File size: 10,190 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 |
// Copyright (c) Microsoft Corporation.
// SPDX-License-Identifier: Apache-2.0
// DeepSpeed Team
#pragma once
#if (__x86_64__ || __i386__)
#include <cpuid.h>
#include <x86intrin.h>
#endif
#define TILE (128 * 1024 * 1024)
#if defined(__AVX512__) or defined(__AVX256__)
#include <immintrin.h>
template <typename T>
inline T readAs(const void* src)
{
T res;
std::memcpy(&res, src, sizeof(T));
return res;
}
template <typename T>
inline void writeAs(void* dst, const T& val)
{
std::memcpy(dst, &val, sizeof(T));
}
#define ROUND_DOWN(size, step) ((size) & ~((step) - 1))
#if defined(__AVX512__)
#define SIMD_STORE(a, d) _mm512_storeu_ps(a, d)
#define SIMD_LOAD(x) _mm512_loadu_ps(x)
#define SIMD_SET(x) _mm512_set1_ps(x)
#define SIMD_ADD(x, y) _mm512_add_ps(x, y)
#define SIMD_MUL(x, y) _mm512_mul_ps(x, y)
#define SIMD_FMA(x, y, c) _mm512_fmadd_ps(x, y, c)
#define SIMD_SQRT(x) _mm512_sqrt_ps(x)
#define SIMD_DIV(x, y) _mm512_div_ps(x, y)
#define SIMD_AND(x, y) _mm512_and_ps(x, y)
#define SIMD_ANDNOT(x, y) _mm512_andnot_ps(x, y)
#define SIMD_OR(x, y) _mm512_or_ps(x, y)
#define SIMD_XOR(x, y) _mm512_xor_ps(x, y)
#define SIMD_WIDTH 16
static __m512 load_16_bf16_as_f32(const void* data)
{
__m256i a = readAs<__m256i>(data); // use memcpy to avoid aliasing
__m512i b = _mm512_cvtepu16_epi32(a); // convert 8 u16 to 8 u32
__m512i c = _mm512_slli_epi32(b, 16); // logical shift left of all u32 by
// 16 bits (representing bf16->f32)
return readAs<__m512>(&c); // use memcpy to avoid aliasing
}
static void store_16_f32_as_bf16_nearest(__m512 v, void* data)
{
__m512i u32 = readAs<__m512i>(&v);
// flow assuming non-nan:
// uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF);
__m512i b = _mm512_srli_epi32(u32, 16);
__m512i lsb_mask = _mm512_set1_epi32(0x00000001);
__m512i c = _mm512_and_si512(b, lsb_mask);
__m512i bias_constant = _mm512_set1_epi32(0x00007fff);
__m512i rounding_bias = _mm512_add_epi32(c, bias_constant);
// uint16_t res = static_cast<uint16_t>((U32 + rounding_bias) >> 16);
__m512i d = _mm512_add_epi32(u32, rounding_bias);
__m512i e = _mm512_srli_epi32(d, 16);
__m256i non_nan_res = _mm512_cvtusepi32_epi16(e);
// handle nan (exp is all 1s and mantissa != 0)
// if ((x & 0x7fffffffU) > 0x7f800000U)
__m512i mask_out_sign = _mm512_set1_epi32(0x7fffffff);
__m512i non_sign_bits = _mm512_and_si512(u32, mask_out_sign);
__m512i nan_threshold = _mm512_set1_epi32(0x7f800000);
__mmask16 nan_mask = _mm512_cmp_epi32_mask(non_sign_bits, nan_threshold, _MM_CMPINT_GT);
// mix in results with nans as needed
__m256i nans = _mm256_set1_epi16(0x7fc0);
__m256i res = _mm256_mask_mov_epi16(non_nan_res, nan_mask, nans);
writeAs(data, res);
}
#define SIMD_LOAD_BF16(x) load_16_bf16_as_f32(x)
#define SIMD_STORE_BF16(x, d) store_16_f32_as_bf16_nearest(d, x)
#define SIMD_LOAD_FP16(x) _mm512_cvtph_ps(_mm256_castps_si256(_mm256_loadu_ps(x)))
#define SIMD_STORE_FP16(x, d) \
_mm256_store_ps(x, _mm256_castsi256_ps(_mm512_cvtps_ph(d, _MM_FROUND_TO_NEAREST_INT)))
#define INTV __m256i
#elif defined(__AVX256__)
#define SIMD_STORE(a, d) _mm256_storeu_ps(a, d)
#define SIMD_LOAD(x) _mm256_loadu_ps(x)
#define SIMD_SET(x) _mm256_set1_ps(x)
#define SIMD_ADD(x, y) _mm256_add_ps(x, y)
#define SIMD_MUL(x, y) _mm256_mul_ps(x, y)
#define SIMD_FMA(x, y, c) _mm256_fmadd_ps(x, y, c)
#define SIMD_SQRT(x) _mm256_sqrt_ps(x)
#define SIMD_DIV(x, y) _mm256_div_ps(x, y)
#define SIMD_AND(x, y) _mm256_and_ps(x, y)
#define SIMD_ANDNOT(x, y) _mm256_andnot_ps(x, y)
#define SIMD_OR(x, y) _mm256_or_ps(x, y)
#define SIMD_XOR(x, y) _mm256_xor_ps(x, y)
#define SIMD_WIDTH 8
#define SIMD_LOAD_BF16(x) static_assert(false && "AVX256 does not support BFloat16")
#define SIMD_STORE_BF16(x, d) static_assert(false && "AVX256 does not support BFloat16")
#define SIMD_LOAD_FP16(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*)x))
#define SIMD_STORE_FP16(x, d) \
_mm_store_ps(x, _mm_castsi128_ps(_mm256_cvtps_ph(d, _MM_FROUND_TO_NEAREST_INT)))
#define INTV __m128i
#endif
union AVX_Data {
#if defined(__AVX512__)
__m512 data;
#elif defined(__AVX256__)
__m256 data;
#endif
// float data_f[16];
};
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, c10::Half>, void> simd_store(T* dst,
AVX_Data* src)
{
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { SIMD_STORE_FP16((float*)(dst + width * i), src[i].data); }
}
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, c10::BFloat16>, void> simd_store(T* dst,
AVX_Data* src)
{
#ifdef __AVX512__
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { SIMD_STORE_BF16((float*)(dst + width * i), src[i].data); }
#else
throw std::runtime_error("AVX512 required for BFloat16");
#endif
}
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, float>, void> simd_store(T* dst, AVX_Data* src)
{
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { SIMD_STORE(dst + width * i, src[i].data); }
}
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, c10::Half>, void> simd_load(AVX_Data* dst,
T* src)
{
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_LOAD_FP16((float*)(src + width * i)); }
}
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, c10::BFloat16>, void> simd_load(AVX_Data* dst,
T* src)
{
#ifdef __AVX512__
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_LOAD_BF16((float*)(src + width * i)); }
#else
throw std::runtime_error("AVX512 required for BFloat16");
#endif
}
template <int span, typename T>
inline typename std::enable_if_t<std::is_same_v<T, float>, void> simd_load(AVX_Data* dst, T* src)
{
size_t width = SIMD_WIDTH;
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_LOAD(src + width * i); }
}
template <int span>
inline void simd_fma(AVX_Data* dst, AVX_Data* src_m_l, AVX_Data src_m_r, AVX_Data* src_a)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) {
dst[i].data = SIMD_FMA(src_m_l[i].data, src_m_r.data, src_a[i].data);
}
}
template <int span>
inline void simd_fma(AVX_Data* dst, AVX_Data* src_m_l, AVX_Data src_m_r, AVX_Data src_a)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) {
dst[i].data = SIMD_FMA(src_m_l[i].data, src_m_r.data, src_a.data);
}
}
template <int span>
inline void simd_fma(AVX_Data* dst, AVX_Data* src_m_l, AVX_Data* src_m_r, AVX_Data* src_a)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) {
dst[i].data = SIMD_FMA(src_m_l[i].data, src_m_r[i].data, src_a[i].data);
}
}
template <int span>
inline void simd_sqrt(AVX_Data* dst, AVX_Data* src)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_SQRT(src[i].data); }
}
template <int span>
inline void simd_add(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_ADD(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_add(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_ADD(src_a_l[i].data, src_a_r[i].data); }
}
template <int span>
inline void simd_mul(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_MUL(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_mul(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_MUL(src_a_l[i].data, src_a_r[i].data); }
}
template <int span>
inline void simd_div(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_DIV(src_a_l[i].data, src_a_r[i].data); }
}
template <int span>
inline void simd_and(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_AND(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_and(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_AND(src_a_l[i].data, src_a_r[i].data); }
}
template <int span>
inline void simd_andnot(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_ANDNOT(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_andnot(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) {
dst[i].data = SIMD_ANDNOT(src_a_l[i].data, src_a_r[i].data);
}
}
template <int span>
inline void simd_or(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_OR(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_or(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_OR(src_a_l[i].data, src_a_r[i].data); }
}
template <int span>
inline void simd_xor(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_XOR(src_a_l[i].data, src_a_r.data); }
}
template <int span>
inline void simd_xor(AVX_Data* dst, AVX_Data* src_a_l, AVX_Data* src_a_r)
{
#pragma unroll
for (size_t i = 0; i < span; ++i) { dst[i].data = SIMD_XOR(src_a_l[i].data, src_a_r[i].data); }
}
#endif
|