rfx_dwt_2d_decode_block_horiz_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband_width) { int y, n; INT16 * l_ptr = l; INT16 * h_ptr = h; INT16 * dst_ptr = dst; for (y = 0; y < subband_width; y++) { /* Even coefficients */ for (n = 0; n < subband_width; n+=8) { // dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); int16x8_t l_n = vld1q_s16(l_ptr); int16x8_t h_n = vld1q_s16(h_ptr); int16x8_t h_n_m = vld1q_s16(h_ptr - 1); if (n == 0) { int16_t first = vgetq_lane_s16(h_n_m, 1); h_n_m = vsetq_lane_s16(first, h_n_m, 0); } int16x8_t tmp_n = vaddq_s16(h_n, h_n_m); tmp_n = vaddq_s16(tmp_n, vdupq_n_s16(1)); tmp_n = vshrq_n_s16(tmp_n, 1); int16x8_t dst_n = vsubq_s16(l_n, tmp_n); vst1q_s16(l_ptr, dst_n); l_ptr+=8; h_ptr+=8; } l_ptr -= subband_width; h_ptr -= subband_width; /* Odd coefficients */ for (n = 0; n < subband_width; n+=8) { // dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); int16x8_t h_n = vld1q_s16(h_ptr); h_n = vshlq_n_s16(h_n, 1); int16x8x2_t dst_n; dst_n.val[0] = vld1q_s16(l_ptr); int16x8_t dst_n_p = vld1q_s16(l_ptr + 1); if (n == subband_width - 8) { int16_t last = vgetq_lane_s16(dst_n_p, 6); dst_n_p = vsetq_lane_s16(last, dst_n_p, 7); } dst_n.val[1] = vaddq_s16(dst_n_p, dst_n.val[0]); dst_n.val[1] = vshrq_n_s16(dst_n.val[1], 1); dst_n.val[1] = vaddq_s16(dst_n.val[1], h_n); vst2q_s16(dst_ptr, dst_n); l_ptr+=8; h_ptr+=8; dst_ptr+=16; } } }
void UpsampleRgbaLinePairNEON(const uint8_t *top_y, const uint8_t *bottom_y, const uint8_t *top_u, const uint8_t *top_v, const uint8_t *cur_u, const uint8_t *cur_v, uint8_t *top_dst, uint8_t *bottom_dst, int len) { int block; uint8_t uv_buf[2 * 32 + 15]; uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); const int uv_len = (len + 1) >> 1; const int num_blocks = (uv_len - 1) >> 3; const int leftover = uv_len - num_blocks * 8; const int last_pos = 1 + 16 * num_blocks; const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; const int16x4_t cf16 = vld1_s16(coef); const int32x2_t cf32 = vmov_n_s32(76283); const uint8x8_t u16 = vmov_n_u8(16); const uint8x8_t u128 = vmov_n_u8(128); for (block = 0; block < num_blocks; ++block) { { uint8x8_t a = vld1_u8(top_u); uint8x8_t b = vld1_u8(top_u + 1); uint8x8_t c = vld1_u8(cur_u); uint8x8_t d = vld1_u8(cur_u + 1); uint16x8_t al = vshll_n_u8(a, 1); uint16x8_t bl = vshll_n_u8(b, 1); uint16x8_t cl = vshll_n_u8(c, 1); uint16x8_t dl = vshll_n_u8(d, 1); uint8x8_t diag1, diag2; uint16x8_t sl; sl = vaddl_u8(a, b); sl = vaddw_u8(sl, c); sl = vaddw_u8(sl, d); al = vaddq_u16(sl, al); bl = vaddq_u16(sl, bl); al = vaddq_u16(al, dl); bl = vaddq_u16(bl, cl); diag2 = vshrn_n_u16(al, 3); diag1 = vshrn_n_u16(bl, 3); a = vrhadd_u8(a, diag1); b = vrhadd_u8(b, diag2); c = vrhadd_u8(c, diag2); d = vrhadd_u8(d, diag1); { const uint8x8x2_t a_b = {{ a, b }}; const uint8x8x2_t c_d = {{ c, d }}; vst2_u8(r_uv, a_b); vst2_u8(r_uv + 32, c_d); } } { uint8x8_t a = vld1_u8(top_v); uint8x8_t b = vld1_u8(top_v + 1); uint8x8_t c = vld1_u8(cur_v); uint8x8_t d = vld1_u8(cur_v + 1); uint16x8_t al = vshll_n_u8(a, 1); uint16x8_t bl = vshll_n_u8(b, 1); uint16x8_t cl = vshll_n_u8(c, 1); uint16x8_t dl = vshll_n_u8(d, 1); uint8x8_t diag1, diag2; uint16x8_t sl; sl = vaddl_u8(a, b); sl = vaddw_u8(sl, c); sl = vaddw_u8(sl, d); al = vaddq_u16(sl, al); bl = vaddq_u16(sl, bl); al = vaddq_u16(al, dl); bl = vaddq_u16(bl, cl); diag2 = vshrn_n_u16(al, 3); diag1 = vshrn_n_u16(bl, 3); a = vrhadd_u8(a, diag1); b = vrhadd_u8(b, diag2); c = vrhadd_u8(c, diag2); d = vrhadd_u8(d, diag1); { const uint8x8x2_t a_b = {{ a, b }}; const uint8x8x2_t c_d = {{ c, d }}; vst2_u8(r_uv + 16, a_b); vst2_u8(r_uv + 16 + 32, c_d); } } { if (top_y) { { int i; for (i = 0; i < 16; i += 8) { int off = ((16 * block + 1) + i) * 4; uint8x8_t y = vld1_u8(top_y + (16 * block + 1) + i); uint8x8_t u = vld1_u8((r_uv) + i); uint8x8_t v = vld1_u8((r_uv) + i + 16); int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); int16x8_t ud = vshlq_n_s16(uu, 1); int16x8_t vd = vshlq_n_s16(vv, 1); int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), vget_low_s16(vd), cf16, 0); int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), vget_high_s16(vd), cf16, 0); int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), vrshrn_n_s32(vrh, 16)); int32x4_t vl = vmovl_s16(vget_low_s16(vv)); int32x4_t vh = vmovl_s16(vget_high_s16(vv)); int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1); int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1); int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2); int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2); int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), vrshrn_n_s32(gch, 16)); int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), vget_low_s16(ud), cf16, 3); int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), vget_high_s16(ud), cf16, 3); int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), vrshrn_n_s32(ubh, 16)); int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr)); int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr)); int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc)); int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc)); int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub)); int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub)); rl = vmulq_lane_s32(rl, cf32, 0); rh = vmulq_lane_s32(rh, cf32, 0); gl = vmulq_lane_s32(gl, cf32, 0); gh = vmulq_lane_s32(gh, cf32, 0); bl = vmulq_lane_s32(bl, cf32, 0); bh = vmulq_lane_s32(bh, cf32, 0); y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), vrshrn_n_s32(rh, 16))); u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), vrshrn_n_s32(gh, 16))); v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), vrshrn_n_s32(bh, 16))); do { const uint8x8x4_t r_g_b_v255 = {{ y, u, v, vmov_n_u8(255) }}; vst4_u8(top_dst + off, r_g_b_v255); } while (0); } } } if (bottom_y) { { int i; for (i = 0; i < 16; i += 8) { int off = ((16 * block + 1) + i) * 4; uint8x8_t y = vld1_u8(bottom_y + (16 * block + 1) + i); uint8x8_t u = vld1_u8(((r_uv) + 32) + i); uint8x8_t v = vld1_u8(((r_uv) + 32) + i + 16); int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); int16x8_t ud = vshlq_n_s16(uu, 1); int16x8_t vd = vshlq_n_s16(vv, 1); int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), vget_low_s16(vd), cf16, 0); int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), vget_high_s16(vd), cf16, 0); int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), vrshrn_n_s32(vrh, 16)); int32x4_t vl = vmovl_s16(vget_low_s16(vv)); int32x4_t vh = vmovl_s16(vget_high_s16(vv)); int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1); int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1); int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2); int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2); int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), vrshrn_n_s32(gch, 16)); int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), vget_low_s16(ud), cf16, 3); int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), vget_high_s16(ud), cf16, 3); int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), vrshrn_n_s32(ubh, 16)); int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr)); int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr)); int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc)); int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc)); int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub)); int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub)); rl = vmulq_lane_s32(rl, cf32, 0); rh = vmulq_lane_s32(rh, cf32, 0); gl = vmulq_lane_s32(gl, cf32, 0); gh = vmulq_lane_s32(gh, cf32, 0); bl = vmulq_lane_s32(bl, cf32, 0); bh = vmulq_lane_s32(bh, cf32, 0); y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), vrshrn_n_s32(rh, 16))); u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), vrshrn_n_s32(gh, 16))); v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), vrshrn_n_s32(bh, 16))); do { const uint8x8x4_t r_g_b_v255 = {{ y, u, v, vmov_n_u8(255) }}; vst4_u8(bottom_dst + off, r_g_b_v255); } while (0); } } } } } }
rfx_dwt_2d_decode_block_vert_NEON(INT16 * l, INT16 * h, INT16 * dst, int subband_width) { int x, n; INT16 * l_ptr = l; INT16 * h_ptr = h; INT16 * dst_ptr = dst; int total_width = subband_width + subband_width; /* Even coefficients */ for (n = 0; n < subband_width; n++) { for (x = 0; x < total_width; x+=8) { // dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1); int16x8_t l_n = vld1q_s16(l_ptr); int16x8_t h_n = vld1q_s16(h_ptr); int16x8_t tmp_n = vaddq_s16(h_n, vdupq_n_s16(1));; if (n == 0) tmp_n = vaddq_s16(tmp_n, h_n); else { int16x8_t h_n_m = vld1q_s16((h_ptr - total_width)); tmp_n = vaddq_s16(tmp_n, h_n_m); } tmp_n = vshrq_n_s16(tmp_n, 1); int16x8_t dst_n = vsubq_s16(l_n, tmp_n); vst1q_s16(dst_ptr, dst_n); l_ptr+=8; h_ptr+=8; dst_ptr+=8; } dst_ptr+=total_width; } h_ptr = h; dst_ptr = dst + total_width; /* Odd coefficients */ for (n = 0; n < subband_width; n++) { for (x = 0; x < total_width; x+=8) { // dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1); int16x8_t h_n = vld1q_s16(h_ptr); int16x8_t dst_n_m = vld1q_s16(dst_ptr - total_width); h_n = vshlq_n_s16(h_n, 1); int16x8_t tmp_n = dst_n_m; if (n == subband_width - 1) tmp_n = vaddq_s16(tmp_n, dst_n_m); else { int16x8_t dst_n_p = vld1q_s16((dst_ptr + total_width)); tmp_n = vaddq_s16(tmp_n, dst_n_p); } tmp_n = vshrq_n_s16(tmp_n, 1); int16x8_t dst_n = vaddq_s16(tmp_n, h_n); vst1q_s16(dst_ptr, dst_n); h_ptr+=8; dst_ptr+=8; } dst_ptr+=total_width; } }
void vp8_short_fdct8x4_neon( int16_t *input, int16_t *output, int pitch) { int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16; int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16; uint16x4_t d28u16, d29u16; uint16x8_t q14u16; int16x8_t q0s16, q1s16, q2s16, q3s16; int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16; int32x4_t q9s32, q10s32, q11s32, q12s32; int16x8x2_t v2tmp0, v2tmp1; int32x4x2_t v2tmp2, v2tmp3; d16s16 = vdup_n_s16(5352); d17s16 = vdup_n_s16(2217); q9s32 = vdupq_n_s32(14500); q10s32 = vdupq_n_s32(7500); // Part one pitch >>= 1; q0s16 = vld1q_s16(input); input += pitch; q1s16 = vld1q_s16(input); input += pitch; q2s16 = vld1q_s16(input); input += pitch; q3s16 = vld1q_s16(input); v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16), vreinterpretq_s32_s16(q2s16)); v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16), vreinterpretq_s32_s16(q3s16)); v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0 vreinterpretq_s16_s32(v2tmp3.val[0])); // q1 v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2 vreinterpretq_s16_s32(v2tmp3.val[1])); // q3 q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]); q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]); q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]); q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]); q11s16 = vshlq_n_s16(q11s16, 3); q12s16 = vshlq_n_s16(q12s16, 3); q13s16 = vshlq_n_s16(q13s16, 3); q14s16 = vshlq_n_s16(q14s16, 3); q0s16 = vaddq_s16(q11s16, q12s16); q2s16 = vsubq_s16(q11s16, q12s16); q11s32 = q9s32; q12s32 = q10s32; d26s16 = vget_low_s16(q13s16); d27s16 = vget_high_s16(q13s16); d28s16 = vget_low_s16(q14s16); d29s16 = vget_high_s16(q14s16); q9s32 = vmlal_s16(q9s32, d28s16, d16s16); q10s32 = vmlal_s16(q10s32, d28s16, d17s16); q11s32 = vmlal_s16(q11s32, d29s16, d16s16); q12s32 = vmlal_s16(q12s32, d29s16, d17s16); q9s32 = vmlal_s16(q9s32, d26s16, d17s16); q10s32 = vmlsl_s16(q10s32, d26s16, d16s16); q11s32 = vmlal_s16(q11s32, d27s16, d17s16); q12s32 = vmlsl_s16(q12s32, d27s16, d16s16); d2s16 = vshrn_n_s32(q9s32, 12); d6s16 = vshrn_n_s32(q10s32, 12); d3s16 = vshrn_n_s32(q11s32, 12); d7s16 = vshrn_n_s32(q12s32, 12); q1s16 = vcombine_s16(d2s16, d3s16); q3s16 = vcombine_s16(d6s16, d7s16); // Part two q9s32 = vdupq_n_s32(12000); q10s32 = vdupq_n_s32(51000); v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16), vreinterpretq_s32_s16(q2s16)); v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16), vreinterpretq_s32_s16(q3s16)); v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]), // q0 vreinterpretq_s16_s32(v2tmp3.val[0])); // q1 v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]), // q2 vreinterpretq_s16_s32(v2tmp3.val[1])); // q3 q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]); q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]); q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]); q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]); q15s16 = vdupq_n_s16(7); q11s16 = vaddq_s16(q11s16, q15s16); q0s16 = vaddq_s16(q11s16, q12s16); q1s16 = vsubq_s16(q11s16, q12s16); q11s32 = q9s32; q12s32 = q10s32; d0s16 = vget_low_s16(q0s16); d1s16 = vget_high_s16(q0s16); d2s16 = vget_low_s16(q1s16); d3s16 = vget_high_s16(q1s16); d0s16 = vshr_n_s16(d0s16, 4); d4s16 = vshr_n_s16(d1s16, 4); d2s16 = vshr_n_s16(d2s16, 4); d6s16 = vshr_n_s16(d3s16, 4); d26s16 = vget_low_s16(q13s16); d27s16 = vget_high_s16(q13s16); d28s16 = vget_low_s16(q14s16); d29s16 = vget_high_s16(q14s16); q9s32 = vmlal_s16(q9s32, d28s16, d16s16); q10s32 = vmlal_s16(q10s32, d28s16, d17s16); q11s32 = vmlal_s16(q11s32, d29s16, d16s16); q12s32 = vmlal_s16(q12s32, d29s16, d17s16); q9s32 = vmlal_s16(q9s32, d26s16, d17s16); q10s32 = vmlsl_s16(q10s32, d26s16, d16s16); q11s32 = vmlal_s16(q11s32, d27s16, d17s16); q12s32 = vmlsl_s16(q12s32, d27s16, d16s16); d1s16 = vshrn_n_s32(q9s32, 16); d3s16 = vshrn_n_s32(q10s32, 16); d5s16 = vshrn_n_s32(q11s32, 16); d7s16 = vshrn_n_s32(q12s32, 16); qEmptys16 = vdupq_n_s16(0); q14u16 = vceqq_s16(q14s16, qEmptys16); q14u16 = vmvnq_u16(q14u16); d28u16 = vget_low_u16(q14u16); d29u16 = vget_high_u16(q14u16); d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16)); d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16)); q0s16 = vcombine_s16(d0s16, d1s16); q1s16 = vcombine_s16(d2s16, d3s16); q2s16 = vcombine_s16(d4s16, d5s16); q3s16 = vcombine_s16(d6s16, d7s16); vst1q_s16(output, q0s16); vst1q_s16(output + 8, q1s16); vst1q_s16(output + 16, q2s16); vst1q_s16(output + 24, q3s16); return; }