Home
last modified time | relevance | path

Searched refs:vb89ABc1 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D4x16c2-minmax-rndnu-neon-mull-ld2r.c97 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
166 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
167 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
168 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
169 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
311 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
319 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
328 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
337 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
346 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mull-ld1r.c105 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
174 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
175 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
176 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
177 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
319 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
327 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
336 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
345 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
354 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
D4x16c2-minmax-rndnu-neon-mull-dup.c93 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
162 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
163 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
164 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
165 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
307 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
315 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
324 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
333 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
342 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mull-ld4r.c93 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
162 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
163 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
164 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
165 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
307 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
315 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
324 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
333 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
342 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-dup.c82 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
137 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
138 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
139 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
250 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
258 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
267 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
276 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
D3x16c2-minmax-rndnu-neon-mull-ld2r.c85 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
140 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
141 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
142 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
253 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
261 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
270 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
279 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
D3x16c2-minmax-rndnu-neon-mull-ld4r.c82 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
137 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
138 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
139 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
250 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
258 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
267 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
276 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-ld1r.c91 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
146 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
147 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
148 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
259 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
267 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
276 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
285 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c71 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
112 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
113 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
193 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
201 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
210 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c77 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
118 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
119 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
199 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
207 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
216 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c71 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
112 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
113 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
193 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
201 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
210 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
D2x16c2-minmax-rndnu-neon-mull-ld2r.c73 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
114 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
115 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
195 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
203 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
212 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
D1x16c2-minmax-rndnu-neon-mull-ld2r.c61 const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
88 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r()
137 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
145 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D4x16c2-minmax-rndnu-neon-mull-dup.c110 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
179 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
180 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
181 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
182 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
324 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
332 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
341 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
350 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
359 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mull-ld4r.c110 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
179 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
180 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
181 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
182 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
324 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
332 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
341 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
350 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
359 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
D4x16c2-minmax-rndnu-neon-mull-ld2r.c114 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
183 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
184 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
185 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
186 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
328 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
336 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
345 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
354 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
363 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mull-ld1r.c122 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
191 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
192 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
193 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
194 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
336 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
344 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
353 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
362 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
371 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
D3x16c2-minmax-rndnu-neon-mull-ld1r.c106 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
161 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
162 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
163 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
274 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r() local
282 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
291 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
300 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
D3x16c2-minmax-rndnu-neon-mull-dup.c97 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
152 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
153 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
154 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
265 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
273 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
282 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
291 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
D3x16c2-minmax-rndnu-neon-mull-ld4r.c97 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
152 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
153 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
154 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
265 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r() local
273 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
282 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
291 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-ld2r.c100 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
155 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
156 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
157 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
268 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
276 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
285 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
294 const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c90 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
131 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
132 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
212 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
220 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
229 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c84 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
125 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
126 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
206 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
214 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
223 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c84 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
125 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
126 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
206 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
214 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
223 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
D2x16c2-minmax-rndnu-neon-mull-ld2r.c86 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
127 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
128 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
208 … const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
216 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
225 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()

123