Home
last modified time | relevance | path

Searched refs:vb89c0x1 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D3x16c4s2-minmax-rndnu-neon-mlal.c150 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
151 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
152 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
153 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D1x16c4s2-minmax-rndnu-neon-mlal.c92 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
93 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal()
D4x16c4s2-minmax-rndnu-neon-mlal.c179 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
180 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
181 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
182 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
183 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D2x16c4s2-minmax-rndnu-neon-mlal.c121 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
122 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
123 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-dup.c189 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
190 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
191 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
192 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
193 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c197 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
198 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
199 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
200 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
201 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c189 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
190 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
191 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
192 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
193 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c158 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
159 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
160 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
161 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c158 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
159 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
160 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
161 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c164 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
165 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
166 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
167 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c131 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
132 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
133 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
D2x16c4-minmax-rndnu-neon-mlal-dup.c127 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
128 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
129 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld2r.c127 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
128 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
129 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D3x16c4s2-minmax-rndnu-neon-mlal.c165 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
166 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
167 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
168 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal()
D4x16c4s2-minmax-rndnu-neon-mlal.c196 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
197 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
198 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
199 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
200 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal()
D2x16c4s2-minmax-rndnu-neon-mlal.c134 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
135 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
136 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal()
D1x16c4s2-minmax-rndnu-neon-mlal.c103 const int8x8_t vb89c0x1 = vld1_s8(w); w = (const int8_t*) w + 8; in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
104 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal()
D4x16c4-minmax-rndnu-neon-mlal-ld1r.c214 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
215 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
216 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
217 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
218 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r()
D4x16c4-minmax-rndnu-neon-mlal-dup.c206 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
207 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
208 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
209 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
210 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup()
D4x16c4-minmax-rndnu-neon-mlal-ld2r.c206 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
207 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
208 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
209 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
210 vprod3x89c0 = vmlal_s8(vprod3x89c0, vb89c0x1, va3c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c179 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
180 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
181 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
182 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c173 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
174 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
175 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
176 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c173 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
174 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
175 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
176 vprod2x89c0 = vmlal_s8(vprod2x89c0, vb89c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-dup.c140 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
141 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
142 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup()
D2x16c4-minmax-rndnu-neon-mlal-ld1r.c144 … const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
145 vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()
146 vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r()

12