aboutsummaryrefslogtreecommitdiff
path: root/src/dsp/x86/mask_blend_sse4.cc
diff options
context:
space:
mode:
Diffstat (limited to 'src/dsp/x86/mask_blend_sse4.cc')
-rw-r--r--src/dsp/x86/mask_blend_sse4.cc507
1 files changed, 505 insertions, 2 deletions
diff --git a/src/dsp/x86/mask_blend_sse4.cc b/src/dsp/x86/mask_blend_sse4.cc
index d8036be..2e836af 100644
--- a/src/dsp/x86/mask_blend_sse4.cc
+++ b/src/dsp/x86/mask_blend_sse4.cc
@@ -430,12 +430,515 @@ void Init8bpp() {
} // namespace
} // namespace low_bitdepth
-void MaskBlendInit_SSE4_1() { low_bitdepth::Init8bpp(); }
+#if LIBGAV1_MAX_BITDEPTH >= 10
+namespace high_bitdepth {
+namespace {
+
+constexpr int kMax10bppSample = (1 << 10) - 1;
+constexpr int kMaskInverse = 64;
+constexpr int kRoundBitsMaskBlend = 4;
+
+inline __m128i RightShiftWithRoundingZero_U16(const __m128i v_val_d, int bits,
+ const __m128i zero) {
+ // Shift out all but the last bit.
+ const __m128i v_tmp_d = _mm_srli_epi16(v_val_d, bits - 1);
+ // Avg with zero will shift by 1 and round.
+ return _mm_avg_epu16(v_tmp_d, zero);
+}
+
+inline __m128i RightShiftWithRoundingConst_S32(const __m128i v_val_d, int bits,
+ const __m128i shift) {
+ const __m128i v_tmp_d = _mm_add_epi32(v_val_d, shift);
+ return _mm_srai_epi32(v_tmp_d, bits);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline __m128i GetMask4x2(const uint8_t* mask, ptrdiff_t mask_stride,
+ const __m128i zero) {
+ if (subsampling_x == 1) {
+ if (subsampling_y == 0) {
+ const __m128i mask_val_0 = _mm_cvtepu8_epi16(LoadLo8(mask));
+ const __m128i mask_val_1 =
+ _mm_cvtepu8_epi16(LoadLo8(mask + (mask_stride << subsampling_y)));
+ __m128i subsampled_mask = _mm_hadd_epi16(mask_val_0, mask_val_1);
+ return RightShiftWithRoundingZero_U16(subsampled_mask, 1, zero);
+ }
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i mask_val_0 =
+ LoadHi8(LoadLo8(mask), mask + (mask_stride << 1));
+ const __m128i mask_val_1 = LoadHi8(LoadLo8(mask + mask_stride),
+ mask + (mask_stride << 1) + mask_stride);
+ const __m128i add = _mm_adds_epu8(mask_val_0, mask_val_1);
+ const __m128i subsampled_mask = _mm_maddubs_epi16(add, one);
+ return RightShiftWithRoundingZero_U16(subsampled_mask, 2, zero);
+ }
+ assert(subsampling_y == 0 && subsampling_x == 0);
+ const __m128i mask_val_0 = Load4(mask);
+ const __m128i mask_val_1 = Load4(mask + mask_stride);
+ return _mm_cvtepu8_epi16(
+ _mm_or_si128(mask_val_0, _mm_slli_si128(mask_val_1, 4)));
+}
+
+template <int subsampling_x, int subsampling_y>
+inline __m128i GetMask8(const uint8_t* mask, const ptrdiff_t stride,
+ const __m128i zero) {
+ if (subsampling_x == 1) {
+ if (subsampling_y == 0) {
+ const __m128i row_vals = LoadUnaligned16(mask);
+ const __m128i mask_val_0 = _mm_cvtepu8_epi16(row_vals);
+ const __m128i mask_val_1 = _mm_cvtepu8_epi16(_mm_srli_si128(row_vals, 8));
+ __m128i subsampled_mask = _mm_hadd_epi16(mask_val_0, mask_val_1);
+ return RightShiftWithRoundingZero_U16(subsampled_mask, 1, zero);
+ }
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i mask_val_0 = LoadUnaligned16(mask);
+ const __m128i mask_val_1 = LoadUnaligned16(mask + stride);
+ const __m128i add_0 = _mm_adds_epu8(mask_val_0, mask_val_1);
+ const __m128i mask_0 = _mm_maddubs_epi16(add_0, one);
+ return RightShiftWithRoundingZero_U16(mask_0, 2, zero);
+ }
+ assert(subsampling_y == 0 && subsampling_x == 0);
+ const __m128i mask_val = LoadLo8(mask);
+ return _mm_cvtepu8_epi16(mask_val);
+}
+
+inline void WriteMaskBlendLine10bpp4x2_SSE4_1(
+ const uint16_t* pred_0, const uint16_t* pred_1,
+ const ptrdiff_t pred_stride_1, const __m128i& pred_mask_0,
+ const __m128i& pred_mask_1, const __m128i& offset, const __m128i& max,
+ const __m128i& shift4, uint16_t* dst, const ptrdiff_t dst_stride) {
+ const __m128i pred_val_0 = LoadUnaligned16(pred_0);
+ const __m128i pred_val_1 = LoadHi8(LoadLo8(pred_1), pred_1 + pred_stride_1);
+
+ // int res = (mask_value * pred_0[x] + (64 - mask_value) * pred_1[x]) >> 6;
+ const __m128i compound_pred_lo_0 = _mm_mullo_epi16(pred_val_0, pred_mask_0);
+ const __m128i compound_pred_hi_0 = _mm_mulhi_epu16(pred_val_0, pred_mask_0);
+ const __m128i compound_pred_lo_1 = _mm_mullo_epi16(pred_val_1, pred_mask_1);
+ const __m128i compound_pred_hi_1 = _mm_mulhi_epu16(pred_val_1, pred_mask_1);
+ const __m128i pack0_lo =
+ _mm_unpacklo_epi16(compound_pred_lo_0, compound_pred_hi_0);
+ const __m128i pack0_hi =
+ _mm_unpackhi_epi16(compound_pred_lo_0, compound_pred_hi_0);
+ const __m128i pack1_lo =
+ _mm_unpacklo_epi16(compound_pred_lo_1, compound_pred_hi_1);
+ const __m128i pack1_hi =
+ _mm_unpackhi_epi16(compound_pred_lo_1, compound_pred_hi_1);
+ const __m128i compound_pred_lo = _mm_add_epi32(pack0_lo, pack1_lo);
+ const __m128i compound_pred_hi = _mm_add_epi32(pack0_hi, pack1_hi);
+ // res -= (bitdepth == 8) ? 0 : kCompoundOffset;
+ const __m128i sub_0 =
+ _mm_sub_epi32(_mm_srli_epi32(compound_pred_lo, 6), offset);
+ const __m128i sub_1 =
+ _mm_sub_epi32(_mm_srli_epi32(compound_pred_hi, 6), offset);
+
+ // dst[x] = static_cast<Pixel>(
+ // Clip3(RightShiftWithRounding(res, inter_post_round_bits), 0,
+ // (1 << kBitdepth8) - 1));
+ const __m128i shift_0 =
+ RightShiftWithRoundingConst_S32(sub_0, kRoundBitsMaskBlend, shift4);
+ const __m128i shift_1 =
+ RightShiftWithRoundingConst_S32(sub_1, kRoundBitsMaskBlend, shift4);
+ const __m128i result = _mm_min_epi16(_mm_packus_epi32(shift_0, shift_1), max);
+ StoreLo8(dst, result);
+ StoreHi8(dst + dst_stride, result);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void MaskBlend10bpp4x4_SSE4_1(const uint16_t* pred_0,
+ const uint16_t* pred_1,
+ const ptrdiff_t pred_stride_1,
+ const uint8_t* mask,
+ const ptrdiff_t mask_stride, uint16_t* dst,
+ const ptrdiff_t dst_stride) {
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i shift4 = _mm_set1_epi32((1 << kRoundBitsMaskBlend) >> 1);
+ const __m128i offset = _mm_set1_epi32(kCompoundOffset);
+ const __m128i max = _mm_set1_epi16(kMax10bppSample);
+ __m128i pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1, pred_mask_0,
+ pred_mask_1, offset, max, shift4, dst,
+ dst_stride);
+ pred_0 += 4 << 1;
+ pred_1 += pred_stride_1 << 1;
+ mask += mask_stride << (1 + subsampling_y);
+ dst += dst_stride << 1;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1, pred_mask_0,
+ pred_mask_1, offset, max, shift4, dst,
+ dst_stride);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void MaskBlend10bpp4xH_SSE4_1(const uint16_t* pred_0,
+ const uint16_t* pred_1,
+ const ptrdiff_t pred_stride_1,
+ const uint8_t* const mask_ptr,
+ const ptrdiff_t mask_stride,
+ const int height, uint16_t* dst,
+ const ptrdiff_t dst_stride) {
+ const uint8_t* mask = mask_ptr;
+ if (height == 4) {
+ MaskBlend10bpp4x4_SSE4_1<subsampling_x, subsampling_y>(
+ pred_0, pred_1, pred_stride_1, mask, mask_stride, dst, dst_stride);
+ return;
+ }
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i zero = _mm_setzero_si128();
+ const uint8_t pred0_stride2 = 4 << 1;
+ const ptrdiff_t pred1_stride2 = pred_stride_1 << 1;
+ const ptrdiff_t mask_stride2 = mask_stride << (1 + subsampling_y);
+ const ptrdiff_t dst_stride2 = dst_stride << 1;
+ const __m128i offset = _mm_set1_epi32(kCompoundOffset);
+ const __m128i max = _mm_set1_epi16(kMax10bppSample);
+ const __m128i shift4 = _mm_set1_epi32((1 << kRoundBitsMaskBlend) >> 1);
+ int y = height;
+ do {
+ __m128i pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, offset, max,
+ shift4, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, offset, max,
+ shift4, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, offset, max,
+ shift4, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ WriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, offset, max,
+ shift4, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+ y -= 8;
+ } while (y != 0);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void MaskBlend10bpp_SSE4_1(const void* prediction_0,
+ const void* prediction_1,
+ const ptrdiff_t prediction_stride_1,
+ const uint8_t* const mask_ptr,
+ const ptrdiff_t mask_stride, const int width,
+ const int height, void* dest,
+ const ptrdiff_t dest_stride) {
+ auto* dst = static_cast<uint16_t*>(dest);
+ const ptrdiff_t dst_stride = dest_stride / sizeof(dst[0]);
+ const auto* pred_0 = static_cast<const uint16_t*>(prediction_0);
+ const auto* pred_1 = static_cast<const uint16_t*>(prediction_1);
+ const ptrdiff_t pred_stride_0 = width;
+ const ptrdiff_t pred_stride_1 = prediction_stride_1;
+ if (width == 4) {
+ MaskBlend10bpp4xH_SSE4_1<subsampling_x, subsampling_y>(
+ pred_0, pred_1, pred_stride_1, mask_ptr, mask_stride, height, dst,
+ dst_stride);
+ return;
+ }
+ const uint8_t* mask = mask_ptr;
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i zero = _mm_setzero_si128();
+ const ptrdiff_t mask_stride_ss = mask_stride << subsampling_y;
+ const __m128i offset = _mm_set1_epi32(kCompoundOffset);
+ const __m128i max = _mm_set1_epi16(kMax10bppSample);
+ const __m128i shift4 = _mm_set1_epi32((1 << kRoundBitsMaskBlend) >> 1);
+ int y = height;
+ do {
+ int x = 0;
+ do {
+ const __m128i pred_mask_0 = GetMask8<subsampling_x, subsampling_y>(
+ mask + (x << subsampling_x), mask_stride, zero);
+ const __m128i pred_val_0 = LoadUnaligned16(pred_0 + x);
+ const __m128i pred_val_1 = LoadUnaligned16(pred_1 + x);
+ // 64 - mask
+ const __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+
+ const __m128i compound_pred_lo_0 =
+ _mm_mullo_epi16(pred_val_0, pred_mask_0);
+ const __m128i compound_pred_hi_0 =
+ _mm_mulhi_epu16(pred_val_0, pred_mask_0);
+ const __m128i compound_pred_lo_1 =
+ _mm_mullo_epi16(pred_val_1, pred_mask_1);
+ const __m128i compound_pred_hi_1 =
+ _mm_mulhi_epu16(pred_val_1, pred_mask_1);
+ const __m128i pack0_lo =
+ _mm_unpacklo_epi16(compound_pred_lo_0, compound_pred_hi_0);
+ const __m128i pack0_hi =
+ _mm_unpackhi_epi16(compound_pred_lo_0, compound_pred_hi_0);
+ const __m128i pack1_lo =
+ _mm_unpacklo_epi16(compound_pred_lo_1, compound_pred_hi_1);
+ const __m128i pack1_hi =
+ _mm_unpackhi_epi16(compound_pred_lo_1, compound_pred_hi_1);
+ const __m128i compound_pred_lo = _mm_add_epi32(pack0_lo, pack1_lo);
+ const __m128i compound_pred_hi = _mm_add_epi32(pack0_hi, pack1_hi);
+
+ const __m128i sub_0 =
+ _mm_sub_epi32(_mm_srli_epi32(compound_pred_lo, 6), offset);
+ const __m128i sub_1 =
+ _mm_sub_epi32(_mm_srli_epi32(compound_pred_hi, 6), offset);
+ const __m128i shift_0 =
+ RightShiftWithRoundingConst_S32(sub_0, kRoundBitsMaskBlend, shift4);
+ const __m128i shift_1 =
+ RightShiftWithRoundingConst_S32(sub_1, kRoundBitsMaskBlend, shift4);
+ const __m128i result =
+ _mm_min_epi16(_mm_packus_epi32(shift_0, shift_1), max);
+ StoreUnaligned16(dst + x, result);
+ x += 8;
+ } while (x < width);
+ dst += dst_stride;
+ pred_0 += pred_stride_0;
+ pred_1 += pred_stride_1;
+ mask += mask_stride_ss;
+ } while (--y != 0);
+}
+
+inline void InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(
+ const uint16_t* prediction_0, const uint16_t* prediction_1,
+ const ptrdiff_t pred_stride_1, const __m128i& pred_mask_0,
+ const __m128i& pred_mask_1, const __m128i& shift6, uint16_t* dst,
+ const ptrdiff_t dst_stride) {
+ const __m128i pred_val_0 = LoadUnaligned16(prediction_0);
+ const __m128i pred_val_1 =
+ LoadHi8(LoadLo8(prediction_1), prediction_1 + pred_stride_1);
+
+ const __m128i mask_0 = _mm_unpacklo_epi16(pred_mask_1, pred_mask_0);
+ const __m128i mask_1 = _mm_unpackhi_epi16(pred_mask_1, pred_mask_0);
+ const __m128i pred_0 = _mm_unpacklo_epi16(pred_val_0, pred_val_1);
+ const __m128i pred_1 = _mm_unpackhi_epi16(pred_val_0, pred_val_1);
+
+ const __m128i compound_pred_0 = _mm_madd_epi16(pred_0, mask_0);
+ const __m128i compound_pred_1 = _mm_madd_epi16(pred_1, mask_1);
+ const __m128i shift_0 =
+ RightShiftWithRoundingConst_S32(compound_pred_0, 6, shift6);
+ const __m128i shift_1 =
+ RightShiftWithRoundingConst_S32(compound_pred_1, 6, shift6);
+ const __m128i res = _mm_packus_epi32(shift_0, shift_1);
+ StoreLo8(dst, res);
+ StoreHi8(dst + dst_stride, res);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void InterIntraMaskBlend10bpp4x4_SSE4_1(
+ const uint16_t* pred_0, const uint16_t* pred_1,
+ const ptrdiff_t pred_stride_1, const uint8_t* mask,
+ const ptrdiff_t mask_stride, uint16_t* dst, const ptrdiff_t dst_stride) {
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i shift6 = _mm_set1_epi32((1 << 6) >> 1);
+ const __m128i zero = _mm_setzero_si128();
+ __m128i pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, shift6,
+ dst, dst_stride);
+ pred_0 += 4 << 1;
+ pred_1 += pred_stride_1 << 1;
+ mask += mask_stride << (1 + subsampling_y);
+ dst += dst_stride << 1;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1, shift6,
+ dst, dst_stride);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void InterIntraMaskBlend10bpp4xH_SSE4_1(const uint16_t* pred_0,
+ const uint16_t* pred_1,
+ const ptrdiff_t pred_stride_1,
+ const uint8_t* const mask_ptr,
+ const ptrdiff_t mask_stride,
+ const int height, uint16_t* dst,
+ const ptrdiff_t dst_stride) {
+ const uint8_t* mask = mask_ptr;
+ if (height == 4) {
+ InterIntraMaskBlend10bpp4x4_SSE4_1<subsampling_x, subsampling_y>(
+ pred_0, pred_1, pred_stride_1, mask, mask_stride, dst, dst_stride);
+ return;
+ }
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i zero = _mm_setzero_si128();
+ const __m128i shift6 = _mm_set1_epi32((1 << 6) >> 1);
+ const uint8_t pred0_stride2 = 4 << 1;
+ const ptrdiff_t pred1_stride2 = pred_stride_1 << 1;
+ const ptrdiff_t mask_stride2 = mask_stride << (1 + subsampling_y);
+ const ptrdiff_t dst_stride2 = dst_stride << 1;
+ int y = height;
+ do {
+ __m128i pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1,
+ shift6, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1,
+ shift6, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1,
+ shift6, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+
+ pred_mask_0 =
+ GetMask4x2<subsampling_x, subsampling_y>(mask, mask_stride, zero);
+ pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ InterIntraWriteMaskBlendLine10bpp4x2_SSE4_1(pred_0, pred_1, pred_stride_1,
+ pred_mask_0, pred_mask_1,
+ shift6, dst, dst_stride);
+ pred_0 += pred0_stride2;
+ pred_1 += pred1_stride2;
+ mask += mask_stride2;
+ dst += dst_stride2;
+ y -= 8;
+ } while (y != 0);
+}
+
+template <int subsampling_x, int subsampling_y>
+inline void InterIntraMaskBlend10bpp_SSE4_1(
+ const void* prediction_0, const void* prediction_1,
+ const ptrdiff_t prediction_stride_1, const uint8_t* const mask_ptr,
+ const ptrdiff_t mask_stride, const int width, const int height, void* dest,
+ const ptrdiff_t dest_stride) {
+ auto* dst = static_cast<uint16_t*>(dest);
+ const ptrdiff_t dst_stride = dest_stride / sizeof(dst[0]);
+ const auto* pred_0 = static_cast<const uint16_t*>(prediction_0);
+ const auto* pred_1 = static_cast<const uint16_t*>(prediction_1);
+ const ptrdiff_t pred_stride_0 = width;
+ const ptrdiff_t pred_stride_1 = prediction_stride_1;
+ if (width == 4) {
+ InterIntraMaskBlend10bpp4xH_SSE4_1<subsampling_x, subsampling_y>(
+ pred_0, pred_1, pred_stride_1, mask_ptr, mask_stride, height, dst,
+ dst_stride);
+ return;
+ }
+ const uint8_t* mask = mask_ptr;
+ const __m128i mask_inverter = _mm_set1_epi16(kMaskInverse);
+ const __m128i shift6 = _mm_set1_epi32((1 << 6) >> 1);
+ const __m128i zero = _mm_setzero_si128();
+ const ptrdiff_t mask_stride_ss = mask_stride << subsampling_y;
+ int y = height;
+ do {
+ int x = 0;
+ do {
+ const __m128i pred_mask_0 = GetMask8<subsampling_x, subsampling_y>(
+ mask + (x << subsampling_x), mask_stride, zero);
+ const __m128i pred_val_0 = LoadUnaligned16(pred_0 + x);
+ const __m128i pred_val_1 = LoadUnaligned16(pred_1 + x);
+ // 64 - mask
+ const __m128i pred_mask_1 = _mm_sub_epi16(mask_inverter, pred_mask_0);
+ const __m128i mask_0 = _mm_unpacklo_epi16(pred_mask_1, pred_mask_0);
+ const __m128i mask_1 = _mm_unpackhi_epi16(pred_mask_1, pred_mask_0);
+ const __m128i pred_0 = _mm_unpacklo_epi16(pred_val_0, pred_val_1);
+ const __m128i pred_1 = _mm_unpackhi_epi16(pred_val_0, pred_val_1);
+
+ const __m128i compound_pred_0 = _mm_madd_epi16(pred_0, mask_0);
+ const __m128i compound_pred_1 = _mm_madd_epi16(pred_1, mask_1);
+ const __m128i shift_0 =
+ RightShiftWithRoundingConst_S32(compound_pred_0, 6, shift6);
+ const __m128i shift_1 =
+ RightShiftWithRoundingConst_S32(compound_pred_1, 6, shift6);
+ StoreUnaligned16(dst + x, _mm_packus_epi32(shift_0, shift_1));
+ x += 8;
+ } while (x < width);
+ dst += dst_stride;
+ pred_0 += pred_stride_0;
+ pred_1 += pred_stride_1;
+ mask += mask_stride_ss;
+ } while (--y != 0);
+}
+
+void Init10bpp() {
+ Dsp* const dsp = dsp_internal::GetWritableDspTable(kBitdepth10);
+ assert(dsp != nullptr);
+
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlend444)
+ dsp->mask_blend[0][0] = MaskBlend10bpp_SSE4_1<0, 0>;
+#endif
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlend422)
+ dsp->mask_blend[1][0] = MaskBlend10bpp_SSE4_1<1, 0>;
+#endif
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlend420)
+ dsp->mask_blend[2][0] = MaskBlend10bpp_SSE4_1<1, 1>;
+#endif
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlendInterIntra444)
+ dsp->mask_blend[0][1] = InterIntraMaskBlend10bpp_SSE4_1<0, 0>;
+#endif
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlendInterIntra422)
+ dsp->mask_blend[1][1] = InterIntraMaskBlend10bpp_SSE4_1<1, 0>;
+#endif
+#if DSP_ENABLED_10BPP_SSE4_1(MaskBlendInterIntra420)
+ dsp->mask_blend[2][1] = InterIntraMaskBlend10bpp_SSE4_1<1, 1>;
+#endif
+}
+
+} // namespace
+} // namespace high_bitdepth
+#endif // LIBGAV1_MAX_BITDEPTH >= 10
+
+void MaskBlendInit_SSE4_1() {
+ low_bitdepth::Init8bpp();
+#if LIBGAV1_MAX_BITDEPTH >= 10
+ high_bitdepth::Init10bpp();
+#endif // LIBGAV1_MAX_BITDEPTH >= 10
+}
} // namespace dsp
} // namespace libgav1
-#else // !LIBGAV1_TARGETING_SSE4_1
+#else // !LIBGAV1_TARGETING_SSE4_1
namespace libgav1 {
namespace dsp {
generated by cgit v1.2.3 (git 2.39.1) at 2025-04-04 11:47:04 +0000