Home
last modified time | relevance | path

Searched refs:vprod2x0 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-rndnu-neon-mlal.c103 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
106 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
109 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
193 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
196 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c119 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
123 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
127 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
234 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
238 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c99 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
102 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
105 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c16-minmax-rndnu-neon-mlal.c114 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
118 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
122 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D3x16c8-minmax-rndnu-neon-mlal.c135 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
138 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
141 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
305 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
308 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D3x16c16-minmax-rndnu-neon-mlal.c131 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
134 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
137 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c92 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
95 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D4x16c8-minmax-rndnu-neon-mlal.c159 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
163 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
167 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
378 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
382 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
D4x16c16-minmax-rndnu-neon-mlal.c154 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
158 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
162 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c107 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
111 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mull.c116 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
119 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D4x16c8-minmax-rndnu-neon-mull.c139 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
143 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-rndnu-neon-mlal.c118 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
121 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
124 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
208 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal() local
211 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mlal()
D4x8c8-minmax-rndnu-neon-mlal.c136 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
140 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
144 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
251 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal() local
255 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mlal()
D3x8c16-minmax-rndnu-neon-mlal.c114 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal() local
117 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
120 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c16__neon_mlal()
D4x8c16-minmax-rndnu-neon-mlal.c131 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal() local
135 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
139 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c16__neon_mlal()
D3x16c8-minmax-rndnu-neon-mlal.c150 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
153 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
156 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
320 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal() local
323 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mlal()
D3x16c16-minmax-rndnu-neon-mlal.c146 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal() local
149 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
152 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal()
D4x16c8-minmax-rndnu-neon-mlal.c176 int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
180 vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
184 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
395 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal() local
399 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c107 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
110 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D4x16c16-minmax-rndnu-neon-mlal.c171 int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal() local
175 vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
179 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal()
D4x8c8-minmax-rndnu-neon-mull.c124 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull() local
128 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c8__neon_mull()
D3x16c8-minmax-rndnu-neon-mull.c131 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull() local
134 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c8__neon_mull()
D4x16c8-minmax-rndnu-neon-mull.c156 const int16x8_t vprod2x0 = vmull_s8(vb0, va2); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull() local
160 vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull()