/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x8-neon-2x2.c | 447 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 451 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 452 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 456 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 457 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 778 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 782 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 783 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 449 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 453 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 454 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 458 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 459 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 778 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 782 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 783 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 449 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 453 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 454 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 458 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 459 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 778 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 782 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 783 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 447 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 451 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c2x4567, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 452 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c2x4567, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 456 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c2x4567, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 457 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c2x4567, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 778 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 782 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 783 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neon-2x1.c | 311 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 315 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 316 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 313 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 317 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 318 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 313 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 317 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 318 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 311 const float32x4_t vk12c2x4567 = vld1q_f32(w + 212); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 315 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c2x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 316 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c2x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|