Home
last modified time | relevance | path

Searched refs:vbCDEFc0 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D4x16c2-minmax-rndnu-neon-mull-ld2r.c94 const int8x8_t vbCDEFc0 = 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
137 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
138 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
139 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
140 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
269 const int8x8_t vbCDEFc0 = 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
278 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
287 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
296 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
305 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mull-ld1r.c102 const int8x8_t vbCDEFc0 = 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
145 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
146 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
147 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
148 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
277 const int8x8_t vbCDEFc0 = 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
286 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
295 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
304 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
313 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
D4x16c2-minmax-rndnu-neon-mull-dup.c90 const int8x8_t vbCDEFc0 = 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
133 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
134 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
135 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
136 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
265 const int8x8_t vbCDEFc0 = 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
274 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
283 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
292 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
301 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mull-ld4r.c90 const int8x8_t vbCDEFc0 = 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
133 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
134 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
135 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
136 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
265 const int8x8_t vbCDEFc0 = 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
274 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
283 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
292 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
301 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-dup.c79 const int8x8_t vbCDEFc0 = 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
115 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
116 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
117 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
217 const int8x8_t vbCDEFc0 = 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
226 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
235 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
244 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
D3x16c2-minmax-rndnu-neon-mull-ld2r.c82 const int8x8_t vbCDEFc0 = 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
118 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
119 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
120 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
220 const int8x8_t vbCDEFc0 = 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
229 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
238 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
247 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
D3x16c2-minmax-rndnu-neon-mull-ld4r.c79 const int8x8_t vbCDEFc0 = 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
115 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
116 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
117 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
217 const int8x8_t vbCDEFc0 = 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
226 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
235 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
244 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-ld1r.c88 const int8x8_t vbCDEFc0 = 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
124 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
125 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
126 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
226 const int8x8_t vbCDEFc0 = 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
235 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
244 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
253 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c68 const int8x8_t vbCDEFc0 = 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
97 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
98 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
169 const int8x8_t vbCDEFc0 = 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
178 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
187 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c74 const int8x8_t vbCDEFc0 = 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
103 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
104 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
175 const int8x8_t vbCDEFc0 = 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
184 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
193 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c68 const int8x8_t vbCDEFc0 = 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
97 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
98 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
169 const int8x8_t vbCDEFc0 = 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
178 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
187 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
D2x16c2-minmax-rndnu-neon-mull-ld2r.c70 const int8x8_t vbCDEFc0 = 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
99 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
100 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
171 const int8x8_t vbCDEFc0 = 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
180 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
189 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
D1x16c2-minmax-rndnu-neon-mull-ld2r.c58 const int8x8_t vbCDEFc0 = 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
80 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r()
122 const int8x8_t vbCDEFc0 = 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
131 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D4x16c2-minmax-rndnu-neon-mull-dup.c107 … const int8x8_t vbCDEFc0 = 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
150 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
151 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
152 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
153 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
282 … const int8x8_t vbCDEFc0 = 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
291 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
300 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
309 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
318 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup()
D4x16c2-minmax-rndnu-neon-mull-ld4r.c107 … const int8x8_t vbCDEFc0 = 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
150 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
151 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
152 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
153 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
282 … const int8x8_t vbCDEFc0 = 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
291 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
300 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
309 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
318 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r()
D4x16c2-minmax-rndnu-neon-mull-ld2r.c111 … const int8x8_t vbCDEFc0 = 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
154 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
155 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
156 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
157 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
286 … const int8x8_t vbCDEFc0 = 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
295 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
304 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
313 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
322 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r()
D4x16c2-minmax-rndnu-neon-mull-ld1r.c119 … const int8x8_t vbCDEFc0 = 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
162 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
163 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
164 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
165 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
294 … const int8x8_t vbCDEFc0 = 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
303 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
312 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
321 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
330 const int16x8_t vprod3xCDEFc0 = vmull_s8(vbCDEFc0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r()
D3x16c2-minmax-rndnu-neon-mull-ld1r.c103 … const int8x8_t vbCDEFc0 = 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
139 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
140 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
141 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
241 … const int8x8_t vbCDEFc0 = 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
250 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
259 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
268 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r()
D3x16c2-minmax-rndnu-neon-mull-dup.c94 … const int8x8_t vbCDEFc0 = 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
130 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
131 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
132 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
232 … const int8x8_t vbCDEFc0 = 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
241 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
250 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
259 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup()
D3x16c2-minmax-rndnu-neon-mull-ld4r.c94 … const int8x8_t vbCDEFc0 = 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
130 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
131 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
132 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
232 … const int8x8_t vbCDEFc0 = 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
241 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
250 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
259 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r()
D3x16c2-minmax-rndnu-neon-mull-ld2r.c97 … const int8x8_t vbCDEFc0 = 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
133 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
134 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
135 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
235 … const int8x8_t vbCDEFc0 = 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
244 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
253 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
262 const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c87 … const int8x8_t vbCDEFc0 = 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
116 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
117 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
188 … const int8x8_t vbCDEFc0 = 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
197 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
206 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c81 … const int8x8_t vbCDEFc0 = 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
110 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
111 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
182 … const int8x8_t vbCDEFc0 = 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
191 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
200 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c81 … const int8x8_t vbCDEFc0 = 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
110 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
111 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
182 … const int8x8_t vbCDEFc0 = 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
191 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
200 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
D2x16c2-minmax-rndnu-neon-mull-ld2r.c83 … const int8x8_t vbCDEFc0 = 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
112 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
113 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
184 … const int8x8_t vbCDEFc0 = 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
193 const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()
202 const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r()

123