// 'do_above' and 'do_left' facilitate branch removal when inlined. static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left, int do_above, int do_left) { uint16x8_t sum_top; uint16x8_t sum_left; uint8x8_t dc0; if (do_above) { const uint8x16_t A0 = vld1q_u8(above); // top row const uint8x16_t A1 = vld1q_u8(above + 16); const uint16x8_t p0 = vpaddlq_u8(A0); // cascading summation of the top const uint16x8_t p1 = vpaddlq_u8(A1); const uint16x8_t p2 = vaddq_u16(p0, p1); const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2)); const uint16x4_t p4 = vpadd_u16(p3, p3); const uint16x4_t p5 = vpadd_u16(p4, p4); sum_top = vcombine_u16(p5, p5); } if (do_left) { const uint8x16_t L0 = vld1q_u8(left); // left row const uint8x16_t L1 = vld1q_u8(left + 16); const uint16x8_t p0 = vpaddlq_u8(L0); // cascading summation of the left const uint16x8_t p1 = vpaddlq_u8(L1); const uint16x8_t p2 = vaddq_u16(p0, p1); const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2)); const uint16x4_t p4 = vpadd_u16(p3, p3); const uint16x4_t p5 = vpadd_u16(p4, p4); sum_left = vcombine_u16(p5, p5); } if (do_above && do_left) { const uint16x8_t sum = vaddq_u16(sum_left, sum_top); dc0 = vrshrn_n_u16(sum, 6); } else if (do_above) { dc0 = vrshrn_n_u16(sum_top, 5); } else if (do_left) { dc0 = vrshrn_n_u16(sum_left, 5); } else { dc0 = vdup_n_u8(0x80); } { const uint8x16_t dc = vdupq_lane_u8(dc0, 0); int i; for (i = 0; i < 32; ++i) { vst1q_u8(dst + i * stride, dc); vst1q_u8(dst + i * stride + 16, dc); } } }
void test_vpaddu16 (void) { uint16x4_t out_uint16x4_t; uint16x4_t arg0_uint16x4_t; uint16x4_t arg1_uint16x4_t; out_uint16x4_t = vpadd_u16 (arg0_uint16x4_t, arg1_uint16x4_t); }
// 'do_above' and 'do_left' facilitate branch removal when inlined. static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left, int do_above, int do_left) { uint16x8_t sum_top; uint16x8_t sum_left; uint8x8_t dc0; if (do_above) { const uint8x16_t A = vld1q_u8(above); // top row const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); const uint16x4_t p2 = vpadd_u16(p1, p1); const uint16x4_t p3 = vpadd_u16(p2, p2); sum_top = vcombine_u16(p3, p3); } if (do_left) { const uint8x16_t L = vld1q_u8(left); // left row const uint16x8_t p0 = vpaddlq_u8(L); // cascading summation of the left const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); const uint16x4_t p2 = vpadd_u16(p1, p1); const uint16x4_t p3 = vpadd_u16(p2, p2); sum_left = vcombine_u16(p3, p3); } if (do_above && do_left) { const uint16x8_t sum = vaddq_u16(sum_left, sum_top); dc0 = vrshrn_n_u16(sum, 5); } else if (do_above) { dc0 = vrshrn_n_u16(sum_top, 4); } else if (do_left) { dc0 = vrshrn_n_u16(sum_left, 4); } else { dc0 = vdup_n_u8(0x80); } { const uint8x16_t dc = vdupq_lane_u8(dc0, 0); int i; for (i = 0; i < 16; ++i) { vst1q_u8(dst + i * stride, dc); } } }
// 'do_above' and 'do_left' facilitate branch removal when inlined. static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left, int do_above, int do_left) { uint16x8_t sum_top; uint16x8_t sum_left; uint8x8_t dc0; if (do_above) { const uint8x8_t A = vld1_u8(above); // top row const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top const uint16x4_t p1 = vpadd_u16(p0, p0); const uint16x4_t p2 = vpadd_u16(p1, p1); sum_top = vcombine_u16(p2, p2); } if (do_left) { const uint8x8_t L = vld1_u8(left); // left border const uint16x4_t p0 = vpaddl_u8(L); // cascading summation of the left const uint16x4_t p1 = vpadd_u16(p0, p0); const uint16x4_t p2 = vpadd_u16(p1, p1); sum_left = vcombine_u16(p2, p2); } if (do_above && do_left) { const uint16x8_t sum = vaddq_u16(sum_left, sum_top); dc0 = vrshrn_n_u16(sum, 4); } else if (do_above) { dc0 = vrshrn_n_u16(sum_top, 3); } else if (do_left) { dc0 = vrshrn_n_u16(sum_left, 3); } else { dc0 = vdup_n_u8(0x80); } { const uint8x8_t dc = vdup_lane_u8(dc0, 0); int i; for (i = 0; i < 8; ++i) { vst1_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc)); } } }
uint16x4_t test_vpadd_u16 (uint16x4_t __a, uint16x4_t __b) { return vpadd_u16(__a, __b); }