Home
last modified time | relevance | path

Searched refs:vprod0x6 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-neon-mlal-padal.c93 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
94 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
95 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
127 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
128 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c85 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
86 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
87 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c127 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
129 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
131 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
180 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
182 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c109 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
110 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
111 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
175 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
176 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c161 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
164 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
167 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
233 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
236 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c118 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
120 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
122 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c101 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
102 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
103 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c78 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
79 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c151 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
154 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
157 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c195 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
199 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
203 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
286 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
290 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c151 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
153 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
155 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
260 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
262 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c142 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
144 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
146 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c184 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
188 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
192 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-neon-mlal-padal.c104 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
105 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
106 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
138 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
139 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c96 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
97 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
98 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c140 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
142 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
144 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
193 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
195 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c120 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
121 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
122 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
186 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
187 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c131 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
133 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
135 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c112 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
113 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
114 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c176 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
179 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
182 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
248 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
251 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c89 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
90 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c166 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
169 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
172 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c164 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
166 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
168 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
273 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
275 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c212 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
216 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
220 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
303 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
307 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c201 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
205 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
209 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()

12