/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p1c3x8-neonfma-2x2.c | 117 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 119 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 120 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 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() 132 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 133 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 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() [all …]
|
D | 3x3s2p1c3x8-neon-2x2.c | 115 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 117 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 118 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 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() 130 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 131 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 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() [all …]
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 117 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 118 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 119 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 120 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 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() 132 vo0x1c4567 = vfmaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 133 vo1x1c4567 = vfmaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 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() [all …]
|
D | 3x3s2p0p1c3x8-neon-2x2.c | 115 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 116 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 117 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk00c0x4567, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 118 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk00c0x4567, vget_high_f32(vi2x1), 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() 130 vo0x1c4567 = vmlaq_lane_f32(vo0x1c4567, vk10c0x4567, vget_high_f32(vi1x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 131 vo1x1c4567 = vmlaq_lane_f32(vo1x1c4567, vk10c0x4567, vget_high_f32(vi3x1), 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() [all …]
|
D | 3x3s2p1c3x4-neonfma-2x2.c | 110 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 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() 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() 138 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 139 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 146 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 147 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() [all …]
|
D | 3x3s2p1c3x4-neon-2x2.c | 108 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 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() 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() 136 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 137 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 144 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 145 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() [all …]
|
D | 3x3s2p0p1c3x4-neon-2x2.c | 108 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 109 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 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() 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() 132 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 133 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 140 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 141 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() [all …]
|
D | 3x3s2p0p1c3x4-neonfma-2x2.c | 110 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c0x0123, vget_high_f32(vi0x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 111 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c0x0123, vget_high_f32(vi2x1), 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() 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() 134 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk00c1x0123, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 135 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk00c1x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 142 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk10c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 143 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk10c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() [all …]
|
D | 3x3s2p1c3x8-neon-2x1.c | 123 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 124 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 125 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 126 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 131 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 132 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 133 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 134 vo1c4567 = vmlaq_lane_f32(vo1c4567, vk10c1x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 139 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 140 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() [all …]
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 125 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk00c1x0123, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 126 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk00c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 127 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk00c1x4567, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 128 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk00c1x4567, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 133 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 134 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk10c1x0123, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 135 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 136 vo1c4567 = vfmaq_lane_f32(vo1c4567, vk10c1x4567, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 141 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk20c1x0123, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 142 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk20c1x0123, vget_high_f32(vi4x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() [all …]
|
/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-neon-4x4-acc2.c | 116 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 117 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 118 float32x4_t vo2p1 = vmulq_lane_f32(vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 119 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 131 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 132 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 133 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 134 vo3p1 = vmlaq_lane_f32(vo3p1, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 136 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() 137 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() [all …]
|
D | 5x5p2-minmax-neon-4x4.c | 116 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 117 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 118 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 119 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 131 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 132 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 133 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 134 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 136 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() 137 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() [all …]
|
D | 5x5p2-minmax-neonfma-4x4.c | 116 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 117 vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 118 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 119 vo3p0 = vfmaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 131 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 132 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 133 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 134 vo3p0 = vfmaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 136 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() 137 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() [all …]
|
D | 5x5p2-minmax-neonfma-4x4-acc2.c | 116 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 117 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 118 float32x4_t vo2p1 = vmulq_lane_f32(vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 119 float32x4_t vo3p1 = vmulq_lane_f32(vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 131 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 132 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 133 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 134 vo3p1 = vfmaq_lane_f32(vo3p1, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 136 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() 137 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() [all …]
|
D | 5x5p2-minmax-neonfma-5x4.c | 126 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 127 vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 128 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 129 vo3p0 = vfmaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 130 vo4p0 = vfmaq_lane_f32(vo4p0, vi4x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 144 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 145 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 146 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 147 vo3p0 = vfmaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() 148 vo4p0 = vfmaq_lane_f32(vo4p0, vi7x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() [all …]
|
D | 5x5p2-minmax-neon-5x4.c | 126 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 127 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 128 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 129 vo3p0 = vmlaq_lane_f32(vo3p0, vi3x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 130 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 144 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 145 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 146 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 147 vo3p0 = vmlaq_lane_f32(vo3p0, vi6x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() 148 vo4p0 = vmlaq_lane_f32(vo4p0, vi7x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() [all …]
|
D | 5x5p2-minmax-neon-3x4.c | 106 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 107 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 108 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 118 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 119 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 120 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 122 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 123 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 124 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() 134 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() [all …]
|
D | 5x5p2-minmax-neonfma-3x4-acc2.c | 106 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 107 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 108 float32x4_t vo2p1 = vmulq_lane_f32(vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 118 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 119 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 120 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 122 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 123 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 124 vo2p0 = vfmaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() 134 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() [all …]
|
D | 5x5p2-minmax-neonfma-3x4.c | 106 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 107 vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 108 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 118 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 119 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 120 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 122 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 123 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 124 vo2p0 = vfmaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() 134 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() [all …]
|
D | 5x5p2-minmax-neon-3x4-acc2.c | 106 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 107 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 108 float32x4_t vo2p1 = vmulq_lane_f32(vi2x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 118 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 119 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 120 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 122 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 123 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 124 vo2p0 = vmlaq_lane_f32(vo2p0, vi6x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() 134 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() [all …]
|
D | 5x5p2-minmax-neonfma-2x4-acc2.c | 96 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 105 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 106 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 108 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 109 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 118 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 119 vo1p1 = vfmaq_lane_f32(vo1p1, vi1x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 121 vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() 122 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() [all …]
|
D | 5x5p2-minmax-neon-2x4.c | 96 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 97 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 105 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 106 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 108 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 109 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 118 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 119 vo1p0 = vmlaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 121 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() 122 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() [all …]
|
D | 5x5p2-minmax-neonfma-2x4.c | 96 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 97 vo1p0 = vfmaq_lane_f32(vo1p0, vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 105 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 106 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 108 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 109 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 118 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 119 vo1p0 = vfmaq_lane_f32(vo1p0, vi1x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 121 vo0p0 = vfmaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() 122 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() [all …]
|
D | 5x5p2-minmax-neon-2x4-acc2.c | 96 float32x4_t vo0p1 = vmulq_lane_f32(vi0x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 97 float32x4_t vo1p1 = vmulq_lane_f32(vi1x4567, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 105 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 106 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x4567, vget_high_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 108 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 109 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x4567, vget_high_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 118 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 119 vo1p1 = vmlaq_lane_f32(vo1p1, vi1x3456, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 121 vo0p0 = vmlaq_lane_f32(vo0p0, vi1x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() 122 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() [all …]
|
/external/XNNPACK/src/f32-conv-hwc2chw/ |
D | 3x3s2p1c3x4-neon-2x2.c | 120 vo0x1 = vmlaq_lane_f32(vo0x1, vk00c0, vget_high_f32(vi0x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 121 vo1x1 = vmlaq_lane_f32(vo1x1, vk00c0, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 127 vo0x1 = vmlaq_lane_f32(vo0x1, vk10c0, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 128 vo1x1 = vmlaq_lane_f32(vo1x1, vk10c0, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 134 vo0x1 = vmlaq_lane_f32(vo0x1, vk20c0, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 135 vo1x1 = vmlaq_lane_f32(vo1x1, vk20c0, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 146 vo0x0 = vmlaq_lane_f32(vo0x0, vk00c1, vget_high_f32(vi0x0), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 147 vo1x0 = vmlaq_lane_f32(vo1x0, vk00c1, vget_high_f32(vi2x0), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 153 vo0x0 = vmlaq_lane_f32(vo0x0, vk10c1, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() 154 vo1x0 = vmlaq_lane_f32(vo1x0, vk10c1, vget_high_f32(vi3x0), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() [all …]
|