/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-sse41-c8-acc2.c | 94 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 99 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() 161 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 166 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c8-acc2.c | 102 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 107 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() 182 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 187 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
D | 7x-minmax-sse2-c8-acc2.c | 105 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 110 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 188 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 193 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 111 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 118 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() 194 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 199 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7p7x-minmax-sse41-c8-acc2.c | 189 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 194 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() 257 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 262 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 127 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 134 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() 227 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 232 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 128 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 137 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() 229 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 234 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 132 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 139 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 235 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 240 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 213 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 218 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() 294 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 299 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
D | 7p7x-minmax-sse2-c8-acc2.c | 216 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 221 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 300 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 305 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 240 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 247 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 324 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 329 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 159 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 168 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 284 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 289 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 152 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 161 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() 274 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 279 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 288 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 295 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 389 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 394 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 293 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 300 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 397 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 402 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 367 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 376 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 469 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 474 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 462 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 471 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 588 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 593 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 455 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 464 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 578 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 583 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up8x9-minmax-sse2-mul16.c | 218 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 220 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() 395 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 400 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
D | up8x9-minmax-ssse3-mul16.c | 218 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 220 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() 395 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 400 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 296 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 298 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 519 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 524 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 296 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 298 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 519 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 524 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 374 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 376 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 633 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 638 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 374 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 376 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 633 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 638 const __m128i vabsprod57 = _mm_mul_epu32(vabsacc57, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|