Home
last modified time | relevance | path

Searched refs:vk11c0x4567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p1c3x8-neonfma-2x2.c247 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
251 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
252 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
256 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
257 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
651 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
655 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
656 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
937 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
941 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
D3x3s2p1c3x8-neon-2x2.c245 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
249 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
250 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
254 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
255 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
651 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
655 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
656 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
939 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
943 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x2.c247 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
251 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
252 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
256 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
257 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
644 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
648 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
649 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
930 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
934 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
D3x3s2p0p1c3x8-neon-2x2.c245 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
249 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
250 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
254 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk11c0x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
255 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk11c0x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
644 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
648 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
649 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
932 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
936 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
D3x3s2p0p1c3x8-neonfma-2x1.c179 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
183 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
184 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
466 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
470 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
471 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neon-2x1.c184 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
188 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
189 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
473 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
477 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
478 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neon-2x1.c177 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
181 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
182 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
466 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
470 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
471 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk11c0x4567, vget_high_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
D3x3s2p1c3x8-neonfma-2x1.c186 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
190 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
191 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
473 const float32x4_t vk11c0x4567 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
477 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
478 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk11c0x4567, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()