Home
last modified time | relevance | path

Searched refs:vacc4x01 (Results 1 – 25 of 50) sorted by relevance

12

/external/XNNPACK/src/f32-igemm/gen/
D6x2-minmax-neonfma-lane-ld64.c71 float32x2_t vacc4x01 = vacc0x01; in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64() local
124 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
137 vacc4x01 = vfma_f32(vacc4x01, va4c0, vb01c0); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
147 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c1, va4, 1); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
160 vacc4x01 = vfma_f32(vacc4x01, va4c1, vb01c1); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
178 vacc4x01 = vfma_f32(vacc4x01, va4, vb01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
189 vacc4x01 = vmin_f32(vacc4x01, vmax); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
197 vacc4x01 = vmax_f32(vacc4x01, vmin); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
203 vst1_f32(c4, vacc4x01); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
219 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x2__neonfma_lane_ld64()
D6x2-minmax-neon-lane-ld64.c71 float32x2_t vacc4x01 = vacc0x01; in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64() local
123 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
131 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c1, va4, 1); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
148 vacc4x01 = vmla_f32(vacc4x01, va4, vb01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
159 vacc4x01 = vmin_f32(vacc4x01, vmax); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
167 vacc4x01 = vmax_f32(vacc4x01, vmin); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
173 vst1_f32(c4, vacc4x01); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
189 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x2__neon_lane_ld64()
D6x8-minmax-neonfma-lane-ld64.c249 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
256 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
263 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
271 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
D6x8-minmax-neon-lane-ld64.c249 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() local
256 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
263 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
271 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
D6x8-minmax-neonfma-dup-ld64.c261 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
268 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
275 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
283 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
D6x8-minmax-neon-dup-ld64.c261 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() local
268 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
275 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
283 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
D6x8-minmax-neon-lane-ld128.c287 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128() local
294 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128()
301 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128()
309 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld128()
/external/XNNPACK/src/f32-gemm/gen/
D6x2-minmax-neonfma-lane-ld64.c77 float32x2_t vacc4x01 = vacc0x01; in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64() local
96 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
109 vacc4x01 = vfma_f32(vacc4x01, va4c0, vb01c0); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
119 vacc4x01 = vfma_lane_f32(vacc4x01, vb01c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
132 vacc4x01 = vfma_f32(vacc4x01, va4c1, vb01c1); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
150 vacc4x01 = vfma_f32(vacc4x01, va4, vb01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
159 vacc4x01 = vmin_f32(vacc4x01, vmax); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
167 vacc4x01 = vmax_f32(vacc4x01, vmin); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
179 vst1_f32(c4, vacc4x01); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
198 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64()
D6x2-minmax-neon-lane-ld64.c77 float32x2_t vacc4x01 = vacc0x01; in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64() local
95 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
103 vacc4x01 = vmla_lane_f32(vacc4x01, vb01c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
120 vacc4x01 = vmla_f32(vacc4x01, va4, vb01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
129 vacc4x01 = vmin_f32(vacc4x01, vmax); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
137 vacc4x01 = vmax_f32(vacc4x01, vmin); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
149 vst1_f32(c4, vacc4x01); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
168 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64()
D5x8-minmax-neon-lane-ld64.c197 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local
203 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
209 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
216 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
D5x8-minmax-neonfma-lane-ld64.c197 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local
203 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
209 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
216 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
D6x8-minmax-neonfma-lane-ld64.c224 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
231 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
238 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
246 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
D6x8-minmax-neon-lane-ld64.c224 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() local
231 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
238 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
246 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
D6x8-minmax-neonfma-dup-ld64.c236 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
243 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
250 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
258 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
D6x8-minmax-neon-dup-ld64.c236 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() local
243 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
250 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
258 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
D6x8-minmax-neon-lane-ld128.c262 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128() local
269 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128()
276 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128()
284 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128()
/external/XNNPACK/src/f32-gemm/gen-inc/
D5x8inc-minmax-neonfma-lane-ld64.c199 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local
205 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
211 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
218 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
D5x8inc-minmax-neon-lane-ld64.c199 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local
205 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
211 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
218 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
D6x8inc-minmax-neonfma-lane-ld64.c226 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local
233 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
240 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
248 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
D6x8inc-minmax-neon-lane-ld64.c226 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() local
233 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
240 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
248 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
D6x8inc-minmax-neon-dup-ld64.c238 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() local
245 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
252 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
260 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
D6x8inc-minmax-neonfma-dup-ld64.c238 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() local
245 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
252 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
260 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
D6x8inc-minmax-neonfma-lane-ld128.c264 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld128() local
271 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld128()
278 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld128()
286 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld128()
/external/XNNPACK/src/f32-ppmm/gen/
D8x8-minmax-neon.c199 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() local
208 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neon()
217 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neon()
227 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_ppmm_minmax_ukernel_8x8__neon()
D8x8-minmax-neonfma.c227 float32x2_t vacc4x01 = vget_low_f32(vacc4x0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() local
236 vst1_f32(c4, vacc4x01); c4 += 2; in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()
245 vacc4x01 = vget_high_f32(vacc4x0123); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()
255 vst1_lane_f32(c4, vacc4x01, 0); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma()

12