| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p1c3x4-neonfma-2x2.c | 113 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
115 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
116 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
118 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
119 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
404 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
406 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
407 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
595 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
597 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neon-2x2.c | 111 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
113 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
114 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
116 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
117 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
404 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
406 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
407 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
597 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
599 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x2.c | 111 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
113 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
114 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
116 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
117 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
404 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
406 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
407 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
597 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
599 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neonfma-2x2.c | 113 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
115 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
116 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
118 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
119 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
404 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
406 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
407 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
595 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
597 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x1.c | 99 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
101 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
102 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
293 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
295 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
296 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
| D | 3x3s2p0p1c3x4-neonfma-2x1.c | 101 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
103 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
104 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
293 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
295 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
296 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neonfma-2x1.c | 101 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
103 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
104 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
293 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
295 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
296 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neon-2x1.c | 99 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
101 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
102 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
293 const float32x4_t vk10c0x0123 = vld1q_f32(w + 8); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
295 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
296 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 122 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
125 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
126 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
130 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
131 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
571 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
574 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
575 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
857 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
860 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neon-2x2.c | 120 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
123 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
124 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
128 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
129 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
571 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
574 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
575 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
859 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
862 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x2.c | 122 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
125 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
126 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
130 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
131 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
571 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
574 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
575 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
857 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
860 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x2.c | 120 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
123 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
124 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
128 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c0x0123, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
129 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c0x0123, vget_high_f32(vi3x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
571 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
574 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
575 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
859 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
862 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 106 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
109 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
110 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
393 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
396 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
397 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
| D | 3x3s2p1c3x8-neon-2x1.c | 104 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
107 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
108 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
393 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
396 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
397 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 104 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
107 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
108 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
393 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
396 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
397 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x1.c | 106 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
109 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
110 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
393 const float32x4_t vk10c0x0123 = vld1q_f32(w + 16); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
396 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
397 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c0x0123, vget_low_f32(vi3x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-sse-1x1.c | 93 const __m128 vk10c0x0123 = _mm_load_ps(w + 8); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
94 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c0x0123, vi10c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
220 const __m128 vk10c0x0123 = _mm_load_ps(w + 8); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
221 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk10c0x0123, vi10c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
|