Home
last modified time | relevance | path

Searched refs:vacc3x01 (Results 1 – 25 of 80) sorted by relevance

1234

/external/XNNPACK/src/f32-gemm/gen/
D4x2-neonfma-lane-ld64.c64 float32x2_t vacc3x01 = vacc0x01; in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64() local
79 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c0, va3, 0); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
88 vacc3x01 = vfma_f32(vacc3x01, va3c0, vb01c0); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
96 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c1, va3, 1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
105 vacc3x01 = vfma_f32(vacc3x01, va3c1, vb01c1); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
119 vacc3x01 = vfma_f32(vacc3x01, va3, vb01); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
126 vacc3x01 = vmin_f32(vacc3x01, vmax); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
132 vacc3x01 = vmax_f32(vacc3x01, vmin); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
141 vst1_f32(c3, vacc3x01); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
155 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64()
D4x2-neon-lane-ld64.c64 float32x2_t vacc3x01 = vacc0x01; in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64() local
78 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c0, va3, 0); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
84 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c1, va3, 1); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
97 vacc3x01 = vmla_f32(vacc3x01, va3, vb01); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
104 vacc3x01 = vmin_f32(vacc3x01, vmax); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
110 vacc3x01 = vmax_f32(vacc3x01, vmin); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
119 vst1_f32(c3, vacc3x01); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
133 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x2__neon_lane_ld64()
D4x8-neon-lane-ld64.c171 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld64() local
176 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neon_lane_ld64()
181 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld64()
187 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld64()
D4x8-neonfma-lane-ld64.c171 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld64() local
176 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld64()
181 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld64()
187 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld64()
D4x8-neon-dup-ld64.c179 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64() local
184 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
189 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
195 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neon_dup_ld64()
D4x8-neonfma-dup-ld64.c179 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64() local
184 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
189 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
195 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64()
D5x8-neon-lane-ld64.c198 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_5x8__neon_lane_ld64() local
204 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_5x8__neon_lane_ld64()
210 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_5x8__neon_lane_ld64()
217 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_5x8__neon_lane_ld64()
D4x8-neon-lane-ld128.c201 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld128() local
206 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neon_lane_ld128()
211 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld128()
217 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neon_lane_ld128()
D4x8-neonfma-lane-ld128.c201 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128() local
206 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
211 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
217 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128()
/external/XNNPACK/src/f32-igemm/gen/
D4x2-neonfma-lane-ld64.c62 float32x2_t vacc3x01 = vacc0x01; in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64() local
101 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c0, va3, 0); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
110 vacc3x01 = vfma_f32(vacc3x01, va3c0, vb01c0); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
118 vacc3x01 = vfma_lane_f32(vacc3x01, vb01c1, va3, 1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
127 vacc3x01 = vfma_f32(vacc3x01, va3c1, vb01c1); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
141 vacc3x01 = vfma_f32(vacc3x01, va3, vb01); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
150 vacc3x01 = vmin_f32(vacc3x01, vmax); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
156 vacc3x01 = vmax_f32(vacc3x01, vmin); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
159 vst1_f32(c3, vacc3x01); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
172 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64()
D4x2-neon-lane-ld64.c62 float32x2_t vacc3x01 = vacc0x01; in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64() local
100 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c0, va3, 0); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
106 vacc3x01 = vmla_lane_f32(vacc3x01, vb01c1, va3, 1); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
119 vacc3x01 = vmla_f32(vacc3x01, va3, vb01); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
128 vacc3x01 = vmin_f32(vacc3x01, vmax); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
134 vacc3x01 = vmax_f32(vacc3x01, vmin); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
137 vst1_f32(c3, vacc3x01); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
150 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x2__neon_lane_ld64()
D4x4-neon-lane-ld64.c150 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x4__neon_lane_ld64() local
155 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x4__neon_lane_ld64()
160 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x4__neon_lane_ld64()
166 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x4__neon_lane_ld64()
D4x4-neonfma-lane-ld64.c150 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x4__neonfma_lane_ld64() local
155 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x4__neonfma_lane_ld64()
160 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x4__neonfma_lane_ld64()
166 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x4__neonfma_lane_ld64()
D4x8-neon-dup-ld64.c200 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64() local
205 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
210 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
216 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x8__neon_dup_ld64()
D4x8-neonfma-dup-ld64.c200 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64() local
205 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
210 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
216 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64()
D4x8-neon-lane-ld64.c192 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neon_lane_ld64() local
197 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x8__neon_lane_ld64()
202 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neon_lane_ld64()
208 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x8__neon_lane_ld64()
D4x8-neonfma-lane-ld64.c192 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld64() local
197 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld64()
202 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld64()
208 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld64()
/external/XNNPACK/src/f32-ppmm/gen/
D4x8-neon.c127 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_4x8__neon() local
132 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_ppmm_ukernel_4x8__neon()
137 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_4x8__neon()
143 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_ppmm_ukernel_4x8__neon()
D4x8-neonfma.c143 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_4x8__neonfma() local
148 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_ppmm_ukernel_4x8__neonfma()
153 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_4x8__neonfma()
159 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_ppmm_ukernel_4x8__neonfma()
D8x8-neon.c200 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_8x8__neon() local
209 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_ppmm_ukernel_8x8__neon()
218 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_ppmm_ukernel_8x8__neon()
228 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_ppmm_ukernel_8x8__neon()
/external/XNNPACK/src/f32-gemm/gen-inc/
D4x8-neonfma-lane-ld64.c173 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64() local
178 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64()
183 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64()
189 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64()
D4x8-neon-lane-ld64.c173 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64() local
178 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64()
183 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64()
189 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64()
D4x8-neon-dup-ld64.c181 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64() local
186 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
191 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
197 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64()
D4x8-neonfma-dup-ld64.c181 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64() local
186 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
191 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
197 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64()
D5x8-neonfma-lane-ld64.c200 float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64() local
206 vst1_f32(c3, vacc3x01); c3 += 2; in xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64()
212 vacc3x01 = vget_high_f32(vacc3x0123); in xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64()
219 vst1_lane_f32(c3, vacc3x01, 0); in xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64()

1234