Home
last modified time | relevance | path

Searched refs:vprod2x23c0 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c107 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
111 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
114 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
212 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
215 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
291 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
292 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c113 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
117 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
120 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
221 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
224 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
300 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
301 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-dup.c107 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
111 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
114 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
212 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
215 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
291 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
292 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4s2-minmax-rndnu-neon-mlal.c99 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
103 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
106 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
198 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
201 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c124 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
129 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
133 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
256 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
351 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
352 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c124 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
129 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
133 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
256 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
351 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
352 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c132 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
137 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
141 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
268 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
272 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
363 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
364 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4s2-minmax-rndnu-neon-mlal.c114 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
119 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
123 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
239 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D3x8c4-minmax-rndnu-neon-mull-ld1r.c101 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
104 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
180 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
181 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c98 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
101 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
177 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
178 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c98 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
101 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
177 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
178 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c4-minmax-rndnu-neon-mlal-dup.c122 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
126 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
227 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
230 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
306 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
307 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c128 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
132 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
135 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
236 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
239 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
315 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
316 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c122 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
126 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
227 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
230 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
306 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
307 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4s2-minmax-rndnu-neon-mlal.c114 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
118 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
121 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
213 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
216 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c149 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
154 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
158 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
285 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
289 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
380 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
381 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
D4x8c4-minmax-rndnu-neon-mlal-dup.c141 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
146 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
150 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
273 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
277 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
368 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
369 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c141 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
146 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
150 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
273 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
277 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
368 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
369 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
D4x8c4s2-minmax-rndnu-neon-mlal.c131 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
136 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
140 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
256 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D3x8c4-minmax-rndnu-neon-mull-ld1r.c116 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
119 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
195 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
196 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c113 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
192 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
193 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c113 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
192 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
193 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-ld1r.c148 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
152 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
155 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
344 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
347 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
491 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
492 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r()
D3x16c4-minmax-rndnu-neon-mlal-ld2r.c142 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
146 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
149 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
335 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
338 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
482 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
483 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r()
D3x16c4-minmax-rndnu-neon-mlal-dup.c142 int16x8_t vprod2x23c0 = vmull_s8(vb23c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
146 vprod2x23c0 = vmlal_s8(vprod2x23c0, vb23c0x1, va2c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
149 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
335 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
338 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()
482 const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
483 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup()

123