/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-sse41-c8-acc2.c | 93 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 97 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() 160 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 164 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c8-acc2.c | 101 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 105 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() 181 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 185 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
D | 7x-minmax-sse2-c8-acc2.c | 104 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 108 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 187 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 191 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
D | 7x-minmax-sse41-c16-acc2.c | 110 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 116 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() 193 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 197 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7p7x-minmax-sse41-c8-acc2.c | 188 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 192 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() 256 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 260 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 126 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 132 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() 226 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 230 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 127 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 135 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() 228 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 232 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 131 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 137 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 234 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 238 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 212 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 216 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() 293 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 297 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
D | 7p7x-minmax-sse2-c8-acc2.c | 215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 219 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 299 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 303 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
D | 7p7x-minmax-sse41-c16-acc2.c | 239 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 245 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 323 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 327 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 158 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 166 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 283 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 287 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-ssse3-c24-acc2.c | 151 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 159 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() 273 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 277 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 287 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 293 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 388 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 392 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 292 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 298 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 396 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 400 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 366 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 374 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 468 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 472 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 461 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 469 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 587 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 591 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-ssse3-c24-acc2.c | 454 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 462 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 577 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 581 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up8x9-minmax-sse2-mul16.c | 215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 217 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() 394 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 398 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
D | up8x9-minmax-ssse3-mul16.c | 215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 217 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() 394 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 398 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 293 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 295 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 518 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 522 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 293 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 295 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 518 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 522 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 371 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 373 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 632 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 636 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 371 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 373 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 632 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 636 const __m128i vabsprod13 = _mm_mul_epu32(vabsacc13, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|