/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3s2p1-minmax-neonfma-4x4.c | 217 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 218 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 219 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 220 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 221 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 222 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 223 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 224 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 225 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() 226 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() [all …]
|
D | 3x3s2p1-minmax-neon-4x4.c | 217 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 218 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 219 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 220 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 221 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 222 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 223 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 224 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 225 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() 226 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() [all …]
|
D | 3x3s2p1-minmax-neon-3x4.c | 183 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 184 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 185 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 186 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 187 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 188 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 189 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 190 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 191 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() 192 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() [all …]
|
D | 3x3s2p1-minmax-neonfma-3x4.c | 183 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 184 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 185 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 186 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 187 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 188 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 189 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 190 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 191 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() 192 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() [all …]
|
D | 3x3s2p1-minmax-neonfma-2x4.c | 149 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 150 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 151 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 152 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 153 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 154 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 155 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 156 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 157 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() 158 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
|
D | 3x3s2p1-minmax-neon-2x4-acc2.c | 151 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 152 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 153 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 154 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 155 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 156 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 157 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 158 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 159 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2() 160 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4_acc2()
|
D | 3x3s2p1-minmax-neon-2x4.c | 149 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 150 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 151 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 152 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 153 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 154 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 155 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 156 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 157 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() 158 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
|
D | 3x3s2p1-minmax-neonfma-2x4-acc2.c | 151 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 152 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 153 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 154 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 155 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 156 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 157 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 158 …loat32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 159 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() 160 …loat32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
|
D | 3x3s2p1-minmax-neon-1x4-acc3.c | 116 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() 117 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() 118 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() 119 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() 120 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() 121 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
|
D | 3x3s2p1-minmax-neon-1x4-acc2.c | 115 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() 116 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() 117 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() 118 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() 119 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() 120 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
|
D | 3x3s2p1-minmax-neonfma-1x4-acc2.c | 115 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() 116 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() 117 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() 118 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() 119 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() 120 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
|
D | 3x3s2p1-minmax-neon-1x4.c | 114 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() 115 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() 116 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() 117 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() 118 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() 119 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
|
D | 3x3s2p1-minmax-neon-1x4-acc4.c | 117 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() 118 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() 119 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() 120 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() 121 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() 122 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
|
D | 3x3s2p1-minmax-neonfma-1x4-acc3.c | 116 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() 117 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() 118 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() 119 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() 120 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() 121 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
|
D | 3x3s2p1-minmax-neonfma-1x4.c | 114 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() 115 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() 116 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() 117 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() 118 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() 119 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
|
D | 3x3s2p1-minmax-neonfma-1x4-acc4.c | 117 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() 118 …loat32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() 119 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() 120 …loat32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() 121 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() 122 …loat32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
|
D | 5x5s2p2-minmax-neon-1x4.c | 196 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 197 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 198 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 199 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 200 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 202 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 203 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 204 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 205 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 206 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 198 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 199 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 200 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 201 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 202 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 204 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 205 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 206 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 207 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 208 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
D | 5x5s2p2-minmax-neon-1x4-acc2.c | 197 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 198 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 199 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 200 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 201 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 203 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 204 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 205 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 206 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 207 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
D | 5x5s2p2-minmax-neon-1x4-acc4.c | 199 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 200 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 201 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 202 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 203 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 205 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 206 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 207 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 208 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 209 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
D | 5x5s2p2-minmax-neon-1x4-acc3.c | 198 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 199 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 200 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 201 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 202 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 204 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 205 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 206 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 207 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 208 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
D | 5x5s2p2-minmax-neonfma-1x4.c | 196 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 197 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 198 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 199 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 200 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 202 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 203 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 204 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 205 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 206 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 199 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 200 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 201 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 202 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 203 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 205 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 206 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 207 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 208 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 209 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 197 …loat32x4_t vi0x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi0x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 198 …loat32x4_t vi1x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi1x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 199 …loat32x4_t vi2x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi2x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 200 …loat32x4_t vi3x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi3x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 201 …loat32x4_t vi4x8ACE = vreinterpretq_f32_u32(vandq_u32(vmask_even, vreinterpretq_u32_f32(vi4x8ACE9B… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 203 …const float32x4_t vi0x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi0x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 204 …const float32x4_t vi1x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi1x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 205 …const float32x4_t vi2x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi2x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 206 …const float32x4_t vi3x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi3x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 207 …const float32x4_t vi4x9BDF = vreinterpretq_f32_u32(vandq_u32(vmask_odd, vreinterpretq_u32_f32(vi4x… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
/external/XNNPACK/src/f32-gavgpool-cw/ |
D | neon-x4.c | 60 vi0 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0))); in xnn_f32_gavgpool_cw_ukernel__neon_x4() 61 vi1 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1))); in xnn_f32_gavgpool_cw_ukernel__neon_x4() 62 vi2 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2))); in xnn_f32_gavgpool_cw_ukernel__neon_x4() 63 vi3 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi3))); in xnn_f32_gavgpool_cw_ukernel__neon_x4() 110 vi0 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0))); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
|