Compress full-quality DXT1 on AVX2 path.

client/TracyDxt1.cpp

@@ -90,56 +90,6 @@ static const uint16_t DivTable[255*3+1] = {
     0x0163, 0x0163, 0x0162, 0x0162, 0x0161, 0x0161, 0x0160, 0x0160, 0x015f, 0x015f, 0x015e, 0x015e, 0x015d, 0x015d, 0x015d, 0x015c,
     0x015c, 0x015b, 0x015b, 0x015a, 0x015a, 0x0159, 0x0159, 0x0158, 0x0158, 0x0158, 0x0157, 0x0157, 0x0156, 0x0156
 };
-static const uint16_t DivTableAVX[255*3+1] = {
-    0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
-    0x0000, 0x38e3, 0x35e5, 0x3333, 0x30c3, 0x2e8b, 0x2c85, 0x2aaa, 0x28f5, 0x2762, 0x25ed, 0x2492, 0x234f, 0x2222, 0x2108, 0x2000,
-    0x1f07, 0x1e1e, 0x1d41, 0x1c71, 0x1bac, 0x1af2, 0x1a41, 0x1999, 0x18f9, 0x1861, 0x17d0, 0x1745, 0x16c1, 0x1642, 0x15c9, 0x1555,
-    0x14e5, 0x147a, 0x1414, 0x13b1, 0x1352, 0x12f6, 0x129e, 0x1249, 0x11f7, 0x11a7, 0x115b, 0x1111, 0x10c9, 0x1084, 0x1041, 0x1000,
-    0x0fc0, 0x0f83, 0x0f48, 0x0f0f, 0x0ed7, 0x0ea0, 0x0e6c, 0x0e38, 0x0e07, 0x0dd6, 0x0da7, 0x0d79, 0x0d4c, 0x0d20, 0x0cf6, 0x0ccc,
-    0x0ca4, 0x0c7c, 0x0c56, 0x0c30, 0x0c0c, 0x0be8, 0x0bc5, 0x0ba2, 0x0b81, 0x0b60, 0x0b40, 0x0b21, 0x0b02, 0x0ae4, 0x0ac7, 0x0aaa,
-    0x0a8e, 0x0a72, 0x0a57, 0x0a3d, 0x0a23, 0x0a0a, 0x09f1, 0x09d8, 0x09c0, 0x09a9, 0x0991, 0x097b, 0x0964, 0x094f, 0x0939, 0x0924,
-    0x090f, 0x08fb, 0x08e7, 0x08d3, 0x08c0, 0x08ad, 0x089a, 0x0888, 0x0876, 0x0864, 0x0853, 0x0842, 0x0831, 0x0820, 0x0810, 0x0800,
-    0x07f0, 0x07e0, 0x07d1, 0x07c1, 0x07b3, 0x07a4, 0x0795, 0x0787, 0x0779, 0x076b, 0x075d, 0x0750, 0x0743, 0x0736, 0x0729, 0x071c,
-    0x070f, 0x0703, 0x06f7, 0x06eb, 0x06df, 0x06d3, 0x06c8, 0x06bc, 0x06b1, 0x06a6, 0x069b, 0x0690, 0x0685, 0x067b, 0x0670, 0x0666,
-    0x065c, 0x0652, 0x0648, 0x063e, 0x0634, 0x062b, 0x0621, 0x0618, 0x060f, 0x0606, 0x05fd, 0x05f4, 0x05eb, 0x05e2, 0x05d9, 0x05d1,
-    0x05c9, 0x05c0, 0x05b8, 0x05b0, 0x05a8, 0x05a0, 0x0598, 0x0590, 0x0588, 0x0581, 0x0579, 0x0572, 0x056b, 0x0563, 0x055c, 0x0555,
-    0x054e, 0x0547, 0x0540, 0x0539, 0x0532, 0x052b, 0x0525, 0x051e, 0x0518, 0x0511, 0x050b, 0x0505, 0x04fe, 0x04f8, 0x04f2, 0x04ec,
-    0x04e6, 0x04e0, 0x04da, 0x04d4, 0x04ce, 0x04c8, 0x04c3, 0x04bd, 0x04b8, 0x04b2, 0x04ad, 0x04a7, 0x04a2, 0x049c, 0x0497, 0x0492,
-    0x048d, 0x0487, 0x0482, 0x047d, 0x0478, 0x0473, 0x046e, 0x0469, 0x0465, 0x0460, 0x045b, 0x0456, 0x0452, 0x044d, 0x0448, 0x0444,
-    0x043f, 0x043b, 0x0436, 0x0432, 0x042d, 0x0429, 0x0425, 0x0421, 0x041c, 0x0418, 0x0414, 0x0410, 0x040c, 0x0408, 0x0404, 0x0400,
-    0x03fc, 0x03f8, 0x03f4, 0x03f0, 0x03ec, 0x03e8, 0x03e4, 0x03e0, 0x03dd, 0x03d9, 0x03d5, 0x03d2, 0x03ce, 0x03ca, 0x03c7, 0x03c3,
-    0x03c0, 0x03bc, 0x03b9, 0x03b5, 0x03b2, 0x03ae, 0x03ab, 0x03a8, 0x03a4, 0x03a1, 0x039e, 0x039b, 0x0397, 0x0394, 0x0391, 0x038e,
-    0x038b, 0x0387, 0x0384, 0x0381, 0x037e, 0x037b, 0x0378, 0x0375, 0x0372, 0x036f, 0x036c, 0x0369, 0x0366, 0x0364, 0x0361, 0x035e,
-    0x035b, 0x0358, 0x0355, 0x0353, 0x0350, 0x034d, 0x034a, 0x0348, 0x0345, 0x0342, 0x0340, 0x033d, 0x033a, 0x0338, 0x0335, 0x0333,
-    0x0330, 0x032e, 0x032b, 0x0329, 0x0326, 0x0324, 0x0321, 0x031f, 0x031c, 0x031a, 0x0317, 0x0315, 0x0313, 0x0310, 0x030e, 0x030c,
-    0x0309, 0x0307, 0x0305, 0x0303, 0x0300, 0x02fe, 0x02fc, 0x02fa, 0x02f7, 0x02f5, 0x02f3, 0x02f1, 0x02ef, 0x02ec, 0x02ea, 0x02e8,
-    0x02e6, 0x02e4, 0x02e2, 0x02e0, 0x02de, 0x02dc, 0x02da, 0x02d8, 0x02d6, 0x02d4, 0x02d2, 0x02d0, 0x02ce, 0x02cc, 0x02ca, 0x02c8,
-    0x02c6, 0x02c4, 0x02c2, 0x02c0, 0x02be, 0x02bc, 0x02bb, 0x02b9, 0x02b7, 0x02b5, 0x02b3, 0x02b1, 0x02b0, 0x02ae, 0x02ac, 0x02aa,
-    0x02a8, 0x02a7, 0x02a5, 0x02a3, 0x02a1, 0x02a0, 0x029e, 0x029c, 0x029b, 0x0299, 0x0297, 0x0295, 0x0294, 0x0292, 0x0291, 0x028f,
-    0x028d, 0x028c, 0x028a, 0x0288, 0x0287, 0x0285, 0x0284, 0x0282, 0x0280, 0x027f, 0x027d, 0x027c, 0x027a, 0x0279, 0x0277, 0x0276,
-    0x0274, 0x0273, 0x0271, 0x0270, 0x026e, 0x026d, 0x026b, 0x026a, 0x0268, 0x0267, 0x0265, 0x0264, 0x0263, 0x0261, 0x0260, 0x025e,
-    0x025d, 0x025c, 0x025a, 0x0259, 0x0257, 0x0256, 0x0255, 0x0253, 0x0252, 0x0251, 0x024f, 0x024e, 0x024d, 0x024b, 0x024a, 0x0249,
-    0x0247, 0x0246, 0x0245, 0x0243, 0x0242, 0x0241, 0x0240, 0x023e, 0x023d, 0x023c, 0x023b, 0x0239, 0x0238, 0x0237, 0x0236, 0x0234,
-    0x0233, 0x0232, 0x0231, 0x0230, 0x022e, 0x022d, 0x022c, 0x022b, 0x022a, 0x0229, 0x0227, 0x0226, 0x0225, 0x0224, 0x0223, 0x0222,
-    0x0220, 0x021f, 0x021e, 0x021d, 0x021c, 0x021b, 0x021a, 0x0219, 0x0218, 0x0216, 0x0215, 0x0214, 0x0213, 0x0212, 0x0211, 0x0210,
-    0x020f, 0x020e, 0x020d, 0x020c, 0x020b, 0x020a, 0x0209, 0x0208, 0x0207, 0x0206, 0x0205, 0x0204, 0x0203, 0x0202, 0x0201, 0x0200,
-    0x01ff, 0x01fe, 0x01fd, 0x01fc, 0x01fb, 0x01fa, 0x01f9, 0x01f8, 0x01f7, 0x01f6, 0x01f5, 0x01f4, 0x01f3, 0x01f2, 0x01f1, 0x01f0,
-    0x01ef, 0x01ee, 0x01ed, 0x01ec, 0x01eb, 0x01ea, 0x01e9, 0x01e9, 0x01e8, 0x01e7, 0x01e6, 0x01e5, 0x01e4, 0x01e3, 0x01e2, 0x01e1,
-    0x01e0, 0x01e0, 0x01df, 0x01de, 0x01dd, 0x01dc, 0x01db, 0x01da, 0x01da, 0x01d9, 0x01d8, 0x01d7, 0x01d6, 0x01d5, 0x01d4, 0x01d4,
-    0x01d3, 0x01d2, 0x01d1, 0x01d0, 0x01cf, 0x01cf, 0x01ce, 0x01cd, 0x01cc, 0x01cb, 0x01cb, 0x01ca, 0x01c9, 0x01c8, 0x01c7, 0x01c7,
-    0x01c6, 0x01c5, 0x01c4, 0x01c3, 0x01c3, 0x01c2, 0x01c1, 0x01c0, 0x01c0, 0x01bf, 0x01be, 0x01bd, 0x01bd, 0x01bc, 0x01bb, 0x01ba,
-    0x01ba, 0x01b9, 0x01b8, 0x01b7, 0x01b7, 0x01b6, 0x01b5, 0x01b4, 0x01b4, 0x01b3, 0x01b2, 0x01b2, 0x01b1, 0x01b0, 0x01af, 0x01af,
-    0x01ae, 0x01ad, 0x01ad, 0x01ac, 0x01ab, 0x01aa, 0x01aa, 0x01a9, 0x01a8, 0x01a8, 0x01a7, 0x01a6, 0x01a6, 0x01a5, 0x01a4, 0x01a4,
-    0x01a3, 0x01a2, 0x01a2, 0x01a1, 0x01a0, 0x01a0, 0x019f, 0x019e, 0x019e, 0x019d, 0x019c, 0x019c, 0x019b, 0x019a, 0x019a, 0x0199,
-    0x0198, 0x0198, 0x0197, 0x0197, 0x0196, 0x0195, 0x0195, 0x0194, 0x0193, 0x0193, 0x0192, 0x0192, 0x0191, 0x0190, 0x0190, 0x018f,
-    0x018f, 0x018e, 0x018d, 0x018d, 0x018c, 0x018b, 0x018b, 0x018a, 0x018a, 0x0189, 0x0189, 0x0188, 0x0187, 0x0187, 0x0186, 0x0186,
-    0x0185, 0x0184, 0x0184, 0x0183, 0x0183, 0x0182, 0x0182, 0x0181, 0x0180, 0x0180, 0x017f, 0x017f, 0x017e, 0x017e, 0x017d, 0x017d,
-    0x017c, 0x017b, 0x017b, 0x017a, 0x017a, 0x0179, 0x0179, 0x0178, 0x0178, 0x0177, 0x0177, 0x0176, 0x0175, 0x0175, 0x0174, 0x0174,
-    0x0173, 0x0173, 0x0172, 0x0172, 0x0171, 0x0171, 0x0170, 0x0170, 0x016f, 0x016f, 0x016e, 0x016e, 0x016d, 0x016d, 0x016c, 0x016c,
-    0x016b, 0x016b, 0x016a, 0x016a, 0x0169, 0x0169, 0x0168, 0x0168, 0x0167, 0x0167, 0x0166, 0x0166, 0x0165, 0x0165, 0x0164, 0x0164,
-    0x0163, 0x0163, 0x0162, 0x0162, 0x0161, 0x0161, 0x0160, 0x0160, 0x015f, 0x015f, 0x015e, 0x015e, 0x015d, 0x015d, 0x015d, 0x015c,
-    0x015c, 0x015b, 0x015b, 0x015a, 0x015a, 0x0159, 0x0159, 0x0158, 0x0158, 0x0158, 0x0157, 0x0157, 0x0156, 0x0156
-};
 static const uint16_t DivTableNEON[255*3+1] = {
     0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
     0x0000, 0x1c71, 0x1af2, 0x1999, 0x1861, 0x1745, 0x1642, 0x1555, 0x147a, 0x13b1, 0x12f6, 0x1249, 0x11a7, 0x1111, 0x1084, 0x1000,
@@ -512,6 +462,36 @@ static tracy_force_inline void ProcessRGB_AVX( const uint8_t* src, char*& dst )
     __m256i px2 = _mm256_loadu_si256(((__m256i*)src) + 2);
     __m256i px3 = _mm256_loadu_si256(((__m256i*)src) + 3);
 
+    __m256i smask = _mm256_set1_epi32( 0xF8FCF8 );
+    __m256i sd0 = _mm256_and_si256( px0, smask );
+    __m256i sd1 = _mm256_and_si256( px1, smask );
+    __m256i sd2 = _mm256_and_si256( px2, smask );
+    __m256i sd3 = _mm256_and_si256( px3, smask );
+
+    __m256i sc = _mm256_shuffle_epi32(sd0, _MM_SHUFFLE(0, 0, 0, 0));
+
+    __m256i sc0 = _mm256_cmpeq_epi8( sd0, sc );
+    __m256i sc1 = _mm256_cmpeq_epi8( sd1, sc );
+    __m256i sc2 = _mm256_cmpeq_epi8( sd2, sc );
+    __m256i sc3 = _mm256_cmpeq_epi8( sd3, sc );
+
+    __m256i sm0 = _mm256_and_si256( sc0, sc1 );
+    __m256i sm1 = _mm256_and_si256( sc2, sc3 );
+    __m256i sm = _mm256_and_si256( sm0, sm1 );
+
+    const int64_t solid0 = 1 - _mm_testc_si128( _mm256_castsi256_si128( sm ), _mm_set1_epi32( -1 ) );
+    const int64_t solid1 = 1 - _mm_testc_si128( _mm256_extracti128_si256( sm, 1 ), _mm_set1_epi32( -1 ) );
+
+    if( !solid0 && !solid1 )
+    {
+        const auto c0 = uint64_t( to565( src[0], src[1], src[2] ) ) << 16;
+        const auto c1 = uint64_t( to565( src[16], src[17], src[18] ) ) << 16;
+        memcpy( dst, &c0, 8 );
+        memcpy( dst+8, &c1, 8 );
+        dst += 16;
+        return;
+    }
+
     __m256i min0 = _mm256_min_epu8( px0, px1 );
     __m256i min1 = _mm256_min_epu8( px2, px3 );
     __m256i min2 = _mm256_min_epu8( min0, min1 );
@@ -533,8 +513,8 @@ static tracy_force_inline void ProcessRGB_AVX( const uint8_t* src, char*& dst )
     __m256i range1 = _mm256_subs_epu8( rmax, rmin );
     __m256i range2 = _mm256_sad_epu8( rmax, rmin );
 
-    uint16_t vrange0 = DivTableAVX[_mm256_cvtsi256_si32( range2 ) >> 1];
+    uint16_t vrange0 = DivTable[_mm256_cvtsi256_si32( range2 ) >> 1];
-    uint16_t vrange1 = DivTableAVX[_mm256_extract_epi16( range2, 8 ) >> 1];
+    uint16_t vrange1 = DivTable[_mm256_extract_epi16( range2, 8 ) >> 1];
     __m256i range00 = _mm256_set1_epi16( vrange0 );
     __m256i range = _mm256_inserti128_si256( range00, _mm_set1_epi16( vrange1 ), 1 );
 
@@ -578,7 +558,11 @@ static tracy_force_inline void ProcessRGB_AVX( const uint8_t* src, char*& dst )
 
     __m256i d0 = _mm256_unpacklo_epi32( mm5, p );
     __m256i d1 = _mm256_permute4x64_epi64( d0, _MM_SHUFFLE( 3, 2, 2, 0 ) );
-    _mm_storeu_si128( (__m128i*)dst, _mm256_castsi256_si128( d1 ) );
+    __m128i d2 = _mm256_castsi256_si128( d1 );
+
+    __m128i mask = _mm_set_epi64x( 0xFFFF0000 | -solid1, 0xFFFF0000 | -solid0 );
+    __m128i d3 = _mm_and_si128( d2, mask );
+    _mm_storeu_si128( (__m128i*)dst, d3 );
     dst += 16;
 }
 #endif

manual/tracy.tex

@@ -689,7 +689,7 @@ To further reduce image data size, frame images are internally compressed using
 \textbf{Implementation} & \textbf{Required define} & \textbf{Time} \\ \hline
 x86 Reference & --- & 198.2 \si{\micro\second} \\
 x86 SSE4.1\textsuperscript{a} & \texttt{\_\_SSE4\_1\_\_} & 25.4 \si{\micro\second} \\
-x86 AVX2 & \texttt{\_\_AVX2\_\_} & 15.3 \si{\micro\second} \\
+x86 AVX2 & \texttt{\_\_AVX2\_\_} & 17.4 \si{\micro\second} \\
 ARM Reference & --- & 1.04 \si{\milli\second} \\
 ARM32 NEON\textsuperscript{b} & \texttt{\_\_ARM\_NEON} & 529 \si{\micro\second} \\
 ARM64 NEON & \texttt{\_\_ARM\_NEON} & 438 \si{\micro\second}