Home
last modified time | relevance | path

Searched refs:vprod0x7 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-neon-mlal-padal.c97 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
98 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
99 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
130 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
131 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c88 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
89 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
90 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c134 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
136 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
138 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
185 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
187 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c113 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
114 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
115 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
178 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
179 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c171 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
174 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
177 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
240 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
243 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c124 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
126 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
128 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c104 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
105 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
106 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c81 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
82 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c160 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
163 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
166 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c208 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
212 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
216 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
295 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
299 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c158 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
160 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
162 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
265 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
267 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c148 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
150 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
152 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c196 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
200 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
204 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-neon-mlal-padal.c108 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
109 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
110 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
141 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
142 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c99 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
100 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
101 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c147 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
149 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
151 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
198 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
200 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c124 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
125 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
126 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
189 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
190 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c137 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
139 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
141 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c115 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
116 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
117 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c186 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
189 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
192 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
255 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
258 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c92 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
93 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c175 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
178 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
181 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c171 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
173 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
175 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
278 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
280 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c225 int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
229 vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
233 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
312 const int16x8_t vprod0x7 = vmull_s8(vb7, va0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
316 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c213 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
217 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
221 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()

12