/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p1c3x8-neonfma-2x2.c | 462 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 466 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 467 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 471 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 472 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 786 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 790 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 791 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 460 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 464 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 465 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 469 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 470 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 786 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 790 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 791 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 462 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 466 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 467 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 471 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 472 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 786 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 790 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 791 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neon-2x2.c | 460 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 464 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c2x4567, vget_low_f32(vi2x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 465 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c2x4567, vget_low_f32(vi4x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 469 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c2x4567, vget_high_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 470 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c2x4567, vget_high_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 786 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 790 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 791 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 321 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 325 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 326 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neon-2x1.c | 319 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 323 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 324 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 319 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 323 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 324 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 321 const float32x4_t vk22c2x4567 = vld1q_f32(w + 220); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 325 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c2x4567, vi2x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 326 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c2x4567, vi4x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|