/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x8-neon-2x2.c | 349 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 360 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 361 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 365 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 366 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 715 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 719 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 720 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 358 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 362 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 363 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 367 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 368 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 715 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 719 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 720 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 351 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 362 vo0x0c4567 = vfmaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 363 vo1x0c4567 = vfmaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 367 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 368 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 715 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 719 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 720 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 356 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 360 vo0x0c4567 = vmlaq_lane_f32(vo0x0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 361 vo1x0c4567 = vmlaq_lane_f32(vo1x0c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 365 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk02c0x4567, vget_low_f32(vi0x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 366 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk02c0x4567, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 715 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 719 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 720 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p1c3x8-neon-2x1.c | 248 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 252 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 253 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 250 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 254 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 255 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 250 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 254 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 255 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 248 const float32x4_t vk02c0x4567 = vld1q_f32(w + 156); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 252 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk02c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 253 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|