| 1 |
/* |
| 2 |
* Copyright (c) 2020, Intel Corporation |
| 3 |
* |
| 4 |
* Permission is hereby granted, free of charge, to any person obtaining a |
| 5 |
* copy of this software and associated documentation files (the "Software"), |
| 6 |
* to deal in the Software without restriction, including without limitation |
| 7 |
* the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 8 |
* and/or sell copies of the Software, and to permit persons to whom the |
| 9 |
* Software is furnished to do so, subject to the following conditions: |
| 10 |
* |
| 11 |
* The above copyright notice and this permission notice shall be included |
| 12 |
* in all copies or substantial portions of the Software. |
| 13 |
* |
| 14 |
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS |
| 15 |
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 16 |
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 17 |
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 18 |
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 19 |
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 20 |
* OTHER DEALINGS IN THE SOFTWARE. |
| 21 |
*/ |
| 22 |
//! |
| 23 |
//! \file cm_mem_sse2_impl.cpp |
| 24 |
//! \brief Contains CM memory function implementations |
| 25 |
//! |
| 26 |
|
| 27 |
#include "cm_mem.h" |
| 28 |
#include "cm_mem_sse2_impl.h" |
| 29 |
|
| 30 |
#if defined(__SSE2__) || !(defined(LINUX) || defined(ANDROID)) |
| 31 |
|
| 32 |
#include <mmintrin.h> |
| 33 |
|
| 34 |
// GCC>=11 gives a warning when dividing sizeof(array) with sizeof(another_type) |
| 35 |
// in case division was mistyped and meant to calculate array size. |
| 36 |
// As that's not the case here, use extra divisor parenthesis to suppress |
| 37 |
// the warning. |
| 38 |
#define DQWORD_PER_PREFETCH(P) ( sizeof(P)/(sizeof(DQWORD)) ) |
| 39 |
|
| 40 |
void FastMemCopy_SSE2_movntdq_movdqa( |
| 41 |
void* dst, |
| 42 |
void* src, |
| 43 |
const size_t doubleQuadWords ) |
| 44 |
{ |
| 45 |
CM_ASSERT( IsAligned( dst, sizeof(DQWORD) ) ); |
| 46 |
CM_ASSERT( IsAligned( src, sizeof(DQWORD) ) ); |
| 47 |
|
| 48 |
|
| 49 |
const size_t doubleQuadWordsPerPrefetch = DQWORD_PER_PREFETCH(PREFETCH); |
| 50 |
|
| 51 |
// Prefetch the src data |
| 52 |
Prefetch( (uint8_t*)src ); |
| 53 |
Prefetch( (uint8_t*)src + sizeof(PREFETCH) ); |
| 54 |
|
| 55 |
// Convert to SSE2 registers |
| 56 |
__m128i* dst128i = (__m128i*)dst; |
| 57 |
__m128i* src128i = (__m128i*)src; |
| 58 |
|
| 59 |
size_t count = doubleQuadWords; |
| 60 |
|
| 61 |
// Copies a cacheline per loop iteration |
| 62 |
while( count >= doubleQuadWordsPerPrefetch ) |
| 63 |
{ |
| 64 |
Prefetch( (uint8_t*)src128i + 2 * sizeof(PREFETCH) ); |
| 65 |
|
| 66 |
count -= doubleQuadWordsPerPrefetch; |
| 67 |
|
| 68 |
// Copy cacheline of data |
| 69 |
for( size_t i = 0; i < doubleQuadWordsPerPrefetch; i++ ) |
| 70 |
{ |
| 71 |
_mm_stream_si128( dst128i++, |
| 72 |
_mm_load_si128( src128i++ ) ); |
| 73 |
} |
| 74 |
} |
| 75 |
|
| 76 |
// Copy DQWORD if not cacheline multiple |
| 77 |
while( count-- ) |
| 78 |
{ |
| 79 |
_mm_stream_si128( dst128i++, |
| 80 |
_mm_load_si128( src128i++ ) ); |
| 81 |
} |
| 82 |
} |
| 83 |
|
| 84 |
void FastMemCopy_SSE2_movdqu_movdqa( |
| 85 |
void* dst, |
| 86 |
void* src, |
| 87 |
const size_t doubleQuadWords ) |
| 88 |
{ |
| 89 |
CM_ASSERT( IsAligned( src, sizeof(DQWORD) ) ); |
| 90 |
|
| 91 |
const size_t doubleQuadWordsPerPrefetch = DQWORD_PER_PREFETCH(PREFETCH); |
| 92 |
|
| 93 |
// Prefetch the src data |
| 94 |
Prefetch( (uint8_t*)src ); |
| 95 |
Prefetch( (uint8_t*)src + sizeof(PREFETCH) ); |
| 96 |
|
| 97 |
// Convert to SSE2 registers |
| 98 |
__m128i* dst128i = (__m128i*)dst; |
| 99 |
__m128i* src128i = (__m128i*)src; |
| 100 |
|
| 101 |
size_t count = doubleQuadWords; |
| 102 |
|
| 103 |
// Copies a cacheline per loop iteration |
| 104 |
while( count >= doubleQuadWordsPerPrefetch ) |
| 105 |
{ |
| 106 |
Prefetch( (uint8_t*)src128i + 2 * sizeof(PREFETCH) ); |
| 107 |
|
| 108 |
count -= doubleQuadWordsPerPrefetch; |
| 109 |
|
| 110 |
// Copy cacheline of data |
| 111 |
for( size_t i = 0; i < doubleQuadWordsPerPrefetch; i++ ) |
| 112 |
{ |
| 113 |
_mm_storeu_si128( dst128i++, |
| 114 |
_mm_load_si128( src128i++ ) ); |
| 115 |
} |
| 116 |
} |
| 117 |
|
| 118 |
// Copy DQWORD if not cacheline multiple |
| 119 |
while( count-- ) |
| 120 |
{ |
| 121 |
_mm_storeu_si128( dst128i++, |
| 122 |
_mm_load_si128( src128i++ ) ); |
| 123 |
} |
| 124 |
} |
| 125 |
|
| 126 |
void FastMemCopy_SSE2_movntdq_movdqu( |
| 127 |
void* dst, |
| 128 |
const void* src, |
| 129 |
const size_t doubleQuadWords ) |
| 130 |
{ |
| 131 |
CM_ASSERT( IsAligned( dst, sizeof(DQWORD) ) ); |
| 132 |
|
| 133 |
const size_t doubleQuadWordsPerPrefetch = DQWORD_PER_PREFETCH(PREFETCH); |
| 134 |
|
| 135 |
// Prefetch the src data |
| 136 |
Prefetch( (uint8_t*)src ); |
| 137 |
Prefetch( (uint8_t*)src + sizeof(PREFETCH) ); |
| 138 |
|
| 139 |
// Convert to SSE2 registers |
| 140 |
__m128i* dst128i = (__m128i*)dst; |
| 141 |
__m128i* src128i = (__m128i*)src; |
| 142 |
|
| 143 |
size_t count = doubleQuadWords; |
| 144 |
|
| 145 |
// Copies a cacheline per loop iteration |
| 146 |
while( count >= doubleQuadWordsPerPrefetch ) |
| 147 |
{ |
| 148 |
Prefetch( (uint8_t*)src128i + 2 * sizeof(PREFETCH) ); |
| 149 |
|
| 150 |
count -= doubleQuadWordsPerPrefetch; |
| 151 |
|
| 152 |
// Copy cacheline of data |
| 153 |
for( size_t i = 0; i < doubleQuadWordsPerPrefetch; i++ ) |
| 154 |
{ |
| 155 |
_mm_stream_si128( dst128i++, |
| 156 |
_mm_loadu_si128( src128i++ ) ); |
| 157 |
} |
| 158 |
} |
| 159 |
|
| 160 |
// Copy DQWORD if not cacheline multiple |
| 161 |
while( count-- ) |
| 162 |
{ |
| 163 |
_mm_stream_si128( dst128i++, |
| 164 |
_mm_loadu_si128( src128i++ ) ); |
| 165 |
} |
| 166 |
} |
| 167 |
|
| 168 |
void FastMemCopy_SSE2_movdqu_movdqu( |
| 169 |
void* dst, |
| 170 |
const void* src, |
| 171 |
const size_t doubleQuadWords ) |
| 172 |
{ |
| 173 |
const size_t doubleQuadWordsPerPrefetch = DQWORD_PER_PREFETCH(PREFETCH); |
| 174 |
|
| 175 |
// Prefetch the src data |
| 176 |
Prefetch( (uint8_t*)src ); |
| 177 |
Prefetch( (uint8_t*)src + sizeof(PREFETCH) ); |
| 178 |
|
| 179 |
// Convert to SSE2 registers |
| 180 |
__m128i* dst128i = (__m128i*)dst; |
| 181 |
__m128i* src128i = (__m128i*)src; |
| 182 |
|
| 183 |
size_t count = doubleQuadWords; |
| 184 |
|
| 185 |
// Copies a cacheline per loop iteration |
| 186 |
while( count >= doubleQuadWordsPerPrefetch ) |
| 187 |
{ |
| 188 |
Prefetch( (uint8_t*)src128i + 2 * sizeof(PREFETCH) ); |
| 189 |
|
| 190 |
count -= doubleQuadWordsPerPrefetch; |
| 191 |
|
| 192 |
// Copy cacheline of data |
| 193 |
for( size_t i = 0; i < doubleQuadWordsPerPrefetch; i++ ) |
| 194 |
{ |
| 195 |
_mm_storeu_si128( dst128i++, |
| 196 |
_mm_loadu_si128( src128i++ ) ); |
| 197 |
} |
| 198 |
} |
| 199 |
|
| 200 |
// Copy DQWORD if not cacheline multiple |
| 201 |
while( count-- ) |
| 202 |
{ |
| 203 |
_mm_storeu_si128( dst128i++, |
| 204 |
_mm_loadu_si128( src128i++ ) ); |
| 205 |
} |
| 206 |
} |
| 207 |
|
| 208 |
void FastMemCopy_SSE2( |
| 209 |
void* dst, |
| 210 |
void* src, |
| 211 |
const size_t doubleQuadWords ) |
| 212 |
{ |
| 213 |
// Determine if the source and destination addresses are 128-bit aligned |
| 214 |
const bool isDstDoubleQuadWordAligned = IsAligned( dst, sizeof(DQWORD) ); |
| 215 |
const bool isSrcDoubleQuadWordAligned = IsAligned( src, sizeof(DQWORD) ); |
| 216 |
|
| 217 |
if( isSrcDoubleQuadWordAligned && isDstDoubleQuadWordAligned ) |
| 218 |
{ |
| 219 |
FastMemCopy_SSE2_movntdq_movdqa( dst, src, doubleQuadWords ); |
| 220 |
} |
| 221 |
else if( isDstDoubleQuadWordAligned ) |
| 222 |
{ |
| 223 |
FastMemCopy_SSE2_movntdq_movdqu( dst, src, doubleQuadWords ); |
| 224 |
} |
| 225 |
else if( isSrcDoubleQuadWordAligned ) |
| 226 |
{ |
| 227 |
FastMemCopy_SSE2_movdqu_movdqa( dst, src, doubleQuadWords ); |
| 228 |
} |
| 229 |
else // if( !isSrcDoubleQuadWordAligned && !isDstDoubleQuadWordAligned ) |
| 230 |
{ |
| 231 |
FastMemCopy_SSE2_movdqu_movdqu( dst, src, doubleQuadWords ); |
| 232 |
} |
| 233 |
} |
| 234 |
|
| 235 |
void CmFastMemCopy_SSE2( void* dst, const void* src, const size_t bytes ) |
| 236 |
{ |
| 237 |
// Cache pointers to memory |
| 238 |
uint8_t *cacheDst = (uint8_t*)dst; |
| 239 |
uint8_t *cacheSrc = (uint8_t*)src; |
| 240 |
|
| 241 |
size_t count = bytes; |
| 242 |
|
| 243 |
// Get the number of DQWORDs to be copied |
| 244 |
const size_t doubleQuadWords = count / sizeof(DQWORD); |
| 245 |
|
| 246 |
if( count >= CM_CPU_FASTCOPY_THRESHOLD && doubleQuadWords ) |
| 247 |
{ |
| 248 |
FastMemCopy_SSE2( cacheDst, cacheSrc, doubleQuadWords ); |
| 249 |
|
| 250 |
cacheDst += doubleQuadWords * sizeof(DQWORD); |
| 251 |
cacheSrc += doubleQuadWords * sizeof(DQWORD); |
| 252 |
count -= doubleQuadWords * sizeof(DQWORD); |
| 253 |
} |
| 254 |
|
| 255 |
// Copy remaining uint8_t(s) |
| 256 |
if( count ) |
| 257 |
{ |
| 258 |
MOS_SecureMemcpy( cacheDst, count, cacheSrc, count ); |
| 259 |
} |
| 260 |
} |
| 261 |
|
| 262 |
void CmFastMemCopyWC_SSE2( void* dst, const void* src, const size_t bytes ) |
| 263 |
{ |
| 264 |
// Cache pointers to memory |
| 265 |
uint8_t *cacheDst = (uint8_t*)dst; |
| 266 |
uint8_t *cacheSrc = (uint8_t*)src; |
| 267 |
|
| 268 |
size_t count = bytes; |
| 269 |
|
| 270 |
if( count >= CM_CPU_FASTCOPY_THRESHOLD ) |
| 271 |
{ |
| 272 |
const size_t doubleQuadwordAlignBytes = |
| 273 |
GetAlignmentOffset( cacheDst, sizeof(DQWORD) ); |
| 274 |
|
| 275 |
// The destination pointer should be 128-bit aligned |
| 276 |
if( doubleQuadwordAlignBytes ) |
| 277 |
{ |
| 278 |
MOS_SecureMemcpy( cacheDst, doubleQuadwordAlignBytes,cacheSrc, doubleQuadwordAlignBytes ); |
| 279 |
|
| 280 |
cacheDst += doubleQuadwordAlignBytes; |
| 281 |
cacheSrc += doubleQuadwordAlignBytes; |
| 282 |
count -= doubleQuadwordAlignBytes; |
| 283 |
} |
| 284 |
|
| 285 |
// Get the number of DQWORDs to be copied |
| 286 |
const size_t doubleQuadWords = count / sizeof(DQWORD); |
| 287 |
|
| 288 |
if( doubleQuadWords && count >= sizeof(PREFETCH)) |
| 289 |
{ |
| 290 |
// Determine if the source and destination addresses are |
| 291 |
// 128-bit aligned |
| 292 |
CM_ASSERT( IsAligned( cacheDst, sizeof(DQWORD) ) ); |
| 293 |
|
| 294 |
const bool isSrcDoubleQuadWordAligned = |
| 295 |
IsAligned( cacheSrc, sizeof(DQWORD) ); |
| 296 |
|
| 297 |
if( isSrcDoubleQuadWordAligned ) |
| 298 |
{ |
| 299 |
FastMemCopy_SSE2_movntdq_movdqa( cacheDst, cacheSrc, |
| 300 |
doubleQuadWords ); |
| 301 |
} |
| 302 |
else |
| 303 |
{ |
| 304 |
FastMemCopy_SSE2_movntdq_movdqu( cacheDst, cacheSrc, |
| 305 |
doubleQuadWords ); |
| 306 |
} |
| 307 |
|
| 308 |
cacheDst += doubleQuadWords * sizeof(DQWORD); |
| 309 |
cacheSrc += doubleQuadWords * sizeof(DQWORD); |
| 310 |
count -= doubleQuadWords * sizeof(DQWORD); |
| 311 |
} |
| 312 |
} |
| 313 |
|
| 314 |
// Copy remaining uint8_t(s) |
| 315 |
if( count ) |
| 316 |
{ |
| 317 |
MOS_SecureMemcpy( cacheDst, count, cacheSrc, count ); |
| 318 |
} |
| 319 |
} |
| 320 |
|
| 321 |
#endif // __SSE2__ || !(LINUX || ANDROID) |