Home
last modified time | relevance | path

Searched refs:vk22c1x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x4-neonfma-2x2.c303 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
305 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
306 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
308 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
309 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
528 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
530 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
531 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c301 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
303 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
304 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
306 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
307 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
528 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
530 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
531 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p0p1c3x4-neon-2x2.c301 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
303 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
304 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
306 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
307 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
521 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
523 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
524 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
D3x3s2p0p1c3x4-neonfma-2x2.c303 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
305 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
306 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
308 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
309 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
521 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
523 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
524 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
D3x3s2p0p1c3x4-neon-2x1.c216 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
218 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
219 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c218 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
220 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
221 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neonfma-2x1.c225 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
227 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
228 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c223 const float32x4_t vk22c1x0123 = vld1q_f32(w + 96); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
225 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
226 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p1c3x8-neonfma-2x2.c422 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
425 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
426 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
430 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
431 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
761 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
764 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
765 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
D3x3s2p1c3x8-neon-2x2.c420 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
423 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
424 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
428 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
429 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
761 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
764 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
765 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
D3x3s2p0p1c3x8-neonfma-2x2.c422 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
425 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
426 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
430 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
431 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
754 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
757 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
758 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
D3x3s2p0p1c3x8-neon-2x2.c420 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
423 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
424 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
428 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c1x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
429 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c1x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
754 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
757 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
758 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
D3x3s2p0p1c3x8-neonfma-2x1.c289 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
292 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
293 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c294 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
297 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
298 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c287 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
290 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
291 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c1x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c296 const float32x4_t vk22c1x0123 = vld1q_f32(w + 192); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
299 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c1x0123, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
300 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c1x0123, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c177 const __m128 vk22c1x0123 = _mm_load_ps(w + 96); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
179 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk22c1x0123, vi22c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()