/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 165 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 203 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 207 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 211 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 215 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 219 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 223 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 227 vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb4567x89AB, va6x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 231 vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb4567x89AB, va7x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
|
D | 6x16c4-minmax-neondot.c | 139 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local 169 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 173 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 177 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 181 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 185 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 189 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
|
D | 4x16c4-minmax-neondot.c | 113 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local 135 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 139 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 143 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 147 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
|
D | 1x16c4-minmax-neondot.c | 74 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local 84 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 142 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 180 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 184 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 188 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 192 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 196 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 200 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 204 vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb4567x89AB, va6x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 208 vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb4567x89AB, va7x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
|
D | 6x16c4-minmax-neondot.c | 120 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local 150 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 154 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 158 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 162 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 166 vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 170 vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
|
D | 4x16c4-minmax-neondot.c | 98 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local 120 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 124 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 128 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 132 vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
|
D | 1x16c4-minmax-neondot.c | 65 const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local 75 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
|