/external/XNNPACK/src/f32-qs8-vcvt/gen/ |
D | vcvt-avx512skx-x32.c | 70 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() local 71 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() 72 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() 93 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() local 94 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32() 95 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x32()
|
D | vcvt-avx512skx-x64.c | 80 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() local 81 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() 82 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() 103 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() local 104 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64() 105 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x64()
|
D | vcvt-avx512skx-x96.c | 95 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() local 96 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() 97 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() 118 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() local 119 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96() 120 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x96()
|
D | vcvt-avx512skx-x128.c | 104 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local 105 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 106 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 127 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local 128 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 129 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128()
|
D | vcvt-avx2-x16.c | 49 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qs8_vcvt_ukernel__avx2_x16() local 51 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx2_x16() 53 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx2_x16()
|
D | vcvt-avx2-x32.c | 58 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qs8_vcvt_ukernel__avx2_x32() local 61 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx2_x32() 64 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657); in xnn_f32_qs8_vcvt_ukernel__avx2_x32()
|
D | vcvt-avx2-x48.c | 66 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qs8_vcvt_ukernel__avx2_x48() local 70 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx2_x48() 74 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657); in xnn_f32_qs8_vcvt_ukernel__avx2_x48()
|
D | vcvt-avx2-x64.c | 74 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qs8_vcvt_ukernel__avx2_x64() local 79 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx2_x64() 84 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657); in xnn_f32_qs8_vcvt_ukernel__avx2_x64()
|
/external/XNNPACK/src/f32-qu8-vcvt/gen/ |
D | vcvt-avx512skx-x32.c | 70 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() local 71 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() 72 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() 93 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() local 94 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32() 95 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x32()
|
D | vcvt-avx512skx-x64.c | 80 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() local 81 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() 82 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() 103 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() local 104 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64() 105 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x64()
|
D | vcvt-avx512skx-x96.c | 95 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() local 96 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() 97 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() 118 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() local 119 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96() 120 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x96()
|
D | vcvt-avx512skx-x128.c | 104 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local 105 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() 106 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() 127 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local 128 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() 129 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128()
|
D | vcvt-avx2-x16.c | 49 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qu8_vcvt_ukernel__avx2_x16() local 51 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx2_x16() 53 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx2_x16()
|
D | vcvt-avx2-x32.c | 58 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qu8_vcvt_ukernel__avx2_x32() local 61 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx2_x32() 64 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657); in xnn_f32_qu8_vcvt_ukernel__avx2_x32()
|
D | vcvt-avx2-x48.c | 66 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qu8_vcvt_ukernel__avx2_x48() local 70 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx2_x48() 74 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657); in xnn_f32_qu8_vcvt_ukernel__avx2_x48()
|
D | vcvt-avx2-x64.c | 74 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qu8_vcvt_ukernel__avx2_x64() local 79 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx2_x64() 84 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657); in xnn_f32_qu8_vcvt_ukernel__avx2_x64()
|
/external/XNNPACK/src/f32-qs8-vcvt/ |
D | avx512skx.c.in | 109 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… 110 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); 111 …i vy0213 = ${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… 135 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… 136 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); 137 …i vy0213 = ${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213…
|
/external/XNNPACK/src/amalgam/ |
D | avx512skx.c | 172 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local 173 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 174 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 195 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() local 196 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 197 …m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qs8_vcvt_ukernel__avx512skx_x128() 290 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local 291 vacc0213 = _mm256_adds_epi16(vacc0213, _mm512_castsi512_si256(voutput_zero_point)); in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() 292 …128i vy0213 = _mm_packus_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() 313 …__m256i vacc0213 = _mm256_packs_epi32(_mm512_castsi512_si256(vacc0123), _mm512_extracti32x8_epi32(… in xnn_f32_qu8_vcvt_ukernel__avx512skx_x128() local [all …]
|
D | avx2.c | 1644 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qs8_vcvt_ukernel__avx2_x64() local 1649 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qs8_vcvt_ukernel__avx2_x64() 1654 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657); in xnn_f32_qs8_vcvt_ukernel__avx2_x64() 1770 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23); in xnn_f32_qu8_vcvt_ukernel__avx2_x64() local 1775 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point); in xnn_f32_qu8_vcvt_ukernel__avx2_x64() 1780 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657); in xnn_f32_qu8_vcvt_ukernel__avx2_x64()
|