| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 2x16c2-minmax-rndnu-neon-mull-dup.c | 121 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
124 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
128 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
132 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
136 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
230 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
231 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
233 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
235 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
237 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
|
| D | 2x16c2-minmax-rndnu-neon-mull-ld1r.c | 127 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
130 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
134 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
138 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
142 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
236 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
237 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
239 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
241 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
243 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
|
| D | 2x16c2-minmax-rndnu-neon-mull-ld4r.c | 121 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
124 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
128 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
132 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
136 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
230 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
231 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
233 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
235 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
237 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
|
| D | 2x16c2-minmax-rndnu-neon-mull-ld2r.c | 123 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
126 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
130 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
134 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
138 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
232 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
233 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
235 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
237 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
239 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld4r.c | 93 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r() local
96 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
100 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
160 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r() local
161 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
163 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-dup.c | 93 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup() local
96 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
100 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
160 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup() local
161 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
163 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld1r.c | 99 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r() local
102 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
106 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
166 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r() local
167 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
169 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld2r.c | 95 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r() local
98 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
102 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
162 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r() local
163 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
165 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
|
| D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 150 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
154 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
160 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
166 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
172 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
296 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
297 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
299 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
301 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
303 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld2r.c | 153 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
157 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
163 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
169 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
175 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
299 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
300 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
302 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
304 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
306 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld4r.c | 150 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
154 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
160 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
166 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
172 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
296 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
297 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
299 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
301 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
303 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld1r.c | 159 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
163 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
169 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
175 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
181 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
305 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
306 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
308 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
310 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
312 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 2x16c2-minmax-rndnu-neon-mull-ld1r.c | 140 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
143 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
147 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
151 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
155 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
249 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
250 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
252 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
254 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
256 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
|
| D | 2x16c2-minmax-rndnu-neon-mull-dup.c | 134 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
137 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
141 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
145 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
149 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
243 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
244 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
246 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
248 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
250 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
|
| D | 2x16c2-minmax-rndnu-neon-mull-ld4r.c | 134 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
137 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
141 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
145 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
149 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
243 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
244 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
246 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
248 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
250 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
|
| D | 2x16c2-minmax-rndnu-neon-mull-ld2r.c | 136 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
139 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
143 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
147 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
151 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
245 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
246 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
248 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
250 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
252 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld4r.c | 106 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r() local
109 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
113 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
173 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r() local
174 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
176 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld2r.c | 108 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r() local
111 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
115 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
175 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r() local
176 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
178 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-ld1r.c | 112 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r() local
115 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
119 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
179 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r() local
180 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
182 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r()
|
| D | 2x8c2-minmax-rndnu-neon-mull-dup.c | 106 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup() local
109 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
113 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
173 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup() local
174 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
176 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld1r.c | 174 const int8x8_t va1c2 = vreinterpret_s8_s16(va12); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
178 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
184 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
190 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
196 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
320 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
321 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
323 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
325 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
327 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
|
| D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 165 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
169 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
175 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
181 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
187 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
311 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
312 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
314 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
316 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
318 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld4r.c | 165 const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
169 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
175 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
181 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
187 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
311 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
312 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
314 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
316 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
318 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
|
| D | 3x16c2-minmax-rndnu-neon-mull-ld2r.c | 168 const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
172 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
178 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
184 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
190 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
314 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
315 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
317 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
319 const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
321 const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
|
| D | 3x8c2-minmax-rndnu-neon-mull-dup.c | 127 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
131 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
137 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
215 const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
216 const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
218 const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup()
|