/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x8-neon-2x2.c | 408 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 412 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 413 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 417 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 418 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 747 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 751 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 752 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 410 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 414 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 415 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 419 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 420 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 754 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 758 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 759 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c1x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 410 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 414 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 415 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 419 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 420 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 747 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 751 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 752 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 408 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 412 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c1x4567, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 413 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c1x4567, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 417 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c1x4567, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 418 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c1x4567, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 754 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 758 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 759 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neon-2x1.c | 287 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 291 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 292 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 282 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 286 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 287 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 289 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 293 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c1x4567, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 294 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c1x4567, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 280 const float32x4_t vk12c1x4567 = vld1q_f32(w + 188); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 284 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c1x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 285 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c1x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|