Home
last modified time | relevance | path

Searched refs:vk02c2x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c311 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
313 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
314 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
316 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
317 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
526 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
535 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
536 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c309 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
311 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
312 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
314 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
315 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
533 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
535 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
536 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p0p1c3x4-neon-2x2.c309 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
311 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
312 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
314 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
315 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
526 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
535 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
536 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
D3x3s2p1c3x4-neonfma-2x2.c311 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
313 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
314 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
316 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
317 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
533 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
535 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
536 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p0p1c3x4-neon-2x1.c221 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
230 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
231 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p1c3x4-neonfma-2x1.c230 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
232 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
233 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c228 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
230 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
231 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c223 const float32x4_t vk02c2x0123 = vld1q_f32(w + 100); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
232 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
233 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x2.c433 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
436 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
437 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
441 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
442 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
762 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
772 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
773 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
D3x3s2p1c3x8-neonfma-2x2.c435 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
438 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
439 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
443 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
444 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
769 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
772 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
773 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
D3x3s2p0p1c3x8-neonfma-2x2.c435 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
438 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
439 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
443 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
444 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
762 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
772 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
773 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
D3x3s2p1c3x8-neon-2x2.c433 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
436 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk02c2x0123, vget_low_f32(vi0x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
437 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk02c2x0123, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
441 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk02c2x0123, vget_high_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
442 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk02c2x0123, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
769 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
772 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
773 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
D3x3s2p1c3x8-neon-2x1.c302 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
305 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
306 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c297 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
307 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
308 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c304 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
307 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
308 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c295 const float32x4_t vk02c2x0123 = vld1q_f32(w + 200); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
305 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk02c2x0123, vi0x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
306 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk02c2x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c185 const __m128 vk02c2x0123 = _mm_load_ps(w + 100); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
187 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk02c2x0123, vi02c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()