diff --git a/source/Lib/CommonLib/InterpolationFilter.cpp b/source/Lib/CommonLib/InterpolationFilter.cpp
index 9ff4a501217b7a83e4d85589292e0633d1078b6d..1c7ea90e2db56433da449942c32d0715aecb02b3 100644
--- a/source/Lib/CommonLib/InterpolationFilter.cpp
+++ b/source/Lib/CommonLib/InterpolationFilter.cpp
@@ -698,14 +698,11 @@ void InterpolationFilter::xWeightedTriangleBlk( const PredictionUnit &pu, const
const int32_t shiftWeighted = std::max<int>(2, (IF_INTERNAL_PREC - clipbd)) + log2WeightBase;
const int32_t offsetWeighted = (1 << (shiftWeighted - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
#if JVET_P0530_TPM_WEIGHT_ALIGN
- int32_t logSubWidthC = getChannelTypeScaleX(CHANNEL_TYPE_CHROMA, pu.chromaFormat);
- int32_t logSubHeightC = getChannelTypeScaleY(CHANNEL_TYPE_CHROMA, pu.chromaFormat);
+ int32_t stepX = 1 << getComponentScaleX(compIdx, pu.chromaFormat);
+ int32_t stepY = 1 << getComponentScaleY(compIdx, pu.chromaFormat);
- int32_t stepX = (compIdx == 0) ? 1 : (1 << logSubWidthC);
- int32_t stepY = (compIdx == 0) ? 1 : (1 << logSubHeightC);
-
- int32_t widthY = (compIdx == 0) ? width : (width << logSubWidthC);
- int32_t heightY = (compIdx == 0) ? height : (height << logSubHeightC);
+ int32_t widthY = width << getComponentScaleX(compIdx, pu.chromaFormat);
+ int32_t heightY = height << getComponentScaleY(compIdx, pu.chromaFormat);
int32_t ratioWH = (widthY > heightY) ? (widthY / heightY) : 1;
int32_t ratioHW = (widthY > heightY) ? 1 : (heightY / widthY);
@@ -737,10 +734,8 @@ void InterpolationFilter::xWeightedTriangleBlk( const PredictionUnit &pu, const
}
for (tmpY = ratioHW; tmpY > 0; tmpY -= stepY)
{
- for (x = 0; x < weightedStartPos; x++)
+ for (x = 0; x < weightedStartPos; x += stepX)
{
- if (x % stepX != 0)
- continue;
#else
for( y = 0; y < height; y+= ratioHW )
{
@@ -789,10 +784,9 @@ void InterpolationFilter::xWeightedTriangleBlk( const PredictionUnit &pu, const
}
#if JVET_P0530_TPM_WEIGHT_ALIGN
- for (x = weightedEndPos + 1; x < widthY; x++)
+ int32_t start = ((weightedEndPos + 1) % stepX != 0) ? (weightedEndPos + 2) : (weightedEndPos + 1);
+ for (x = start; x < widthY; x += stepX)
{
- if (x % stepX != 0)
- continue;
#else
for( x = weightedEndPos + 1; x < width; x++ )
{
diff --git a/source/Lib/CommonLib/Rom.cpp b/source/Lib/CommonLib/Rom.cpp
index dd7a0331ba42f4b591f9bc5e9d74e2bbf5491c84..38c8963c9dcc270c97ae3f7aa14f784d34665967 100644
--- a/source/Lib/CommonLib/Rom.cpp
+++ b/source/Lib/CommonLib/Rom.cpp
@@ -430,33 +430,25 @@ void initROM()
const int nCbR = (nCbW > nCbH) ? nCbW / nCbH : nCbH / nCbW;
// let SIMD can read at least 64-bit when at last row
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ g_triangleWeights[0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+ g_triangleWeights[1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
+#else
g_triangleWeights[0][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
g_triangleWeights[0][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
g_triangleWeights[1][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
g_triangleWeights[1][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
-#if JVET_P0530_TPM_WEIGHT_ALIGN
- g_triangleWeights[2][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
- g_triangleWeights[2][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
- g_triangleWeights[3][0][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
- g_triangleWeights[3][1][idxH][idxW] = new int16_t[nCbH * nCbW + 4];
- int nCbR_422 = (nCbW * 2 > nCbH) ? (nCbW * 2) / nCbH : nCbH / (nCbW * 2);
#endif
for (int y = 0; y < nCbH; y++)
{
for (int x = 0; x < nCbW; x++)
{
- g_triangleWeights[0][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, x / nCbR - y + 4) : Clip3(0, 8, x - y / nCbR + 4);
- g_triangleWeights[0][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH - 1 - x / nCbR - y + 4) : Clip3(0, 8, nCbW - 1 - x - y / nCbR + 4);
#if JVET_P0530_TPM_WEIGHT_ALIGN
- g_triangleWeights[CHROMA_420][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, (x * 2) / nCbR - (y * 2) + 4) : Clip3(0, 8, (x * 2) - (y * 2) / nCbR + 4);
- g_triangleWeights[CHROMA_420][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH * 2 - 1 - (x * 2) / nCbR - (y * 2) + 4) : Clip3(0, 8, nCbW * 2 - 1 - (x * 2) - (y * 2) / nCbR + 4);
-
- g_triangleWeights[CHROMA_422][0][idxH][idxW][y*nCbW + x] = (nCbW * 2 > nCbH) ? Clip3(0, 8, (x * 2) / nCbR_422 - y + 4) : Clip3(0, 8, (x * 2) - y / nCbR_422 + 4);
- g_triangleWeights[CHROMA_422][1][idxH][idxW][y*nCbW + x] = (nCbW * 2 > nCbH) ? Clip3(0, 8, nCbH - 1 - (x * 2) / nCbR_422 - y + 4) : Clip3(0, 8, nCbW * 2 - 1 - (x * 2) - y / nCbR_422 + 4);
-
- g_triangleWeights[CHROMA_444][0][idxH][idxW][y*nCbW + x] = g_triangleWeights[0][0][idxH][idxW][y*nCbW + x];
- g_triangleWeights[CHROMA_444][1][idxH][idxW][y*nCbW + x] = g_triangleWeights[0][1][idxH][idxW][y*nCbW + x];
+ g_triangleWeights[0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, x / nCbR - y + 4) : Clip3(0, 8, x - y / nCbR + 4);
+ g_triangleWeights[1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH - 1 - x / nCbR - y + 4) : Clip3(0, 8, nCbW - 1 - x - y / nCbR + 4);
#else
+ g_triangleWeights[0][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, x / nCbR - y + 4) : Clip3(0, 8, x - y / nCbR + 4);
+ g_triangleWeights[0][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 8, nCbH - 1 - x / nCbR - y + 4) : Clip3(0, 8, nCbW - 1 - x - y / nCbR + 4);
g_triangleWeights[1][0][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 4, x / nCbR - y + 2) * 2 : Clip3(0, 4, x - y / nCbR + 2) * 2;
g_triangleWeights[1][1][idxH][idxW][y*nCbW + x] = (nCbW > nCbH) ? Clip3(0, 4, nCbH - 1 - x / nCbR - y + 2) * 2 : Clip3(0, 4, nCbW - 1 - x - y / nCbR + 2) * 2;
#endif
@@ -495,6 +487,12 @@ void destroyROM()
{
for (int idxW = 0; idxW < MAX_CU_DEPTH - MIN_CU_LOG2 + 2; ++idxW)
{
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ delete[] g_triangleWeights[0][idxH][idxW];
+ delete[] g_triangleWeights[1][idxH][idxW];
+ g_triangleWeights[0][idxH][idxW] = nullptr;
+ g_triangleWeights[1][idxH][idxW] = nullptr;
+#else
delete[] g_triangleWeights[0][0][idxH][idxW];
delete[] g_triangleWeights[0][1][idxH][idxW];
delete[] g_triangleWeights[1][0][idxH][idxW];
@@ -503,15 +501,6 @@ void destroyROM()
g_triangleWeights[0][1][idxH][idxW] = nullptr;
g_triangleWeights[1][0][idxH][idxW] = nullptr;
g_triangleWeights[1][1][idxH][idxW] = nullptr;
-#if JVET_P0530_TPM_WEIGHT_ALIGN
- delete[] g_triangleWeights[2][0][idxH][idxW];
- delete[] g_triangleWeights[2][1][idxH][idxW];
- delete[] g_triangleWeights[3][0][idxH][idxW];
- delete[] g_triangleWeights[3][1][idxH][idxW];
- g_triangleWeights[2][0][idxH][idxW] = nullptr;
- g_triangleWeights[2][1][idxH][idxW] = nullptr;
- g_triangleWeights[3][0][idxH][idxW] = nullptr;
- g_triangleWeights[3][1][idxH][idxW] = nullptr;
#endif
}
}
@@ -756,7 +745,7 @@ const uint32_t g_scalingListSizeX[SCALING_LIST_SIZE_NUM] = { 1, 2, 4, 8, 16,
uint8_t g_triangleMvStorage[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_SIZE >> MIN_CU_LOG2][MAX_CU_SIZE >> MIN_CU_LOG2];
#if JVET_P0530_TPM_WEIGHT_ALIGN
-int16_t *g_triangleWeights[4][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
+int16_t *g_triangleWeights[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
#else
int16_t *g_triangleWeights[2][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
#endif
diff --git a/source/Lib/CommonLib/Rom.h b/source/Lib/CommonLib/Rom.h
index af8b3d519d7458e1facf71320552afec5183a2e3..fe144b2dfe8dde685fb8d40bf447019d4dc3b66b 100644
--- a/source/Lib/CommonLib/Rom.h
+++ b/source/Lib/CommonLib/Rom.h
@@ -212,7 +212,7 @@ constexpr uint8_t g_tbMax[257] = { 0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3,
extern uint8_t g_triangleMvStorage[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_DEPTH - MIN_CU_LOG2 + 1][MAX_CU_SIZE >> MIN_CU_LOG2][MAX_CU_SIZE >> MIN_CU_LOG2];
// 7-tap/3-tap, direction, 2/4/8/16/32/64/128
#if JVET_P0530_TPM_WEIGHT_ALIGN
-extern int16_t *g_triangleWeights[4][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
+extern int16_t *g_triangleWeights[TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
#else
extern int16_t *g_triangleWeights[2][TRIANGLE_DIR_NUM][MAX_CU_DEPTH - MIN_CU_LOG2 + 2][MAX_CU_DEPTH - MIN_CU_LOG2 + 2];
#endif
diff --git a/source/Lib/CommonLib/x86/InterpolationFilterX86.h b/source/Lib/CommonLib/x86/InterpolationFilterX86.h
index d78580bbc1a7a6f34080625b7c7030a03276df20..7d61dab311e66b6cd8f9a12a5d10cfd264ef4dd5 100644
--- a/source/Lib/CommonLib/x86/InterpolationFilterX86.h
+++ b/source/Lib/CommonLib/x86/InterpolationFilterX86.h
@@ -1227,15 +1227,22 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
int32_t strideSrc0 = predSrc0.get(compIdx).stride;
int32_t strideSrc1 = predSrc1.get(compIdx).stride;
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ int32_t chromaScaleX = getComponentScaleX(compIdx, pu.chromaFormat);
+ int32_t chromaScaleY = getComponentScaleY(compIdx, pu.chromaFormat);
+ int8_t log2WidthY = floorLog2(width << chromaScaleX) - 1;
+ int8_t log2HeightY = floorLog2(height << chromaScaleY) - 1;
+#else
int8_t log2Width = floorLog2(width) - 1;
int8_t log2Height = floorLog2(height) - 1;
+#endif
const char log2WeightBase = 3;
const ClpRng clpRng = pu.cu->slice->clpRngs().comp[compIdx];
const int32_t shiftWeighted = std::max<int>(2, (IF_INTERNAL_PREC - clpRng.bd)) + log2WeightBase;
const int32_t offsetWeighted = (1 << (shiftWeighted - 1)) + (IF_INTERNAL_OFFS << log2WeightBase);
#if JVET_P0530_TPM_WEIGHT_ALIGN
- int wIdx = (compIdx == COMPONENT_Y) ? 0 : pu.cs->sps->getChromaFormatIdc();
- int16_t *weight = g_triangleWeights[wIdx][splitDir][log2Height][log2Width];
+ int16_t *weight = g_triangleWeights[splitDir][log2HeightY][log2WidthY];
+ int16_t stepY = width << (chromaScaleX + chromaScaleY);
#else
const bool longWeight = (compIdx == COMPONENT_Y);
const bool shortWeight = !longWeight;
@@ -1251,13 +1258,23 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
if (width == 2)
{
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ const __m128i mask = _mm_set_epi16( (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080, 0x0504, 0x0100 );
+#endif
for (int y = 0; y < height; y++)
{
__m128i s0 = _mm_cvtsi32_si128(*(uint32_t *) src0);
__m128i s1 = _mm_cvtsi32_si128(*(uint32_t *) src1);
- __m128i w0 = _mm_cvtsi32_si128(*(uint32_t *) weight);
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ __m128i w0 = _mm_loadl_epi64((__m128i *) (weight));
+ if (chromaScaleX == 1)
+ {
+ w0 = _mm_shuffle_epi8(w0, mask);
+ }
+#else
+ __m128i w0 = _mm_cvtsi32_si128(*(uint32_t *)weight);
+#endif
__m128i w1 = _mm_sub_epi16(mmEight, w0);
-
s0 = _mm_unpacklo_epi16(s0, s1);
w0 = _mm_unpacklo_epi16(w0, w1);
s0 = _mm_add_epi32(_mm_madd_epi16(s0, w0), mmOffset);
@@ -1269,16 +1286,31 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
dst += strideDst;
src0 += strideSrc0;
src1 += strideSrc1;
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ weight += stepY;
+#else
weight += 2;
+#endif
}
}
else if(width == 4)
{
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ const __m128i mask = _mm_set_epi16( (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080, 0x0D0C, 0x0908, 0x0504, 0x0100 );
+#endif
for (int y = 0; y < height; y++)
{
__m128i s0 = _mm_loadl_epi64((__m128i *) (src0));
__m128i s1 = _mm_loadl_epi64((__m128i *) (src1));
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ __m128i w0 = _mm_loadu_si128((__m128i *) (weight));
+ if (chromaScaleX == 1)
+ {
+ w0 = _mm_shuffle_epi8(w0, mask);
+ }
+#else
__m128i w0 = _mm_loadl_epi64((__m128i *) (weight));
+#endif
__m128i w1 = _mm_sub_epi16(mmEight, w0);
s0 = _mm_unpacklo_epi16(s0, s1);
w0 = _mm_unpacklo_epi16(w0, w1);
@@ -1290,18 +1322,38 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
dst += strideDst;
src0 += strideSrc0;
src1 += strideSrc1;
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ weight += stepY;
+#else
weight += 4;
+#endif
}
}
else
{
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ const __m128i mask1 = _mm_set_epi16( 0x0D0C, 0x0908, 0x0504, 0x0100, (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080 );
+ const __m128i mask2 = _mm_set_epi16( (short) 0x8080, (short) 0x8080, (short) 0x8080, (short) 0x8080, 0x0D0C, 0x0908, 0x0504, 0x0100 );
+#endif
for (int y = 0; y < height; y++)
{
for (int x = 0; x < width; x += 8)
{
__m128i s0 = _mm_loadu_si128((__m128i *) (src0 + x));
__m128i s1 = _mm_loadu_si128((__m128i *) (src1 + x));
+
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ __m128i w0 = _mm_loadu_si128((__m128i *) (weight + (x << chromaScaleX)));
+ if (chromaScaleX == 1)
+ {
+ __m128i w01 = _mm_loadu_si128((__m128i *) (weight + (x << chromaScaleX) + 8));
+ w0 = _mm_shuffle_epi8(w0, mask1);
+ w01 = _mm_shuffle_epi8(w01, mask2);
+ w0 = _mm_alignr_epi8(w01, w0, 8);
+ }
+#else
__m128i w0 = _mm_loadu_si128((__m128i *) (weight + x));
+#endif
__m128i w1 = _mm_sub_epi16(mmEight, w0);
__m128i s0tmp = _mm_unpacklo_epi16(s0, s1);
@@ -1321,7 +1373,11 @@ void xWeightedTriangleBlk_SSE(const PredictionUnit &pu, const uint32_t width, co
dst += strideDst;
src0 += strideSrc0;
src1 += strideSrc1;
+#if JVET_P0530_TPM_WEIGHT_ALIGN
+ weight += stepY;
+#else
weight += width;
+#endif
}
}
}