/external/XNNPACK/src/qc8-gemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 99 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 102 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 114 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 116 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 117 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 121 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 136 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 139 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 135 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 140 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 158 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 162 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 100 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 103 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 112 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 114 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 118 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 122 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 134 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 137 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 136 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 141 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 156 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 160 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 101 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 104 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 113 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx() 115 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 119 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 123 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 135 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx() 138 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 137 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 142 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 157 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 161 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 113 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 116 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 125 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 127 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 133 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 137 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 149 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 152 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 153 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 158 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 173 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 177 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 114 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 117 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 126 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 128 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 134 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 138 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 150 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 153 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 154 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 159 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 174 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 178 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 2x16c8-minmax-fp32-avx512skx.c | 112 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() local 115 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 127 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx() 129 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx()
|
D | 3x16c8-minmax-fp32-avx512skx.c | 132 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() local 136 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 151 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx() 154 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx()
|
D | 4x16c8-minmax-fp32-avx512skx.c | 152 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 157 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 175 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 179 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx()
|
/external/XNNPACK/src/amalgam/ |
D | avx512skx.c | 1596 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 1601 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 1619 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 1623 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() 1908 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 1913 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 1931 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 1935 …F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_… in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx() 3133 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() local 3138 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F); in xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx() [all …]
|