Home
last modified time | relevance | path

Searched refs:vprod1x89ABc0 (Results 1 – 25 of 60) sorted by relevance

123

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c128 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
131 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
133 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
283 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
285 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
374 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
375 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c116 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
119 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
121 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
265 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
267 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
356 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
357 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c116 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
119 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
121 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
265 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
267 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
356 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
357 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c120 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
123 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
125 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
271 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
273 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
362 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
363 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x16c2s4-minmax-rndnu-neon-mlal.c111 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
114 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
116 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
253 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
255 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
D3x16c2-minmax-rndnu-neon-mlal-ld4r.c138 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
142 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
145 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
341 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
344 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
464 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
465 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
D3x16c2-minmax-rndnu-neon-mlal-dup.c138 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
142 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
145 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
341 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
344 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
464 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
465 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
D3x16c2-minmax-rndnu-neon-mlal-ld1r.c156 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
160 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
163 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
368 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
371 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
491 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
492 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
D3x16c2-minmax-rndnu-neon-mlal-ld2r.c144 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
148 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
151 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
350 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
353 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
473 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
474 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
D3x16c2s4-minmax-rndnu-neon-mlal.c131 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
135 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
138 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
326 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
329 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c113 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
115 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
204 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
205 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
D2x16c2-minmax-rndnu-neon-mull-dup.c107 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
109 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
198 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
199 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld4r.c107 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
109 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
198 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
199 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c2-minmax-rndnu-neon-mlal-ld2r.c106 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
109 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
111 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
257 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
259 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
348 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
349 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r()
D2x16c2-minmax-rndnu-neon-mlal-ld1r.c114 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
117 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
119 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
269 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
271 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
360 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
361 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r()
D2x16c2-minmax-rndnu-neon-mlal-dup.c102 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
105 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
107 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
251 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
253 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
342 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
343 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup()
D2x16c2-minmax-rndnu-neon-mlal-ld4r.c102 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
105 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
107 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
251 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
253 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
342 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
343 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r()
D2x16c2s4-minmax-rndnu-neon-mlal.c98 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
101 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
103 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
240 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
242 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal()
D3x16c2-minmax-rndnu-neon-mlal-dup.c122 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
126 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
129 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
325 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
328 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
448 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
449 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup()
D3x16c2-minmax-rndnu-neon-mlal-ld4r.c122 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
126 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
129 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
325 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
328 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
448 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
449 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r()
D3x16c2-minmax-rndnu-neon-mlal-ld2r.c128 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
132 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
135 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
334 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
337 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
457 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
458 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r()
D3x16c2-minmax-rndnu-neon-mlal-ld1r.c140 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
144 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
147 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
352 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
355 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
475 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
476 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r()
D3x16c2s4-minmax-rndnu-neon-mlal.c116 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
120 vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
123 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
311 int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
314 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal()
D2x16c2-minmax-rndnu-neon-mull-dup.c94 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
96 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
185 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
186 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup()
D2x16c2-minmax-rndnu-neon-mull-ld1r.c100 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
102 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()
191 const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
192 vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r()

123