/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x8-neon-2x2.c | 382 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 386 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 387 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 391 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 392 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 731 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 735 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 736 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 384 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 388 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 389 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 393 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 394 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 731 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 735 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 736 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 384 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 388 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 389 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 393 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 394 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 731 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 735 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 736 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 382 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 386 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 387 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 391 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk22c0x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 392 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk22c0x4567, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 731 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 735 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 736 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neon-2x1.c | 264 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 268 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 269 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 266 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 270 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 271 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 266 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 270 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 271 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 264 const float32x4_t vk22c0x4567 = vld1q_f32(w + 172); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 268 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 269 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk22c0x4567, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|