/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p1c3x8-neonfma-2x2.c | 371 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 375 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 376 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 380 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 381 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 723 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 727 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 728 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 369 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 373 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 374 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 378 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 379 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 723 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 727 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 728 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 371 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 375 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 376 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 380 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 381 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 723 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 727 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 728 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neon-2x2.c | 369 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 373 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 374 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 378 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk12c0x4567, vget_low_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 379 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk12c0x4567, vget_low_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 723 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 727 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 728 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 258 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 262 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 263 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neon-2x1.c | 256 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 260 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 261 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 256 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 260 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 261 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 258 const float32x4_t vk12c0x4567 = vld1q_f32(w + 164); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 262 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk12c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 263 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk12c0x4567, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|