| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p0p1c3x4-neonfma-2x2.c | 121 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
123 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
124 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
126 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
127 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
409 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
411 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
412 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
600 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
602 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neon-2x2.c | 119 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
121 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
122 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
124 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
125 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
409 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
411 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
412 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
602 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
604 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x2.c | 119 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
121 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
122 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
124 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
125 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
409 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
411 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
412 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
602 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
604 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
[all …]
|
| D | 3x3s2p1c3x4-neonfma-2x2.c | 121 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
123 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
124 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
126 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
127 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
409 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
411 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
412 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
600 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
602 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x4-neon-2x1.c | 104 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
106 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
107 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
298 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
300 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
301 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
| D | 3x3s2p1c3x4-neonfma-2x1.c | 106 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
108 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
109 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
298 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
300 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
301 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
| D | 3x3s2p1c3x4-neon-2x1.c | 104 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
106 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
107 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
298 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
300 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
301 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
| D | 3x3s2p0p1c3x4-neonfma-2x1.c | 106 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
108 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
109 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
298 const float32x4_t vk20c0x0123 = vld1q_f32(w + 12); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
300 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
301 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
| D | 3x3s2p0p1c3x8-neon-2x2.c | 133 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
141 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
142 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
579 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
582 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
583 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
867 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
870 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neonfma-2x2.c | 135 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
143 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
144 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
579 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
582 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
583 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
865 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
868 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p0p1c3x8-neonfma-2x2.c | 135 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
143 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
144 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
579 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
582 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
583 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
865 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
868 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neon-2x2.c | 133 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
141 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk20c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
142 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk20c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
579 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
582 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
583 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
867 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
870 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
[all …]
|
| D | 3x3s2p1c3x8-neon-2x1.c | 112 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
115 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
116 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
401 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
404 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
405 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 114 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
117 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
118 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
401 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
404 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
405 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
| D | 3x3s2p1c3x8-neonfma-2x1.c | 114 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
117 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
118 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
401 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
404 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
405 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 112 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
115 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
116 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
401 const float32x4_t vk20c0x0123 = vld1q_f32(w + 24); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
404 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c0x0123, vget_low_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
405 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c0x0123, vget_low_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
| /external/XNNPACK/src/f32-conv-hwc2chw/ |
| D | 3x3s2p1c3x4-sse-1x1.c | 96 const __m128 vk20c0x0123 = _mm_load_ps(w + 12); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
97 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c0x0123, vi20c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
223 const __m128 vk20c0x0123 = _mm_load_ps(w + 12); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
224 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk20c0x0123, vi20c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
|