/external/XNNPACK/src/f32-vbinary/gen/ |
D | vmaxc-neon-x4.c | 47 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmaxc_ukernel__neon_x4() local 49 vst1_f32(y, vy01); y += 2; in xnn_f32_vmaxc_ukernel__neon_x4() 50 vy01 = vget_high_f32(vy0123); in xnn_f32_vmaxc_ukernel__neon_x4() 53 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmaxc_ukernel__neon_x4()
|
D | vminc-neon-x4.c | 47 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vminc_ukernel__neon_x4() local 49 vst1_f32(y, vy01); y += 2; in xnn_f32_vminc_ukernel__neon_x4() 50 vy01 = vget_high_f32(vy0123); in xnn_f32_vminc_ukernel__neon_x4() 53 vst1_lane_f32(y, vy01, 0); in xnn_f32_vminc_ukernel__neon_x4()
|
D | vmin-neon-x4.c | 48 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmin_ukernel__neon_x4() local 50 vst1_f32(y, vy01); y += 2; in xnn_f32_vmin_ukernel__neon_x4() 51 vy01 = vget_high_f32(vy0123); in xnn_f32_vmin_ukernel__neon_x4() 54 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmin_ukernel__neon_x4()
|
D | vmax-neon-x4.c | 48 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmax_ukernel__neon_x4() local 50 vst1_f32(y, vy01); y += 2; in xnn_f32_vmax_ukernel__neon_x4() 51 vy01 = vget_high_f32(vy0123); in xnn_f32_vmax_ukernel__neon_x4() 54 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmax_ukernel__neon_x4()
|
D | vsqrdiffc-neon-x4.c | 49 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsqrdiffc_ukernel__neon_x4() local 51 vst1_f32(y, vy01); y += 2; in xnn_f32_vsqrdiffc_ukernel__neon_x4() 52 vy01 = vget_high_f32(vy0123); in xnn_f32_vsqrdiffc_ukernel__neon_x4() 55 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsqrdiffc_ukernel__neon_x4()
|
D | vsqrdiff-neon-x4.c | 50 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsqrdiff_ukernel__neon_x4() local 52 vst1_f32(y, vy01); y += 2; in xnn_f32_vsqrdiff_ukernel__neon_x4() 53 vy01 = vget_high_f32(vy0123); in xnn_f32_vsqrdiff_ukernel__neon_x4() 56 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsqrdiff_ukernel__neon_x4()
|
D | vadd-minmax-neon-x4.c | 55 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vadd_minmax_ukernel__neon_x4() local 57 vst1_f32(y, vy01); y += 2; in xnn_f32_vadd_minmax_ukernel__neon_x4() 58 vy01 = vget_high_f32(vy0123); in xnn_f32_vadd_minmax_ukernel__neon_x4() 61 vst1_lane_f32(y, vy01, 0); in xnn_f32_vadd_minmax_ukernel__neon_x4()
|
D | vmulc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmulc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vmulc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vmulc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmulc_minmax_ukernel__neon_x4()
|
D | vrdivc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vrdivc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vrdivc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vrdivc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vrdivc_minmax_ukernel__neon_x4()
|
D | vdivc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vdivc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vdivc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vdivc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vdivc_minmax_ukernel__neon_x4()
|
D | vrsubc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vrsubc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vrsubc_minmax_ukernel__neon_x4()
|
D | vmul-minmax-neon-x4.c | 55 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmul_minmax_ukernel__neon_x4() local 57 vst1_f32(y, vy01); y += 2; in xnn_f32_vmul_minmax_ukernel__neon_x4() 58 vy01 = vget_high_f32(vy0123); in xnn_f32_vmul_minmax_ukernel__neon_x4() 61 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmul_minmax_ukernel__neon_x4()
|
D | vminc-neon-x8.c | 56 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vminc_ukernel__neon_x8() local 58 vst1_f32(y, vy01); y += 2; in xnn_f32_vminc_ukernel__neon_x8() 59 vy01 = vget_high_f32(vy0123); in xnn_f32_vminc_ukernel__neon_x8() 62 vst1_lane_f32(y, vy01, 0); in xnn_f32_vminc_ukernel__neon_x8()
|
D | vaddc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vaddc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vaddc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vaddc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vaddc_minmax_ukernel__neon_x4()
|
D | vsub-minmax-neon-x4.c | 55 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsub_minmax_ukernel__neon_x4() local 57 vst1_f32(y, vy01); y += 2; in xnn_f32_vsub_minmax_ukernel__neon_x4() 58 vy01 = vget_high_f32(vy0123); in xnn_f32_vsub_minmax_ukernel__neon_x4() 61 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsub_minmax_ukernel__neon_x4()
|
D | vmaxc-neon-x8.c | 56 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmaxc_ukernel__neon_x8() local 58 vst1_f32(y, vy01); y += 2; in xnn_f32_vmaxc_ukernel__neon_x8() 59 vy01 = vget_high_f32(vy0123); in xnn_f32_vmaxc_ukernel__neon_x8() 62 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmaxc_ukernel__neon_x8()
|
D | vdiv-minmax-neon-x4.c | 55 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vdiv_minmax_ukernel__neon_x4() local 57 vst1_f32(y, vy01); y += 2; in xnn_f32_vdiv_minmax_ukernel__neon_x4() 58 vy01 = vget_high_f32(vy0123); in xnn_f32_vdiv_minmax_ukernel__neon_x4() 61 vst1_lane_f32(y, vy01, 0); in xnn_f32_vdiv_minmax_ukernel__neon_x4()
|
D | vsubc-minmax-neon-x4.c | 54 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsubc_minmax_ukernel__neon_x4() local 56 vst1_f32(y, vy01); y += 2; in xnn_f32_vsubc_minmax_ukernel__neon_x4() 57 vy01 = vget_high_f32(vy0123); in xnn_f32_vsubc_minmax_ukernel__neon_x4() 60 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsubc_minmax_ukernel__neon_x4()
|
D | vmin-neon-x8.c | 59 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmin_ukernel__neon_x8() local 61 vst1_f32(y, vy01); y += 2; in xnn_f32_vmin_ukernel__neon_x8() 62 vy01 = vget_high_f32(vy0123); in xnn_f32_vmin_ukernel__neon_x8() 65 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmin_ukernel__neon_x8()
|
D | vmax-neon-x8.c | 59 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vmax_ukernel__neon_x8() local 61 vst1_f32(y, vy01); y += 2; in xnn_f32_vmax_ukernel__neon_x8() 62 vy01 = vget_high_f32(vy0123); in xnn_f32_vmax_ukernel__neon_x8() 65 vst1_lane_f32(y, vy01, 0); in xnn_f32_vmax_ukernel__neon_x8()
|
D | vsqrdiffc-neon-x8.c | 60 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsqrdiffc_ukernel__neon_x8() local 62 vst1_f32(y, vy01); y += 2; in xnn_f32_vsqrdiffc_ukernel__neon_x8() 63 vy01 = vget_high_f32(vy0123); in xnn_f32_vsqrdiffc_ukernel__neon_x8() 66 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsqrdiffc_ukernel__neon_x8()
|
D | vsqrdiff-neon-x8.c | 63 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vsqrdiff_ukernel__neon_x8() local 65 vst1_f32(y, vy01); y += 2; in xnn_f32_vsqrdiff_ukernel__neon_x8() 66 vy01 = vget_high_f32(vy0123); in xnn_f32_vsqrdiff_ukernel__neon_x8() 69 vst1_lane_f32(y, vy01, 0); in xnn_f32_vsqrdiff_ukernel__neon_x8()
|
D | vrsubc-minmax-neon-x8.c | 67 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x8() local 69 vst1_f32(y, vy01); y += 2; in xnn_f32_vrsubc_minmax_ukernel__neon_x8() 70 vy01 = vget_high_f32(vy0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x8() 73 vst1_lane_f32(y, vy01, 0); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
|
D | vdivc-minmax-neon-x8.c | 67 float32x2_t vy01 = vget_low_f32(vy0123); in xnn_f32_vdivc_minmax_ukernel__neon_x8() local 69 vst1_f32(y, vy01); y += 2; in xnn_f32_vdivc_minmax_ukernel__neon_x8() 70 vy01 = vget_high_f32(vy0123); in xnn_f32_vdivc_minmax_ukernel__neon_x8() 73 vst1_lane_f32(y, vy01, 0); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
|
/external/XNNPACK/src/qu8-vadd/ |
D | minmax-neon.c | 92 uint8x16_t vy01 = vqmovun_high_s16(vqmovun_s16(vacc0), vacc1); in xnn_qu8_vadd_minmax_ukernel__neon() local 95 vy01 = vmaxq_u8(vy01, vy_min); in xnn_qu8_vadd_minmax_ukernel__neon() 97 vy01 = vminq_u8(vy01, vy_max); in xnn_qu8_vadd_minmax_ukernel__neon() 100 vst1q_u8(y, vy01); y += 16; in xnn_qu8_vadd_minmax_ukernel__neon() 143 uint8x16_t vy01 = vcombine_u8(vqmovun_s16(vacc0), vqmovun_s16(vacc1)); in xnn_qu8_vadd_minmax_ukernel__neon() local 144 vy01 = vmaxq_u8(vy01, vy_min); in xnn_qu8_vadd_minmax_ukernel__neon() 145 vy01 = vminq_u8(vy01, vy_max); in xnn_qu8_vadd_minmax_ukernel__neon() 147 vst1q_u8(y, vy01); y += 16; in xnn_qu8_vadd_minmax_ukernel__neon()
|