/external/XNNPACK/src/qs8-igemm/gen/ |
D | 8x16c4-minmax-neondot.c | 166 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() local 204 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 208 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 212 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 216 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 220 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 224 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 228 vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb4567xCDEF, va6x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot() 232 vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb4567xCDEF, va7x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_8x16c4__neondot()
|
D | 6x16c4-minmax-neondot.c | 140 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() local 170 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 174 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 178 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 182 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 186 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot() 190 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot()
|
D | 4x16c4-minmax-neondot.c | 114 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() local 136 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 140 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 144 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot() 148 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot()
|
D | 1x16c4-minmax-neondot.c | 75 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot() local 85 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 8x16c4-minmax-neondot.c | 143 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() local 181 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 185 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 189 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 193 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 197 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 201 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 205 vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb4567xCDEF, va6x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot() 209 vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb4567xCDEF, va7x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_8x16c4__neondot()
|
D | 6x16c4-minmax-neondot.c | 121 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() local 151 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 155 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 159 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 163 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 167 vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot() 171 vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_6x16c4__neondot()
|
D | 4x16c4-minmax-neondot.c | 99 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() local 121 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 125 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 129 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot() 133 vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot()
|
D | 1x16c4-minmax-neondot.c | 66 const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot() local 76 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1); in xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot()
|