Home
last modified time | relevance | path

Searched refs:vacc7x4567 (Results 1 – 10 of 10) sorted by relevance

/external/XNNPACK/src/q8-igemm/
D8x8-neon.c86 int32x4_t vacc7x4567 = vacc0x4567; in xnn_q8_igemm_ukernel_8x8__neon() local
162 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa7), 0); in xnn_q8_igemm_ukernel_8x8__neon()
184 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa7), 1); in xnn_q8_igemm_ukernel_8x8__neon()
206 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa7), 2); in xnn_q8_igemm_ukernel_8x8__neon()
228 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa7), 3); in xnn_q8_igemm_ukernel_8x8__neon()
250 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa7), 0); in xnn_q8_igemm_ukernel_8x8__neon()
272 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa7), 1); in xnn_q8_igemm_ukernel_8x8__neon()
294 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa7), 2); in xnn_q8_igemm_ukernel_8x8__neon()
316 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa7), 3); in xnn_q8_igemm_ukernel_8x8__neon()
358 … vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa7), 0); in xnn_q8_igemm_ukernel_8x8__neon()
[all …]
/external/XNNPACK/src/q8-gemm/
D8x8-neon.c97 int32x4_t vacc7x4567 = vacc0x4567; in xnn_q8_gemm_ukernel_8x8__neon() local
136 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa7), 0); in xnn_q8_gemm_ukernel_8x8__neon()
156 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa7), 1); in xnn_q8_gemm_ukernel_8x8__neon()
176 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa7), 2); in xnn_q8_gemm_ukernel_8x8__neon()
196 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa7), 3); in xnn_q8_gemm_ukernel_8x8__neon()
216 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa7), 0); in xnn_q8_gemm_ukernel_8x8__neon()
236 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa7), 1); in xnn_q8_gemm_ukernel_8x8__neon()
256 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa7), 2); in xnn_q8_gemm_ukernel_8x8__neon()
276 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa7), 3); in xnn_q8_gemm_ukernel_8x8__neon()
316 vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa7), 0); in xnn_q8_gemm_ukernel_8x8__neon()
[all …]
/external/XNNPACK/src/f32-gemm/gen-inc/
D8x8s4-neon.c102 float32x4_t vacc7x4567 = vld1q_f32(acc); acc += 4; in xnn_f32_gemminc_ukernel_8x8s4__neon() local
134 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_gemminc_ukernel_8x8s4__neon()
163 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_gemminc_ukernel_8x8s4__neon()
192 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_gemminc_ukernel_8x8s4__neon()
221 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_gemminc_ukernel_8x8s4__neon()
256 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_gemminc_ukernel_8x8s4__neon()
277 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_gemminc_ukernel_8x8s4__neon()
295 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_gemminc_ukernel_8x8s4__neon()
299 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_gemminc_ukernel_8x8s4__neon()
345 vacc7x0123 = vacc7x4567; in xnn_f32_gemminc_ukernel_8x8s4__neon()
D8x8s4-neonfma.c102 float32x4_t vacc7x4567 = vld1q_f32(acc); acc += 4; in xnn_f32_gemminc_ukernel_8x8s4__neonfma() local
134 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
163 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
192 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
221 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
256 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
277 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
295 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
299 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
345 vacc7x0123 = vacc7x4567; in xnn_f32_gemminc_ukernel_8x8s4__neonfma()
/external/XNNPACK/src/f32-igemm/gen/
D8x8s4-neon.c89 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_igemm_ukernel_8x8s4__neon() local
165 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_igemm_ukernel_8x8s4__neon()
194 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_igemm_ukernel_8x8s4__neon()
223 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_igemm_ukernel_8x8s4__neon()
252 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_igemm_ukernel_8x8s4__neon()
287 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_igemm_ukernel_8x8s4__neon()
312 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_igemm_ukernel_8x8s4__neon()
330 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_igemm_ukernel_8x8s4__neon()
334 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_igemm_ukernel_8x8s4__neon()
371 vacc7x0123 = vacc7x4567; in xnn_f32_igemm_ukernel_8x8s4__neon()
D8x8s4-neonfma.c89 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_igemm_ukernel_8x8s4__neonfma() local
165 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
194 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
223 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
252 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
287 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
312 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
330 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
334 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_igemm_ukernel_8x8s4__neonfma()
371 vacc7x0123 = vacc7x4567; in xnn_f32_igemm_ukernel_8x8s4__neonfma()
/external/XNNPACK/src/f32-gemm/gen/
D8x8s4-neonfma.c100 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_gemm_ukernel_8x8s4__neonfma() local
132 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
161 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
190 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
219 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
254 vacc7x4567 = vfmaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
275 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
293 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
297 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_gemm_ukernel_8x8s4__neonfma()
343 vacc7x0123 = vacc7x4567; in xnn_f32_gemm_ukernel_8x8s4__neonfma()
D8x8s4-neon.c100 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_gemm_ukernel_8x8s4__neon() local
132 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c0); in xnn_f32_gemm_ukernel_8x8s4__neon()
161 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c1); in xnn_f32_gemm_ukernel_8x8s4__neon()
190 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c2); in xnn_f32_gemm_ukernel_8x8s4__neon()
219 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567c3); in xnn_f32_gemm_ukernel_8x8s4__neon()
254 vacc7x4567 = vmlaq_f32(vacc7x4567, va7, vb4567); in xnn_f32_gemm_ukernel_8x8s4__neon()
275 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_gemm_ukernel_8x8s4__neon()
293 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_gemm_ukernel_8x8s4__neon()
297 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_gemm_ukernel_8x8s4__neon()
343 vacc7x0123 = vacc7x4567; in xnn_f32_gemm_ukernel_8x8s4__neon()
/external/XNNPACK/src/f32-ppmm/gen/
D8x8-neonfma.c81 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_ppmm_ukernel_8x8__neonfma() local
107 vacc7x4567 = vfmaq_laneq_f32(vacc7x4567, vb4567, va4567, 3); in xnn_f32_ppmm_ukernel_8x8__neonfma()
133 vacc7x4567 = vfmaq_f32(vacc7x4567, va7777, vb4567); in xnn_f32_ppmm_ukernel_8x8__neonfma()
155 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_ppmm_ukernel_8x8__neonfma()
173 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_ppmm_ukernel_8x8__neonfma()
177 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_ppmm_ukernel_8x8__neonfma()
215 vacc7x0123 = vacc7x4567; in xnn_f32_ppmm_ukernel_8x8__neonfma()
D8x8-neon.c81 float32x4_t vacc7x4567 = vacc0x4567; in xnn_f32_ppmm_ukernel_8x8__neon() local
106 vacc7x4567 = vmlaq_lane_f32(vacc7x4567, vb4567, vget_high_f32(va4567), 1); in xnn_f32_ppmm_ukernel_8x8__neon()
127 vacc7x4567 = vminq_f32(vacc7x4567, vmax); in xnn_f32_ppmm_ukernel_8x8__neon()
145 vacc7x4567 = vmaxq_f32(vacc7x4567, vmin); in xnn_f32_ppmm_ukernel_8x8__neon()
149 vst1q_f32(c7 + 4, vacc7x4567); in xnn_f32_ppmm_ukernel_8x8__neon()
187 vacc7x0123 = vacc7x4567; in xnn_f32_ppmm_ukernel_8x8__neon()