*)dst + i, f16_vec1);
vst1_f16((__fp16*)dst + i + 4, f16_vec2);
vst1_f16((__fp16*)dst + i + 8, f16_vec3);
vst1_f16((__fp16*)dst + i + 12, f16_vec4);
}
for (; i + 8 < size; i += 8) {
auto f32_vec1 = vld1q_f32(src + i);
auto f32_vec2 = vld1q_f32(src + i + 4);
f32_vec1 = vmaxq_f32(vminq_f32(f32_vec1, vpos), vneg);
f32_vec2 = vmaxq_f32(vminq_f32(f32_vec2, vpos), vneg);
auto f16_vec1 = vcvt_f16_f32(f32_vec1);
auto f16_vec2 = vcvt_f16_f32(f32_vec2);
vst1_f16((__fp16*)dst + i, f16_vec1);
vst1_f16((__fp16*)dst + i + 4, f16_vec2);
}
for (; i + 4 < size; i += 4) {
auto f32_vec = vld1q_f32(src + i);
f32_vec = vmaxq_f32(vminq_f32(f32_vec, vpos), vneg);
auto f16_vec = vcvt_f16_f32(f32_vec);
vst1_f16((__fp16*)dst + i, f16_vec);
}
FloatToFloat16_ref(src + i, dst + i, size - i, do_clip);
} else {
size_t i = 0;
for (; i + 16 < size; i += 16) {
auto f32_vec1 = vld1q_f32(src + i);
auto f32_vec2 = vld1q_f32(src + i + 4);
auto f32_vec3 = vld1q_f32(src + i + 8);
auto f32_vec4 = vld1q_f32(src + i + 12);
auto f16_vec1 = vcvt_f16_f32(f32_vec1);
auto f16_vec2 = vcvt_f16_f32(f32_vec2);
auto f16_vec3 = vcvt_f16_f32(f32_vec3);
auto f16_vec4 = vcvt_f16_f32(f32_vec4);
vst1_f16((__fp16*)dst + i, f16_vec1);
vst1_f16((__fp16*)dst + i + 4, f16_vec2);
vst1_f16((__fp16*)dst + i + 8, f16_vec3);
vst1_f16((__fp16*)dst + i + 12, f16_vec4);
}
for (; i + 8 < size; i += 8) {
auto f32_vec1 = vld1q_f32(src + i);
auto f32_vec2 = vld1q_f32(src + i + 4);
auto f16_vec1 = vcvt_f16_f32(f32_vec1);
auto f16_vec2 = vcvt_f16_f32(f32_vec2);
vst1_f16((__fp16*)dst + i, f16_vec1);
vst1_f16((__fp16*)dst + i + 4, f16_vec2);
}
for (; i + 4 < size; i += 4) {
auto f32_vec = vld1q_f32(src + i);
auto f16_vec = vcvt_f16_f32(f32_vec);
vst1_f16((__fp16*)dst + i, f16_vec);
}
FloatToFloat16_ref(src + i, dst + i, size - i);
}
}

void Float16ToFloat_neon(const float16* src, float* dst, size_t size) {
size_t i = 0;
for (; i + 16 < size; i += 16) {
auto f16_vec1 = vld1_f16((__fp16*)src + i);
auto f16_vec2 = vld1_f16((__fp16*)src + i + 4);
auto f16_vec3 = vld1_f16((__fp16*)src + i + 8);
auto f16_vec4 = vld1_f16((__fp16*)src + i + 12);
auto f32_vec1 = vcvt_f32_f16(f16_vec1);
auto f32_vec2 = vcvt_f32_f16(f16_vec2);
auto f32_vec3 = vcvt_f32_f16(f16_vec3);
auto f32_vec4 = vcvt_f32_f16(f16_vec4);
vst1q_f32(dst + i, f32_vec1);
vst1q_f32(dst + i + 4, f32_vec2);
vst1q_f32(dst + i + 8, f32_vec3);
vst1q_f32(dst + i + 12, f32_vec4);
}
for (; i + 8 < size; i += 8) {
auto f16_vec1 = vld1_f16((__fp16*)src + i);
auto f16_vec2 = vld1_f16((__fp16*)src + i + 4);
auto f32_vec1 = vcvt_f32_f16(f16_vec1);
auto f32_vec2 = vcvt_f32_f16(f16_vec2);
vst1q_f32(dst + i, f32_vec1);
vst1q_f32(dst + i + 4, f32_vec2);
}
for (; i + 4 < size; i += 4) {
auto f16_vec = vld1_f16((__fp16*)src + i);
auto f32_vec = vcvt_f32_f16(f16_vec);
vst1q_f32(dst + i, f32_vec);
}
Float16ToFloat_ref(src + i, dst + i, size - i);
}

} // namespace fbgemm
99 changes: 99 additions & 0 deletions src/FbgemmFloat16ConvertSve.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,99 @@
/*
* SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliate <open-source-office@arm.com>
* SPDX-License-Identifier: BSD-3-Clause
*/

#include <arm_sve.h>
#define FBGEMM_EXPORTS
#include "fbgemm/FbgemmConvert.h"

namespace fbgemm {

void FloatToFloat16_sve(
const float* src,
float16* dst,
size_t size,
bool do_clip) {
if (do_clip) {
constexpr float FP16_MAX = 65504.f;
size_t i = 0;
int lanes = svcntw();
auto p_32 = svptrue_b32();
auto p_16 = svptrue_b16();
auto pfalse = svpfalse();
auto p_16_half = svuzp1_b16(p_16, pfalse);
while (i + 2 * lanes < size) {
auto f32_vec1 = svld1_f32(p_32, src + i);
auto f32_vec2 = svld1_f32(p_32, src + i + lanes);
f32_vec1 = svmax_n_f32_x(p_32, svmin_n_f32_x(p_32, f32_vec1, FP16_MAX), -FP16_MAX);
f32_vec2 = svmax_n_f32_x(p_32, svmin_n_f32_x(p_32, f32_vec2, FP16_MAX), -FP16_MAX);
auto f16_vec1 = svcvt_f16_f32_x(p_32, f32_vec1);
auto f16_vec2 = svcvt_f16_f32_x(p_32, f32_vec2);
auto f16_vec = svuzp1_f16(f16_vec1, f16_vec2);
svst1_f16(p_16, (__fp16*)dst + i, f16_vec);
i += 2 * lanes;
}
while (i + lanes < size) {
auto f32_vec = svld1_f32(p_32, src + i);
f32_vec = svmax_n_f32_x(p_32, svmin_n_f32_x(p_32, f32_vec, FP16_MAX), -FP16_MAX);
auto f16_vec = svcvt_f16_f32_x(p_16, f32_vec);
f16_vec = svuzp1_f16(f16_vec, f16_vec);
svst1_f16(p_16_half, (__fp16*)dst + i, f16_vec);
i += lanes;
}
FloatToFloat16_ref(src + i, dst + i, size - i, do_clip);
} else {
size_t i = 0;
int lanes = svcntw();
auto p_32 = svptrue_b32();
auto p_16 = svptrue_b16();
auto pfalse = svpfalse();
auto p_16_half = svuzp1_b16(p_16, pfalse);
while (i + 2 * lanes < size) {
auto f32_vec1 = svld1_f32(p_32, src + i);
auto f32_vec2 = svld1_f32(p_32, src + i + lanes);
auto f16_vec1 = svcvt_f16_f32_x(p_32, f32_vec1);
auto f16_vec2 = svcvt_f16_f32_x(p_32, f32_vec2);
auto f16_vec = svuzp1_f16(f16_vec1, f16_vec2);
svst1_f16(p_16, (__fp16*)dst + i, f16_vec);
i += 2 * lanes;
}
while (i + lanes < size) {
auto f32_vec = svld1_f32(p_32, src + i);
auto f16_vec = svcvt_f16_f32_x(p_32, f32_vec);
f16_vec = svuzp1_f16(f16_vec, f16_vec);
svst1_f16(p_16_half, (__fp16*)dst + i, f16_vec);
i += lanes;
}
FloatToFloat16_ref(src + i, dst + i, size - i);
}
}

void Float16ToFloat_sve(const float16* src, float* dst, size_t size) {
size_t i = 0;
int lanes = svcntw();
auto p_32 = svptrue_b32();
auto p_16 = svptrue_b16();
auto pfalse = svpfalse();
auto p_16_half = svuzp1_b16(p_16, pfalse);
while (i + 2 * lanes < size) {
auto f16_vec = svld1_f16(p_16, (__fp16*)src + i);
auto f16_vec1 = svzip1(f16_vec, f16_vec);
auto f16_vec2 = svzip2(f16_vec, f16_vec);
auto f32_vec1 = svcvt_f32_f16_x(p_16, f16_vec1);
auto f32_vec2 = svcvt_f32_f16_x(p_16, f16_vec2);
svst1_f32(p_32, dst + i, f32_vec1);
svst1_f32(p_32, dst + i + lanes, f32_vec2);
i += 2 * lanes;
}
while (i + lanes < size) {
auto f16_vec = svld1_f16(p_16_half, (__fp16*)src + i);
f16_vec = svzip1_f16(f16_vec, f16_vec);
auto f32_vec = svcvt_f32_f16_x(p_32, f16_vec);
svst1_f32(p_32, dst + i, f32_vec);
i += lanes;
}
Float16ToFloat_ref(src + i, dst + i, size - i);
}

} // namespace fbgemm