Showing the changes to test/encoder/EncUT_Sample.cpp (1 of 85 files). Show all
Unified · Split
| @@ -11,134 +11,138 @@ · 10 unmodified lines | |||
| 11 | 11 | ||
| 12 | using namespace WelsSVCEnc; | 12 | using namespace WelsSVCEnc; |
| 13 | #ifdef X86_ASM | 13 | #ifdef X86_ASM |
| 14 | TEST(IntraSadSatdFuncTest, WelsIntra16x16Combined3Sad_ssse3){ | 14 | TEST (IntraSadSatdFuncTest, WelsIntra16x16Combined3Sad_ssse3) { |
| 15 | const int32_t iLineSizeDec = 32; | 15 | const int32_t iLineSizeDec = 32; |
| 16 | const int32_t iLineSizeEnc = 32; | 16 | const int32_t iLineSizeEnc = 32; |
| 17 | int32_t tmpa, tmpb; | 17 | int32_t tmpa, tmpb; |
| 18 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; | 18 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; |
| 19 | CMemoryAlign cMemoryAlign(0); | 19 | CMemoryAlign cMemoryAlign (0); |
| 20 | int32_t iCpuCores = 0; | 20 | int32_t iCpuCores = 0; |
| 21 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect(&iCpuCores); | 21 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect (&iCpuCores); |
| 22 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSSE3)) | 22 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSSE3)) |
| 23 | return; | 23 | return; |
| 24 | uint8_t* pDec = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeDec<<5,"pDec"); | 24 | uint8_t* pDec = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeDec << 5, "pDec"); |
| 25 | uint8_t* pEnc = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeEnc<<5,"pEnc"); | 25 | uint8_t* pEnc = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeEnc << 5, "pEnc"); |
| 26 | uint8_t* pDst = (uint8_t *)cMemoryAlign.WelsMalloc(512,"pDst"); | 26 | uint8_t* pDst = (uint8_t*)cMemoryAlign.WelsMalloc (512, "pDst"); |
| 27 | srand((uint32_t)time(NULL)); | 27 | srand ((uint32_t)time (NULL)); |
| 28 | for(int i=0;i<(iLineSizeDec<<5);i++) | 28 | for (int i = 0; i < (iLineSizeDec << 5); i++) |
| 29 | pDec[i]=rand()%256; | 29 | pDec[i] = rand() % 256; |
| 30 | for(int i=0;i<(iLineSizeEnc<<5);i++) | 30 | for (int i = 0; i < (iLineSizeEnc << 5); i++) |
| 31 | pEnc[i]=rand()%256; | 31 | pEnc[i] = rand() % 256; |
| 32 | 32 | ||
| 33 | for(int i=0;i<512;i++) | 33 | for (int i = 0; i < 512; i++) |
| 34 | pDst[i]=rand()%256; | 34 | pDst[i] = rand() % 256; |
| 35 | tmpa = WelsSampleSadIntra16x16Combined3_c(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc,&iBestMode_c, iLambda, pDst); | 35 | tmpa = WelsSampleSadIntra16x16Combined3_c (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, &iBestMode_c, iLambda, pDst); |
| 36 | tmpb = WelsIntra16x16Combined3Sad_ssse3(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc,&iBestMode_a, iLambda, pDst); | 36 | tmpb = WelsIntra16x16Combined3Sad_ssse3 (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, &iBestMode_a, iLambda, pDst); |
| 37 | 37 | ||
| 38 | ASSERT_EQ(tmpa, tmpb); | 38 | ASSERT_EQ (tmpa, tmpb); |
| 39 | ASSERT_EQ(iBestMode_c, iBestMode_a); | 39 | ASSERT_EQ (iBestMode_c, iBestMode_a); |
| 40 | 40 | ||
| 41 | cMemoryAlign.WelsFree(pDec,"pDec"); | 41 | cMemoryAlign.WelsFree (pDec, "pDec"); |
| 42 | cMemoryAlign.WelsFree(pEnc,"pEnc"); | 42 | cMemoryAlign.WelsFree (pEnc, "pEnc"); |
| 43 | cMemoryAlign.WelsFree(pDst,"pDst"); | 43 | cMemoryAlign.WelsFree (pDst, "pDst"); |
| 44 | } | 44 | } |
| 45 | 45 | ||
| 46 | TEST(IntraSadSatdFuncTest, WelsIntra16x16Combined3Satd_sse41){ | 46 | TEST (IntraSadSatdFuncTest, WelsIntra16x16Combined3Satd_sse41) { |
| 47 | const int32_t iLineSizeDec = 32; | 47 | const int32_t iLineSizeDec = 32; |
| 48 | const int32_t iLineSizeEnc = 32; | 48 | const int32_t iLineSizeEnc = 32; |
| 49 | int32_t tmpa, tmpb; | 49 | int32_t tmpa, tmpb; |
| 50 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; | 50 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; |
| 51 | CMemoryAlign cMemoryAlign(0); | 51 | CMemoryAlign cMemoryAlign (0); |
| 52 | int32_t iCpuCores = 0; | 52 | int32_t iCpuCores = 0; |
| 53 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect(&iCpuCores); | 53 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect (&iCpuCores); |
| 54 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 54 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 55 | return; | 55 | return; |
| 56 | uint8_t* pDec = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeDec<<5,"pDec"); | 56 | uint8_t* pDec = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeDec << 5, "pDec"); |
| 57 | uint8_t* pEnc = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeEnc<<5,"pEnc"); | 57 | uint8_t* pEnc = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeEnc << 5, "pEnc"); |
| 58 | uint8_t* pDst = (uint8_t *)cMemoryAlign.WelsMalloc(512,"pDst"); | 58 | uint8_t* pDst = (uint8_t*)cMemoryAlign.WelsMalloc (512, "pDst"); |
| 59 | srand((uint32_t)time(NULL)); | 59 | srand ((uint32_t)time (NULL)); |
| 60 | for(int i=0;i<(iLineSizeDec<<5);i++) | 60 | for (int i = 0; i < (iLineSizeDec << 5); i++) |
| 61 | pDec[i]=rand()%256; | 61 | pDec[i] = rand() % 256; |
| 62 | for(int i=0;i<(iLineSizeEnc<<5);i++) | 62 | for (int i = 0; i < (iLineSizeEnc << 5); i++) |
| 63 | pEnc[i]=rand()%256; | 63 | pEnc[i] = rand() % 256; |
| 64 | for(int i=0;i<512;i++) | 64 | for (int i = 0; i < 512; i++) |
| 65 | pDst[i]=rand()%256; | 65 | pDst[i] = rand() % 256; |
| 66 | tmpa = WelsSampleSatdIntra16x16Combined3_c(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc,&iBestMode_c, iLambda, pDst); | 66 | tmpa = WelsSampleSatdIntra16x16Combined3_c (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, &iBestMode_c, iLambda, pDst); |
| 67 | tmpb = WelsIntra16x16Combined3Satd_sse41(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc,&iBestMode_a, iLambda, pDst); | 67 | tmpb = WelsIntra16x16Combined3Satd_sse41 (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, &iBestMode_a, iLambda, pDst); |
| 68 | ASSERT_EQ(tmpa, tmpb); | 68 | ASSERT_EQ (tmpa, tmpb); |
| 69 | ASSERT_EQ(iBestMode_c, iBestMode_a); | 69 | ASSERT_EQ (iBestMode_c, iBestMode_a); |
| 70 | cMemoryAlign.WelsFree(pDec,"pDec"); | 70 | cMemoryAlign.WelsFree (pDec, "pDec"); |
| 71 | cMemoryAlign.WelsFree(pEnc,"pEnc"); | 71 | cMemoryAlign.WelsFree (pEnc, "pEnc"); |
| 72 | cMemoryAlign.WelsFree(pDst,"pDst"); | 72 | cMemoryAlign.WelsFree (pDst, "pDst"); |
| 73 | } | 73 | } |
| 74 | 74 | ||
| 75 | TEST(IntraSadSatdFuncTest, WelsSampleSatdThree4x4_sse2){ | 75 | TEST (IntraSadSatdFuncTest, WelsSampleSatdThree4x4_sse2) { |
| 76 | const int32_t iLineSizeDec = 32; | 76 | const int32_t iLineSizeDec = 32; |
| 77 | const int32_t iLineSizeEnc = 32; | 77 | const int32_t iLineSizeEnc = 32; |
| 78 | int32_t tmpa, tmpb; | 78 | int32_t tmpa, tmpb; |
| 79 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; | 79 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; |
| 80 | int32_t lambda[2] = {iLambda << 2, iLambda}; | 80 | int32_t lambda[2] = {iLambda << 2, iLambda}; |
| 81 | int32_t iPredMode = rand()%3; | 81 | int32_t iPredMode = rand() % 3; |
| 82 | CMemoryAlign cMemoryAlign(0); | 82 | CMemoryAlign cMemoryAlign (0); |
| 83 | int32_t iCpuCores = 0; | 83 | int32_t iCpuCores = 0; |
| 84 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect(&iCpuCores); | 84 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect (&iCpuCores); |
| 85 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 85 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 86 | return; | 86 | return; |
| 87 | uint8_t* pDec = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeDec<<5,"pDec"); | 87 | uint8_t* pDec = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeDec << 5, "pDec"); |
| 88 | uint8_t* pEnc = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeEnc<<5,"pEnc"); | 88 | uint8_t* pEnc = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeEnc << 5, "pEnc"); |
| 89 | uint8_t* pDst = (uint8_t *)cMemoryAlign.WelsMalloc(512,"pDst"); | 89 | uint8_t* pDst = (uint8_t*)cMemoryAlign.WelsMalloc (512, "pDst"); |
| 90 | WelsInitFillingPredFuncs(WELS_CPU_SSE2); | 90 | WelsInitFillingPredFuncs (WELS_CPU_SSE2); |
| 91 | srand((uint32_t)time(NULL)); | 91 | srand ((uint32_t)time (NULL)); |
| 92 | for(int i=0;i<(iLineSizeDec<<5);i++) | 92 | for (int i = 0; i < (iLineSizeDec << 5); i++) |
| 93 | pDec[i]=rand()%256; | 93 | pDec[i] = rand() % 256; |
| 94 | for(int i=0;i<(iLineSizeEnc<<5);i++) | 94 | for (int i = 0; i < (iLineSizeEnc << 5); i++) |
| 95 | pEnc[i]=rand()%256; | 95 | pEnc[i] = rand() % 256; |
| 96 | for(int i=0;i<512;i++) | 96 | for (int i = 0; i < 512; i++) |
| 97 | pDst[i]=rand()%256; | 97 | pDst[i] = rand() % 256; |
| 98 | tmpa = WelsSampleSatdIntra4x4Combined3_c(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc, pDst, &iBestMode_c, lambda[iPredMode == 2], lambda[iPredMode == 1], lambda[iPredMode == 0]); | 98 | tmpa = WelsSampleSatdIntra4x4Combined3_c (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, pDst, &iBestMode_c, |
| 99 | tmpb = WelsSampleSatdThree4x4_sse2(pDec+128, iLineSizeDec, pEnc,iLineSizeEnc, pDst, &iBestMode_a, lambda[iPredMode == 2], lambda[iPredMode == 1], lambda[iPredMode == 0]); | 99 | lambda[iPredMode == 2], lambda[iPredMode == 1], lambda[iPredMode == 0]); |
| 100 | ASSERT_EQ(tmpa, tmpb); | 100 | tmpb = WelsSampleSatdThree4x4_sse2 (pDec + 128, iLineSizeDec, pEnc, iLineSizeEnc, pDst, &iBestMode_a, |
| 101 | ASSERT_EQ(iBestMode_c, iBestMode_a); | 101 | lambda[iPredMode == 2], lambda[iPredMode == 1], lambda[iPredMode == 0]); |
| 102 | cMemoryAlign.WelsFree(pDec,"pDec"); | 102 | ASSERT_EQ (tmpa, tmpb); |
| 103 | cMemoryAlign.WelsFree(pEnc,"pEnc"); | 103 | ASSERT_EQ (iBestMode_c, iBestMode_a); |
| 104 | cMemoryAlign.WelsFree(pDst,"pDst"); | 104 | cMemoryAlign.WelsFree (pDec, "pDec"); |
| 105 | } | 105 | cMemoryAlign.WelsFree (pEnc, "pEnc"); |
| 106 | 106 | cMemoryAlign.WelsFree (pDst, "pDst"); | |
| 107 | TEST(IntraSadSatdFuncTest, WelsIntraChroma8x8Combined3Satd_sse41){ | 107 | } |
| 108 | |||
| 109 | TEST (IntraSadSatdFuncTest, WelsIntraChroma8x8Combined3Satd_sse41) { | ||
| 108 | const int32_t iLineSizeDec = 32; | 110 | const int32_t iLineSizeDec = 32; |
| 109 | const int32_t iLineSizeEnc = 32; | 111 | const int32_t iLineSizeEnc = 32; |
| 110 | int32_t tmpa, tmpb; | 112 | int32_t tmpa, tmpb; |
| 111 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; | 113 | int32_t iBestMode_c, iBestMode_a, iLambda = 50; |
| 112 | CMemoryAlign cMemoryAlign(0); | 114 | CMemoryAlign cMemoryAlign (0); |
| 113 | int32_t iCpuCores = 0; | 115 | int32_t iCpuCores = 0; |
| 114 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect(&iCpuCores); | 116 | uint32_t m_uiCpuFeatureFlag = WelsCPUFeatureDetect (&iCpuCores); |
| 115 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 117 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 116 | return; | 118 | return; |
| 117 | uint8_t* pDecCb = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeDec<<5,"pDecCb"); | 119 | uint8_t* pDecCb = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeDec << 5, "pDecCb"); |
| 118 | uint8_t* pEncCb = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeEnc<<5,"pEncCb"); | 120 | uint8_t* pEncCb = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeEnc << 5, "pEncCb"); |
| 119 | uint8_t* pDecCr = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeDec<<5,"pDecCr"); | 121 | uint8_t* pDecCr = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeDec << 5, "pDecCr"); |
| 120 | uint8_t* pEncCr = (uint8_t *)cMemoryAlign.WelsMalloc(iLineSizeEnc<<5,"pEncCr"); | 122 | uint8_t* pEncCr = (uint8_t*)cMemoryAlign.WelsMalloc (iLineSizeEnc << 5, "pEncCr"); |
| 121 | uint8_t* pDstChma = (uint8_t *)cMemoryAlign.WelsMalloc(512,"pDstChma"); | 123 | uint8_t* pDstChma = (uint8_t*)cMemoryAlign.WelsMalloc (512, "pDstChma"); |
| 122 | srand((uint32_t)time(NULL)); | 124 | srand ((uint32_t)time (NULL)); |
| 123 | for(int i=0;i<(iLineSizeDec<<5);i++){ | 125 | for (int i = 0; i < (iLineSizeDec << 5); i++) { |
| 124 | pDecCb[i]=rand()%256; | 126 | pDecCb[i] = rand() % 256; |
| 125 | pDecCr[i]=rand()%256; | 127 | pDecCr[i] = rand() % 256; |
| 126 | } | 128 | } |
| 127 | for(int i=0;i<(iLineSizeEnc<<5);i++){ | 129 | for (int i = 0; i < (iLineSizeEnc << 5); i++) { |
| 128 | pEncCb[i]=rand()%256; | 130 | pEncCb[i] = rand() % 256; |
| 129 | pEncCr[i]=rand()%256; | 131 | pEncCr[i] = rand() % 256; |
| 130 | } | 132 | } |
| 131 | for(int i=0;i<512;i++) | 133 | for (int i = 0; i < 512; i++) |
| 132 | pDstChma[i]=rand()%256; | 134 | pDstChma[i] = rand() % 256; |
| 133 | tmpa = WelsSampleSatdIntra8x8Combined3_c(pDecCb+128, iLineSizeDec, pEncCb,iLineSizeEnc,&iBestMode_c, iLambda, pDstChma, pDecCr+128, pEncCr); | 135 | tmpa = WelsSampleSatdIntra8x8Combined3_c (pDecCb + 128, iLineSizeDec, pEncCb, iLineSizeEnc, &iBestMode_c, iLambda, |
| 134 | tmpb = WelsIntraChroma8x8Combined3Satd_sse41(pDecCb+128, iLineSizeDec, pEncCb,iLineSizeEnc,&iBestMode_a, iLambda, pDstChma, pDecCr+128, pEncCr); | 136 | pDstChma, pDecCr + 128, pEncCr); |
| 135 | ASSERT_EQ(tmpa, tmpb); | 137 | tmpb = WelsIntraChroma8x8Combined3Satd_sse41 (pDecCb + 128, iLineSizeDec, pEncCb, iLineSizeEnc, &iBestMode_a, iLambda, |
| 136 | ASSERT_EQ(iBestMode_c, iBestMode_a); | 138 | pDstChma, pDecCr + 128, pEncCr); |
| 137 | cMemoryAlign.WelsFree(pDecCb,"pDecCb"); | 139 | ASSERT_EQ (tmpa, tmpb); |
| 138 | cMemoryAlign.WelsFree(pEncCb,"pEncCb"); | 140 | ASSERT_EQ (iBestMode_c, iBestMode_a); |
| 139 | cMemoryAlign.WelsFree(pDecCr,"pDecCr"); | 141 | cMemoryAlign.WelsFree (pDecCb, "pDecCb"); |
| 140 | cMemoryAlign.WelsFree(pEncCr,"pEncCr"); | 142 | cMemoryAlign.WelsFree (pEncCb, "pEncCb"); |
| 141 | cMemoryAlign.WelsFree(pDstChma,"pDstChma"); | 143 | cMemoryAlign.WelsFree (pDecCr, "pDecCr"); |
| 144 | cMemoryAlign.WelsFree (pEncCr, "pEncCr"); | ||
| 145 | cMemoryAlign.WelsFree (pDstChma, "pDstChma"); | ||
| 142 | } | 146 | } |
| 143 | #endif | 147 | #endif |
| 144 | #define ASSERT_MEMORY_FAIL2X(A, B) \ | 148 | #define ASSERT_MEMORY_FAIL2X(A, B) \ |
| @@ -157,27 +161,27 @@ TEST(IntraSadSatdFuncTest, WelsIntraChroma8x8Combined3Satd_sse41){ · 12 unmodified lines | |||
| 157 | #define PIXEL_STRIDE 32 | 161 | #define PIXEL_STRIDE 32 |
| 158 | 162 | ||
| 159 | class SadSatdCFuncTest : public testing::Test { | 163 | class SadSatdCFuncTest : public testing::Test { |
| 160 | public: | 164 | public: |
| 161 | virtual void SetUp() { | 165 | virtual void SetUp() { |
| 162 | pMemAlign = new CMemoryAlign(0); | 166 | pMemAlign = new CMemoryAlign (0); |
| 163 | 167 | ||
| 164 | srand((uint32_t)time(NULL)); | 168 | srand ((uint32_t)time (NULL)); |
| 165 | m_iStrideA = rand()%256 + PIXEL_STRIDE; | 169 | m_iStrideA = rand() % 256 + PIXEL_STRIDE; |
| 166 | m_iStrideB = rand()%256 + PIXEL_STRIDE; | 170 | m_iStrideB = rand() % 256 + PIXEL_STRIDE; |
| 167 | m_pPixSrcA = (uint8_t *)pMemAlign->WelsMalloc(m_iStrideA<<5,"Sad_m_pPixSrcA"); | 171 | m_pPixSrcA = (uint8_t*)pMemAlign->WelsMalloc (m_iStrideA << 5, "Sad_m_pPixSrcA"); |
| 168 | ASSERT_TRUE(NULL != m_pPixSrcA); | 172 | ASSERT_TRUE (NULL != m_pPixSrcA); |
| 169 | m_pPixSrcB = (uint8_t *)pMemAlign->WelsMalloc(m_iStrideB<<5,"Sad_m_pPixSrcB"); | 173 | m_pPixSrcB = (uint8_t*)pMemAlign->WelsMalloc (m_iStrideB << 5, "Sad_m_pPixSrcB"); |
| 170 | ASSERT_MEMORY_FAIL2X(m_pPixSrcA, m_pPixSrcB) | 174 | ASSERT_MEMORY_FAIL2X (m_pPixSrcA, m_pPixSrcB) |
| 171 | m_pSad = (int32_t*)pMemAlign->WelsMalloc(4*sizeof(int32_t), "m_pSad"); | 175 | m_pSad = (int32_t*)pMemAlign->WelsMalloc (4 * sizeof (int32_t), "m_pSad"); |
| 172 | ASSERT_MEMORY_FAIL3X(m_pPixSrcA, m_pPixSrcB, m_pSad) | 176 | ASSERT_MEMORY_FAIL3X (m_pPixSrcA, m_pPixSrcB, m_pSad) |
| 173 | } | 177 | } |
| 174 | virtual void TearDown() { | 178 | virtual void TearDown() { |
| 175 | pMemAlign->WelsFree(m_pPixSrcA,"Sad_m_pPixSrcA"); | 179 | pMemAlign->WelsFree (m_pPixSrcA, "Sad_m_pPixSrcA"); |
| 176 | pMemAlign->WelsFree(m_pPixSrcB,"Sad_m_pPixSrcB"); | 180 | pMemAlign->WelsFree (m_pPixSrcB, "Sad_m_pPixSrcB"); |
| 177 | pMemAlign->WelsFree(m_pSad,"m_pSad"); | 181 | pMemAlign->WelsFree (m_pSad, "m_pSad"); |
| 178 | delete pMemAlign; | 182 | delete pMemAlign; |
| 179 | } | 183 | } |
| 180 | public: | 184 | public: |
| 181 | uint8_t* m_pPixSrcA; | 185 | uint8_t* m_pPixSrcA; |
| 182 | uint8_t* m_pPixSrcB; | 186 | uint8_t* m_pPixSrcB; |
| 183 | int32_t m_iStrideA; | 187 | int32_t m_iStrideA; |
| @@ -187,313 +191,313 @@ public: · 3 unmodified lines | |||
| 187 | CMemoryAlign* pMemAlign; | 191 | CMemoryAlign* pMemAlign; |
| 188 | }; | 192 | }; |
| 189 | 193 | ||
| 190 | TEST_F(SadSatdCFuncTest, WelsSampleSad4x4_c) { | 194 | TEST_F (SadSatdCFuncTest, WelsSampleSad4x4_c) { |
| 191 | srand((uint32_t)time(NULL)); | 195 | srand ((uint32_t)time (NULL)); |
| 192 | for(int i=0; i<(m_iStrideA<<2); i++) | 196 | for (int i = 0; i < (m_iStrideA << 2); i++) |
| 193 | m_pPixSrcA[i]=rand()%256; | 197 | m_pPixSrcA[i] = rand() % 256; |
| 194 | for(int i=0; i<(m_iStrideB<<2); i++) | 198 | for (int i = 0; i < (m_iStrideB << 2); i++) |
| 195 | m_pPixSrcB[i]=rand()%256; | 199 | m_pPixSrcB[i] = rand() % 256; |
| 196 | uint8_t *pPixA=m_pPixSrcA; | 200 | uint8_t* pPixA = m_pPixSrcA; |
| 197 | uint8_t *pPixB=m_pPixSrcB; | 201 | uint8_t* pPixB = m_pPixSrcB; |
| 198 | 202 | ||
| 199 | int32_t iSumSad = 0; | 203 | int32_t iSumSad = 0; |
| 200 | for (int i = 0; i < 4; i++ ) { | 204 | for (int i = 0; i < 4; i++) { |
| 201 | for(int j=0; j<4; j++) | 205 | for (int j = 0; j < 4; j++) |
| 202 | iSumSad+=abs(pPixA[j]-pPixB[j]); | 206 | iSumSad += abs (pPixA[j] - pPixB[j]); |
| 203 | pPixA += m_iStrideA; | 207 | pPixA += m_iStrideA; |
| 204 | pPixB += m_iStrideB; | 208 | pPixB += m_iStrideB; |
| 205 | } | 209 | } |
| 206 | EXPECT_EQ(WelsSampleSad4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); | 210 | EXPECT_EQ (WelsSampleSad4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); |
| 207 | } | 211 | } |
| 208 | 212 | ||
| 209 | TEST_F(SadSatdCFuncTest, WelsSampleSad8x8_c) { | 213 | TEST_F (SadSatdCFuncTest, WelsSampleSad8x8_c) { |
| 210 | srand((uint32_t)time(NULL)); | 214 | srand ((uint32_t)time (NULL)); |
| 211 | for(int i=0; i<(m_iStrideA<<3); i++) | 215 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 212 | m_pPixSrcA[i]=rand()%256; | 216 | m_pPixSrcA[i] = rand() % 256; |
| 213 | for(int i=0; i<(m_iStrideB<<3); i++) | 217 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 214 | m_pPixSrcB[i]=rand()%256; | 218 | m_pPixSrcB[i] = rand() % 256; |
| 215 | uint8_t *pPixA=m_pPixSrcA; | 219 | uint8_t* pPixA = m_pPixSrcA; |
| 216 | uint8_t *pPixB=m_pPixSrcB; | 220 | uint8_t* pPixB = m_pPixSrcB; |
| 217 | 221 | ||
| 218 | int32_t iSumSad = 0; | 222 | int32_t iSumSad = 0; |
| 219 | for (int i = 0; i < 8; i++ ) { | 223 | for (int i = 0; i < 8; i++) { |
| 220 | for(int j=0; j<8; j++) | 224 | for (int j = 0; j < 8; j++) |
| 221 | iSumSad+=abs(pPixA[j]-pPixB[j]); | 225 | iSumSad += abs (pPixA[j] - pPixB[j]); |
| 222 | 226 | ||
| 223 | pPixA += m_iStrideA; | 227 | pPixA += m_iStrideA; |
| 224 | pPixB += m_iStrideB; | 228 | pPixB += m_iStrideB; |
| 225 | } | 229 | } |
| 226 | EXPECT_EQ(WelsSampleSad8x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); | 230 | EXPECT_EQ (WelsSampleSad8x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); |
| 227 | } | 231 | } |
| 228 | 232 | ||
| 229 | TEST_F(SadSatdCFuncTest, WelsSampleSad16x8_c) { | 233 | TEST_F (SadSatdCFuncTest, WelsSampleSad16x8_c) { |
| 230 | srand((uint32_t)time(NULL)); | 234 | srand ((uint32_t)time (NULL)); |
| 231 | for(int i=0; i<(m_iStrideA<<3); i++) | 235 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 232 | m_pPixSrcA[i]=rand()%256; | 236 | m_pPixSrcA[i] = rand() % 256; |
| 233 | for(int i=0; i<(m_iStrideB<<3); i++) | 237 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 234 | m_pPixSrcB[i]=rand()%256; | 238 | m_pPixSrcB[i] = rand() % 256; |
| 235 | uint8_t *pPixA=m_pPixSrcA; | 239 | uint8_t* pPixA = m_pPixSrcA; |
| 236 | uint8_t *pPixB=m_pPixSrcB; | 240 | uint8_t* pPixB = m_pPixSrcB; |
| 237 | 241 | ||
| 238 | int32_t iSumSad = 0; | 242 | int32_t iSumSad = 0; |
| 239 | for (int i = 0; i <8; i++ ) { | 243 | for (int i = 0; i < 8; i++) { |
| 240 | for(int j=0; j<16; j++) | 244 | for (int j = 0; j < 16; j++) |
| 241 | iSumSad+=abs(pPixA[j]-pPixB[j]); | 245 | iSumSad += abs (pPixA[j] - pPixB[j]); |
| 242 | 246 | ||
| 243 | pPixA += m_iStrideA; | 247 | pPixA += m_iStrideA; |
| 244 | pPixB += m_iStrideB; | 248 | pPixB += m_iStrideB; |
| 245 | } | 249 | } |
| 246 | EXPECT_EQ(WelsSampleSad16x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); | 250 | EXPECT_EQ (WelsSampleSad16x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); |
| 247 | } | 251 | } |
| 248 | 252 | ||
| 249 | TEST_F(SadSatdCFuncTest, WelsSampleSad8x16_c) { | 253 | TEST_F (SadSatdCFuncTest, WelsSampleSad8x16_c) { |
| 250 | srand((uint32_t)time(NULL)); | 254 | srand ((uint32_t)time (NULL)); |
| 251 | for(int i=0; i<(m_iStrideA<<4); i++) | 255 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 252 | m_pPixSrcA[i]=rand()%256; | 256 | m_pPixSrcA[i] = rand() % 256; |
| 253 | for(int i=0; i<(m_iStrideB<<4); i++) | 257 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 254 | m_pPixSrcB[i]=rand()%256; | 258 | m_pPixSrcB[i] = rand() % 256; |
| 255 | uint8_t *pPixA=m_pPixSrcA; | 259 | uint8_t* pPixA = m_pPixSrcA; |
| 256 | uint8_t *pPixB=m_pPixSrcB; | 260 | uint8_t* pPixB = m_pPixSrcB; |
| 257 | 261 | ||
| 258 | int32_t iSumSad = 0; | 262 | int32_t iSumSad = 0; |
| 259 | for (int i = 0; i <16; i++ ) { | 263 | for (int i = 0; i < 16; i++) { |
| 260 | for(int j=0; j<8; j++) | 264 | for (int j = 0; j < 8; j++) |
| 261 | iSumSad+=abs(pPixA[j]-pPixB[j]); | 265 | iSumSad += abs (pPixA[j] - pPixB[j]); |
| 262 | 266 | ||
| 263 | pPixA += m_iStrideA; | 267 | pPixA += m_iStrideA; |
| 264 | pPixB += m_iStrideB; | 268 | pPixB += m_iStrideB; |
| 265 | } | 269 | } |
| 266 | EXPECT_EQ(WelsSampleSad8x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); | 270 | EXPECT_EQ (WelsSampleSad8x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); |
| 267 | } | 271 | } |
| 268 | 272 | ||
| 269 | TEST_F(SadSatdCFuncTest, WelsSampleSad16x16_c) { | 273 | TEST_F (SadSatdCFuncTest, WelsSampleSad16x16_c) { |
| 270 | srand((uint32_t)time(NULL)); | 274 | srand ((uint32_t)time (NULL)); |
| 271 | for(int i=0; i<(m_iStrideA<<4); i++) | 275 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 272 | m_pPixSrcA[i]=rand()%256; | 276 | m_pPixSrcA[i] = rand() % 256; |
| 273 | for(int i=0; i<(m_iStrideB<<4); i++) | 277 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 274 | m_pPixSrcB[i]=rand()%256; | 278 | m_pPixSrcB[i] = rand() % 256; |
| 275 | uint8_t *pPixA=m_pPixSrcA; | 279 | uint8_t* pPixA = m_pPixSrcA; |
| 276 | uint8_t *pPixB=m_pPixSrcB; | 280 | uint8_t* pPixB = m_pPixSrcB; |
| 277 | 281 | ||
| 278 | int32_t iSumSad = 0; | 282 | int32_t iSumSad = 0; |
| 279 | for (int i = 0; i <16; i++ ) { | 283 | for (int i = 0; i < 16; i++) { |
| 280 | for(int j=0; j<16; j++) | 284 | for (int j = 0; j < 16; j++) |
| 281 | iSumSad+=abs(pPixA[j]-pPixB[j]); | 285 | iSumSad += abs (pPixA[j] - pPixB[j]); |
| 282 | 286 | ||
| 283 | pPixA += m_iStrideA; | 287 | pPixA += m_iStrideA; |
| 284 | pPixB += m_iStrideB; | 288 | pPixB += m_iStrideB; |
| 285 | } | 289 | } |
| 286 | EXPECT_EQ(WelsSampleSad16x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); | 290 | EXPECT_EQ (WelsSampleSad16x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), iSumSad); |
| 287 | } | 291 | } |
| 288 | 292 | ||
| 289 | TEST_F(SadSatdCFuncTest, WelsSampleSatd4x4_c) { | 293 | TEST_F (SadSatdCFuncTest, WelsSampleSatd4x4_c) { |
| 290 | srand((uint32_t)time(NULL)); | 294 | srand ((uint32_t)time (NULL)); |
| 291 | for(int i=0; i<(m_iStrideA<<2); i++) | 295 | for (int i = 0; i < (m_iStrideA << 2); i++) |
| 292 | m_pPixSrcA[i]=rand()%256; | 296 | m_pPixSrcA[i] = rand() % 256; |
| 293 | for(int i=0; i<(m_iStrideB<<2); i++) | 297 | for (int i = 0; i < (m_iStrideB << 2); i++) |
| 294 | m_pPixSrcB[i]=rand()%256; | 298 | m_pPixSrcB[i] = rand() % 256; |
| 295 | uint8_t *pPixA=m_pPixSrcA; | 299 | uint8_t* pPixA = m_pPixSrcA; |
| 296 | uint8_t *pPixB=m_pPixSrcB; | 300 | uint8_t* pPixB = m_pPixSrcB; |
| 297 | 301 | ||
| 298 | int32_t W[16],T[16],Y[16],k=0; | 302 | int32_t W[16], T[16], Y[16], k = 0; |
| 299 | for(int i=0; i<4; i++) { | 303 | for (int i = 0; i < 4; i++) { |
| 300 | for(int j=0; j<4; j++) | 304 | for (int j = 0; j < 4; j++) |
| 301 | W[k++]=pPixA[j]-pPixB[j]; | 305 | W[k++] = pPixA[j] - pPixB[j]; |
| 302 | pPixA += m_iStrideA; | 306 | pPixA += m_iStrideA; |
| 303 | pPixB += m_iStrideB; | 307 | pPixB += m_iStrideB; |
| 304 | } | 308 | } |
| 305 | 309 | ||
| 306 | T[0]=W[0]+W[4]+W[8]+W[12]; | 310 | T[0] = W[0] + W[4] + W[8] + W[12]; |
| 307 | T[1]=W[1]+W[5]+W[9]+W[13]; | 311 | T[1] = W[1] + W[5] + W[9] + W[13]; |
| 308 | T[2]=W[2]+W[6]+W[10]+W[14]; | 312 | T[2] = W[2] + W[6] + W[10] + W[14]; |
| 309 | T[3]=W[3]+W[7]+W[11]+W[15]; | 313 | T[3] = W[3] + W[7] + W[11] + W[15]; |
| 310 | 314 | ||
| 311 | T[4]=W[0]+W[4]-W[8]-W[12]; | 315 | T[4] = W[0] + W[4] - W[8] - W[12]; |
| 312 | T[5]=W[1]+W[5]-W[9]-W[13]; | 316 | T[5] = W[1] + W[5] - W[9] - W[13]; |
| 313 | T[6]=W[2]+W[6]-W[10]-W[14]; | 317 | T[6] = W[2] + W[6] - W[10] - W[14]; |
| 314 | T[7]=W[3]+W[7]-W[11]-W[15]; | 318 | T[7] = W[3] + W[7] - W[11] - W[15]; |
| 315 | 319 | ||
| 316 | T[8]=W[0]-W[4]-W[8]+W[12]; | 320 | T[8] = W[0] - W[4] - W[8] + W[12]; |
| 317 | T[9]=W[1]-W[5]-W[9]+W[13]; | 321 | T[9] = W[1] - W[5] - W[9] + W[13]; |
| 318 | T[10]=W[2]-W[6]-W[10]+W[14]; | 322 | T[10] = W[2] - W[6] - W[10] + W[14]; |
| 319 | T[11]=W[3]-W[7]-W[11]+W[15]; | 323 | T[11] = W[3] - W[7] - W[11] + W[15]; |
| 320 | 324 | ||
| 321 | T[12]=W[0]-W[4]+W[8]-W[12]; | 325 | T[12] = W[0] - W[4] + W[8] - W[12]; |
| 322 | T[13]=W[1]-W[5]+W[9]-W[13]; | 326 | T[13] = W[1] - W[5] + W[9] - W[13]; |
| 323 | T[14]=W[2]-W[6]+W[10]-W[14]; | 327 | T[14] = W[2] - W[6] + W[10] - W[14]; |
| 324 | T[15]=W[3]-W[7]+W[11]-W[15]; | 328 | T[15] = W[3] - W[7] + W[11] - W[15]; |
| 325 | 329 | ||
| 326 | Y[0]=T[0]+T[1]+T[2]+T[3]; | 330 | Y[0] = T[0] + T[1] + T[2] + T[3]; |
| 327 | Y[1]=T[0]+T[1]-T[2]-T[3]; | 331 | Y[1] = T[0] + T[1] - T[2] - T[3]; |
| 328 | Y[2]=T[0]-T[1]-T[2]+T[3]; | 332 | Y[2] = T[0] - T[1] - T[2] + T[3]; |
| 329 | Y[3]=T[0]-T[1]+T[2]-T[3]; | 333 | Y[3] = T[0] - T[1] + T[2] - T[3]; |
| 330 | 334 | ||
| 331 | Y[4]=T[4]+T[5]+T[6]+T[7]; | 335 | Y[4] = T[4] + T[5] + T[6] + T[7]; |
| 332 | Y[5]=T[4]+T[5]-T[6]-T[7]; | 336 | Y[5] = T[4] + T[5] - T[6] - T[7]; |
| 333 | Y[6]=T[4]-T[5]-T[6]+T[7]; | 337 | Y[6] = T[4] - T[5] - T[6] + T[7]; |
| 334 | Y[7]=T[4]-T[5]+T[6]-T[7]; | 338 | Y[7] = T[4] - T[5] + T[6] - T[7]; |
| 335 | 339 | ||
| 336 | Y[8]=T[8]+T[9]+T[10]+T[11]; | 340 | Y[8] = T[8] + T[9] + T[10] + T[11]; |
| 337 | Y[9]=T[8]+T[9]-T[10]-T[11]; | 341 | Y[9] = T[8] + T[9] - T[10] - T[11]; |
| 338 | Y[10]=T[8]-T[9]-T[10]+T[11]; | 342 | Y[10] = T[8] - T[9] - T[10] + T[11]; |
| 339 | Y[11]=T[8]-T[9]+T[10]-T[11]; | 343 | Y[11] = T[8] - T[9] + T[10] - T[11]; |
| 340 | 344 | ||
| 341 | Y[12]=T[12]+T[13]+T[14]+T[15]; | 345 | Y[12] = T[12] + T[13] + T[14] + T[15]; |
| 342 | Y[13]=T[12]+T[13]-T[14]-T[15]; | 346 | Y[13] = T[12] + T[13] - T[14] - T[15]; |
| 343 | Y[14]=T[12]-T[13]-T[14]+T[15]; | 347 | Y[14] = T[12] - T[13] - T[14] + T[15]; |
| 344 | Y[15]=T[12]-T[13]+T[14]-T[15]; | 348 | Y[15] = T[12] - T[13] + T[14] - T[15]; |
| 345 | 349 | ||
| 346 | int32_t iSumSatd = 0; | 350 | int32_t iSumSatd = 0; |
| 347 | for(int i=0; i<16; i++) | 351 | for (int i = 0; i < 16; i++) |
| 348 | iSumSatd+=abs(Y[i]); | 352 | iSumSatd += abs (Y[i]); |
| 349 | 353 | ||
| 350 | EXPECT_EQ(WelsSampleSatd4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), (iSumSatd+1)>>1); | 354 | EXPECT_EQ (WelsSampleSatd4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), (iSumSatd + 1) >> 1); |
| 351 | } | 355 | } |
| 352 | 356 | ||
| 353 | TEST_F(SadSatdCFuncTest, WelsSampleSadFour16x16_c) { | 357 | TEST_F (SadSatdCFuncTest, WelsSampleSadFour16x16_c) { |
| 354 | srand((uint32_t)time(NULL)); | 358 | srand ((uint32_t)time (NULL)); |
| 355 | for(int i=0; i<(m_iStrideA<<5); i++) | 359 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 356 | m_pPixSrcA[i]=rand()%256; | 360 | m_pPixSrcA[i] = rand() % 256; |
| 357 | for(int i=0; i<(m_iStrideB<<5); i++) | 361 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 358 | m_pPixSrcB[i]=rand()%256; | 362 | m_pPixSrcB[i] = rand() % 256; |
| 359 | uint8_t *pPixA=m_pPixSrcA; | 363 | uint8_t* pPixA = m_pPixSrcA; |
| 360 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 364 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 361 | 365 | ||
| 362 | int32_t iSumSad = 0; | 366 | int32_t iSumSad = 0; |
| 363 | for (int i = 0; i <16; i++ ) { | 367 | for (int i = 0; i < 16; i++) { |
| 364 | for(int j=0; j<16; j++) { | 368 | for (int j = 0; j < 16; j++) { |
| 365 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 369 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 366 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 370 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 367 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 371 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 368 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 372 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 369 | } | 373 | } |
| 370 | pPixA += m_iStrideA; | 374 | pPixA += m_iStrideA; |
| 371 | pPixB += m_iStrideB; | 375 | pPixB += m_iStrideB; |
| 372 | } | 376 | } |
| 373 | WelsSampleSadFour16x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 377 | WelsSampleSadFour16x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 374 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 378 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 375 | } | 379 | } |
| 376 | 380 | ||
| 377 | TEST_F(SadSatdCFuncTest, WelsSampleSadFour16x8_c) { | 381 | TEST_F (SadSatdCFuncTest, WelsSampleSadFour16x8_c) { |
| 378 | srand((uint32_t)time(NULL)); | 382 | srand ((uint32_t)time (NULL)); |
| 379 | for(int i=0; i<(m_iStrideA<<5); i++) | 383 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 380 | m_pPixSrcA[i]=rand()%256; | 384 | m_pPixSrcA[i] = rand() % 256; |
| 381 | for(int i=0; i<(m_iStrideB<<5); i++) | 385 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 382 | m_pPixSrcB[i]=rand()%256; | 386 | m_pPixSrcB[i] = rand() % 256; |
| 383 | uint8_t *pPixA=m_pPixSrcA; | 387 | uint8_t* pPixA = m_pPixSrcA; |
| 384 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 388 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 385 | 389 | ||
| 386 | int32_t iSumSad = 0; | 390 | int32_t iSumSad = 0; |
| 387 | for (int i = 0; i <8; i++ ) { | 391 | for (int i = 0; i < 8; i++) { |
| 388 | for(int j=0; j<16; j++) { | 392 | for (int j = 0; j < 16; j++) { |
| 389 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 393 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 390 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 394 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 391 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 395 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 392 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 396 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 393 | } | 397 | } |
| 394 | pPixA += m_iStrideA; | 398 | pPixA += m_iStrideA; |
| 395 | pPixB += m_iStrideB; | 399 | pPixB += m_iStrideB; |
| 396 | } | 400 | } |
| 397 | 401 | ||
| 398 | WelsSampleSadFour16x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 402 | WelsSampleSadFour16x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 399 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 403 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 400 | } | 404 | } |
| 401 | 405 | ||
| 402 | TEST_F(SadSatdCFuncTest, WelsSampleSadFour8x16_c) { | 406 | TEST_F (SadSatdCFuncTest, WelsSampleSadFour8x16_c) { |
| 403 | srand((uint32_t)time(NULL)); | 407 | srand ((uint32_t)time (NULL)); |
| 404 | for(int i=0; i<(m_iStrideA<<5); i++) | 408 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 405 | m_pPixSrcA[i]=rand()%256; | 409 | m_pPixSrcA[i] = rand() % 256; |
| 406 | for(int i=0; i<(m_iStrideB<<5); i++) | 410 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 407 | m_pPixSrcB[i]=rand()%256; | 411 | m_pPixSrcB[i] = rand() % 256; |
| 408 | uint8_t *pPixA=m_pPixSrcA; | 412 | uint8_t* pPixA = m_pPixSrcA; |
| 409 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 413 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 410 | 414 | ||
| 411 | int32_t iSumSad = 0; | 415 | int32_t iSumSad = 0; |
| 412 | for (int i = 0; i <16; i++ ) { | 416 | for (int i = 0; i < 16; i++) { |
| 413 | for(int j=0; j<8; j++) { | 417 | for (int j = 0; j < 8; j++) { |
| 414 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 418 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 415 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 419 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 416 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 420 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 417 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 421 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 418 | } | 422 | } |
| 419 | pPixA += m_iStrideA; | 423 | pPixA += m_iStrideA; |
| 420 | pPixB += m_iStrideB; | 424 | pPixB += m_iStrideB; |
| 421 | } | 425 | } |
| 422 | 426 | ||
| 423 | WelsSampleSadFour8x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 427 | WelsSampleSadFour8x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 424 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 428 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 425 | } | 429 | } |
| 426 | 430 | ||
| 427 | TEST_F(SadSatdCFuncTest, WelsSampleSadFour8x8_c) { | 431 | TEST_F (SadSatdCFuncTest, WelsSampleSadFour8x8_c) { |
| 428 | srand((uint32_t)time(NULL)); | 432 | srand ((uint32_t)time (NULL)); |
| 429 | for(int i=0; i<(m_iStrideA<<4); i++) | 433 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 430 | m_pPixSrcA[i]=rand()%256; | 434 | m_pPixSrcA[i] = rand() % 256; |
| 431 | for(int i=0; i<(m_iStrideB<<4); i++) | 435 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 432 | m_pPixSrcB[i]=rand()%256; | 436 | m_pPixSrcB[i] = rand() % 256; |
| 433 | uint8_t *pPixA=m_pPixSrcA; | 437 | uint8_t* pPixA = m_pPixSrcA; |
| 434 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 438 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 435 | 439 | ||
| 436 | int32_t iSumSad = 0; | 440 | int32_t iSumSad = 0; |
| 437 | for (int i = 0; i <8; i++ ) { | 441 | for (int i = 0; i < 8; i++) { |
| 438 | for(int j=0; j<8; j++) { | 442 | for (int j = 0; j < 8; j++) { |
| 439 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 443 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 440 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 444 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 441 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 445 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 442 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 446 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 443 | } | 447 | } |
| 444 | pPixA += m_iStrideA; | 448 | pPixA += m_iStrideA; |
| 445 | pPixB += m_iStrideB; | 449 | pPixB += m_iStrideB; |
| 446 | } | 450 | } |
| 447 | WelsSampleSadFour8x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 451 | WelsSampleSadFour8x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 448 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 452 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 449 | } | 453 | } |
| 450 | 454 | ||
| 451 | TEST_F(SadSatdCFuncTest, WelsSampleSadFour4x4_c) { | 455 | TEST_F (SadSatdCFuncTest, WelsSampleSadFour4x4_c) { |
| 452 | srand((uint32_t)time(NULL)); | 456 | srand ((uint32_t)time (NULL)); |
| 453 | for(int i=0; i<(m_iStrideA<<3); i++) | 457 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 454 | m_pPixSrcA[i]=rand()%256; | 458 | m_pPixSrcA[i] = rand() % 256; |
| 455 | for(int i=0; i<(m_iStrideB<<3); i++) | 459 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 456 | m_pPixSrcB[i]=rand()%256; | 460 | m_pPixSrcB[i] = rand() % 256; |
| 457 | uint8_t *pPixA=m_pPixSrcA; | 461 | uint8_t* pPixA = m_pPixSrcA; |
| 458 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 462 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 459 | 463 | ||
| 460 | int32_t iSumSad = 0; | 464 | int32_t iSumSad = 0; |
| 461 | for (int i = 0; i <4; i++ ) { | 465 | for (int i = 0; i < 4; i++) { |
| 462 | for(int j=0; j<4; j++) { | 466 | for (int j = 0; j < 4; j++) { |
| 463 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 467 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 464 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 468 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 465 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 469 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 466 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 470 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 467 | } | 471 | } |
| 468 | pPixA += m_iStrideA; | 472 | pPixA += m_iStrideA; |
| 469 | pPixB += m_iStrideB; | 473 | pPixB += m_iStrideB; |
| 470 | } | 474 | } |
| 471 | WelsSampleSadFour4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 475 | WelsSampleSadFour4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 472 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 476 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 473 | } | 477 | } |
| 474 | 478 | ||
| 475 | #ifdef X86_ASM | 479 | #ifdef X86_ASM |
| 476 | class SadSatdAssemblyFuncTest : public testing::Test { | 480 | class SadSatdAssemblyFuncTest : public testing::Test { |
| 477 | public: | 481 | public: |
| 478 | virtual void SetUp() { | 482 | virtual void SetUp() { |
| 479 | int32_t iCpuCores = 0; | 483 | int32_t iCpuCores = 0; |
| 480 | m_uiCpuFeatureFlag = WelsCPUFeatureDetect(&iCpuCores); | 484 | m_uiCpuFeatureFlag = WelsCPUFeatureDetect (&iCpuCores); |
| 481 | pMemAlign = new CMemoryAlign(16); | 485 | pMemAlign = new CMemoryAlign (16); |
| 482 | m_iStrideA = m_iStrideB = PIXEL_STRIDE; | 486 | m_iStrideA = m_iStrideB = PIXEL_STRIDE; |
| 483 | m_pPixSrcA = (uint8_t *)pMemAlign->WelsMalloc(m_iStrideA<<5,"Sad_m_pPixSrcA"); | 487 | m_pPixSrcA = (uint8_t*)pMemAlign->WelsMalloc (m_iStrideA << 5, "Sad_m_pPixSrcA"); |
| 484 | ASSERT_TRUE(NULL != m_pPixSrcA); | 488 | ASSERT_TRUE (NULL != m_pPixSrcA); |
| 485 | m_pPixSrcB = (uint8_t *)pMemAlign->WelsMalloc(m_iStrideB<<5,"Sad_m_pPixSrcB"); | 489 | m_pPixSrcB = (uint8_t*)pMemAlign->WelsMalloc (m_iStrideB << 5, "Sad_m_pPixSrcB"); |
| 486 | ASSERT_MEMORY_FAIL2X(m_pPixSrcA, m_pPixSrcB) | 490 | ASSERT_MEMORY_FAIL2X (m_pPixSrcA, m_pPixSrcB) |
| 487 | m_pSad = (int32_t*)pMemAlign->WelsMalloc(4*sizeof(int32_t), "m_pSad"); | 491 | m_pSad = (int32_t*)pMemAlign->WelsMalloc (4 * sizeof (int32_t), "m_pSad"); |
| 488 | ASSERT_MEMORY_FAIL3X(m_pPixSrcA, m_pPixSrcB, m_pSad) | 492 | ASSERT_MEMORY_FAIL3X (m_pPixSrcA, m_pPixSrcB, m_pSad) |
| 489 | } | 493 | } |
| 490 | virtual void TearDown() { | 494 | virtual void TearDown() { |
| 491 | pMemAlign->WelsFree(m_pPixSrcA,"Sad_m_pPixSrcA"); | 495 | pMemAlign->WelsFree (m_pPixSrcA, "Sad_m_pPixSrcA"); |
| 492 | pMemAlign->WelsFree(m_pPixSrcB,"Sad_m_pPixSrcB"); | 496 | pMemAlign->WelsFree (m_pPixSrcB, "Sad_m_pPixSrcB"); |
| 493 | pMemAlign->WelsFree(m_pSad,"m_pSad"); | 497 | pMemAlign->WelsFree (m_pSad, "m_pSad"); |
| 494 | delete pMemAlign; | 498 | delete pMemAlign; |
| 495 | } | 499 | } |
| 496 | public: | 500 | public: |
| 497 | uint32_t m_uiCpuFeatureFlag; | 501 | uint32_t m_uiCpuFeatureFlag; |
| 498 | uint8_t* m_pPixSrcA; | 502 | uint8_t* m_pPixSrcA; |
| 499 | uint8_t* m_pPixSrcB; | 503 | uint8_t* m_pPixSrcB; |
| @@ -504,332 +508,347 @@ public: · 4 unmodified lines | |||
| 504 | CMemoryAlign* pMemAlign; | 508 | CMemoryAlign* pMemAlign; |
| 505 | }; | 509 | }; |
| 506 | 510 | ||
| 507 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSad4x4_mmx) { | 511 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSad4x4_mmx) { |
| 508 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_MMXEXT)) | 512 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_MMXEXT)) |
| 509 | return; | 513 | return; |
| 510 | 514 | ||
| 511 | srand((uint32_t)time(NULL)); | 515 | srand ((uint32_t)time (NULL)); |
| 512 | for(int i=0; i<(m_iStrideA<<2); i++) | 516 | for (int i = 0; i < (m_iStrideA << 2); i++) |
| 513 | m_pPixSrcA[i]=rand()%256; | 517 | m_pPixSrcA[i] = rand() % 256; |
| 514 | for(int i=0; i<(m_iStrideB<<2); i++) | 518 | for (int i = 0; i < (m_iStrideB << 2); i++) |
| 515 | m_pPixSrcB[i]=rand()%256; | 519 | m_pPixSrcB[i] = rand() % 256; |
| 516 | 520 | ||
| 517 | EXPECT_EQ(WelsSampleSad4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad4x4_mmx(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 521 | EXPECT_EQ (WelsSampleSad4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad4x4_mmx (m_pPixSrcA, |
| 522 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 518 | } | 523 | } |
| 519 | 524 | ||
| 520 | 525 | ||
| 521 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSad8x8_sse21) { | 526 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSad8x8_sse21) { |
| 522 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 527 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 523 | return; | 528 | return; |
| 524 | 529 | ||
| 525 | srand((uint32_t)time(NULL)); | 530 | srand ((uint32_t)time (NULL)); |
| 526 | for(int i=0; i<(m_iStrideA<<3); i++) | 531 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 527 | m_pPixSrcA[i]=rand()%256; | 532 | m_pPixSrcA[i] = rand() % 256; |
| 528 | for(int i=0; i<(m_iStrideB<<3); i++) | 533 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 529 | m_pPixSrcB[i]=rand()%256; | 534 | m_pPixSrcB[i] = rand() % 256; |
| 530 | 535 | ||
| 531 | EXPECT_EQ(WelsSampleSad8x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad8x8_sse21(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB )); | 536 | EXPECT_EQ (WelsSampleSad8x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad8x8_sse21 (m_pPixSrcA, |
| 537 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 532 | } | 538 | } |
| 533 | 539 | ||
| 534 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSad8x16_sse2) { | 540 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSad8x16_sse2) { |
| 535 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 541 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 536 | return; | 542 | return; |
| 537 | srand((uint32_t)time(NULL)); | 543 | srand ((uint32_t)time (NULL)); |
| 538 | for(int i=0; i<(m_iStrideA<<4); i++) | 544 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 539 | m_pPixSrcA[i]=rand()%256; | 545 | m_pPixSrcA[i] = rand() % 256; |
| 540 | for(int i=0; i<(m_iStrideB<<4); i++) | 546 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 541 | m_pPixSrcB[i]=rand()%256; | 547 | m_pPixSrcB[i] = rand() % 256; |
| 542 | 548 | ||
| 543 | EXPECT_EQ(WelsSampleSad8x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad8x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 549 | EXPECT_EQ (WelsSampleSad8x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad8x16_sse2 (m_pPixSrcA, |
| 550 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 544 | } | 551 | } |
| 545 | 552 | ||
| 546 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSad16x8_sse2) { | 553 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSad16x8_sse2) { |
| 547 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 554 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 548 | return; | 555 | return; |
| 549 | srand((uint32_t)time(NULL)); | 556 | srand ((uint32_t)time (NULL)); |
| 550 | for(int i=0; i<(m_iStrideA<<3); i++) | 557 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 551 | m_pPixSrcA[i]=rand()%256; | 558 | m_pPixSrcA[i] = rand() % 256; |
| 552 | for(int i=0; i<(m_iStrideB<<3); i++) | 559 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 553 | m_pPixSrcB[i]=rand()%256; | 560 | m_pPixSrcB[i] = rand() % 256; |
| 554 | 561 | ||
| 555 | EXPECT_EQ(WelsSampleSad16x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad16x8_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 562 | EXPECT_EQ (WelsSampleSad16x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad16x8_sse2 (m_pPixSrcA, |
| 563 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 556 | } | 564 | } |
| 557 | 565 | ||
| 558 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSad16x16_sse2) { | 566 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSad16x16_sse2) { |
| 559 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 567 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 560 | return; | 568 | return; |
| 561 | srand((uint32_t)time(NULL)); | 569 | srand ((uint32_t)time (NULL)); |
| 562 | for(int i=0; i<(m_iStrideA<<4); i++) | 570 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 563 | m_pPixSrcA[i]=rand()%256; | 571 | m_pPixSrcA[i] = rand() % 256; |
| 564 | for(int i=0; i<(m_iStrideB<<4); i++) | 572 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 565 | m_pPixSrcB[i]=rand()%256; | 573 | m_pPixSrcB[i] = rand() % 256; |
| 566 | 574 | ||
| 567 | EXPECT_EQ(WelsSampleSad16x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad16x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 575 | EXPECT_EQ (WelsSampleSad16x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSad16x16_sse2 (m_pPixSrcA, |
| 576 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 568 | } | 577 | } |
| 569 | 578 | ||
| 570 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd4x4_sse2) { | 579 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd4x4_sse2) { |
| 571 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 580 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 572 | return; | 581 | return; |
| 573 | srand((uint32_t)time(NULL)); | 582 | srand ((uint32_t)time (NULL)); |
| 574 | for(int i=0; i<(m_iStrideA<<2); i++) | 583 | for (int i = 0; i < (m_iStrideA << 2); i++) |
| 575 | m_pPixSrcA[i]=rand()%256; | 584 | m_pPixSrcA[i] = rand() % 256; |
| 576 | for(int i=0; i<(m_iStrideB<<2); i++) | 585 | for (int i = 0; i < (m_iStrideB << 2); i++) |
| 577 | m_pPixSrcB[i]=rand()%256; | 586 | m_pPixSrcB[i] = rand() % 256; |
| 578 | 587 | ||
| 579 | EXPECT_EQ(WelsSampleSatd4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd4x4_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 588 | EXPECT_EQ (WelsSampleSatd4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd4x4_sse2 (m_pPixSrcA, |
| 589 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 580 | } | 590 | } |
| 581 | 591 | ||
| 582 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd8x8_sse2) { | 592 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd8x8_sse2) { |
| 583 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 593 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 584 | return; | 594 | return; |
| 585 | srand((uint32_t)time(NULL)); | 595 | srand ((uint32_t)time (NULL)); |
| 586 | for(int i=0; i<(m_iStrideA<<3); i++) | 596 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 587 | m_pPixSrcA[i]=rand()%256; | 597 | m_pPixSrcA[i] = rand() % 256; |
| 588 | for(int i=0; i<(m_iStrideB<<3); i++) | 598 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 589 | m_pPixSrcB[i]=rand()%256; | 599 | m_pPixSrcB[i] = rand() % 256; |
| 590 | 600 | ||
| 591 | EXPECT_EQ(WelsSampleSatd8x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x8_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB )); | 601 | EXPECT_EQ (WelsSampleSatd8x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x8_sse2 (m_pPixSrcA, |
| 602 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 592 | } | 603 | } |
| 593 | 604 | ||
| 594 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd8x16_sse2) { | 605 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd8x16_sse2) { |
| 595 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 606 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 596 | return; | 607 | return; |
| 597 | srand((uint32_t)time(NULL)); | 608 | srand ((uint32_t)time (NULL)); |
| 598 | for(int i=0; i<(m_iStrideA<<4); i++) | 609 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 599 | m_pPixSrcA[i]=rand()%256; | 610 | m_pPixSrcA[i] = rand() % 256; |
| 600 | for(int i=0; i<(m_iStrideB<<4); i++) | 611 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 601 | m_pPixSrcB[i]=rand()%256; | 612 | m_pPixSrcB[i] = rand() % 256; |
| 602 | 613 | ||
| 603 | EXPECT_EQ(WelsSampleSatd8x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 614 | EXPECT_EQ (WelsSampleSatd8x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x16_sse2 (m_pPixSrcA, |
| 615 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 604 | } | 616 | } |
| 605 | 617 | ||
| 606 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd16x8_sse2) { | 618 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd16x8_sse2) { |
| 607 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 619 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 608 | return; | 620 | return; |
| 609 | srand((uint32_t)time(NULL)); | 621 | srand ((uint32_t)time (NULL)); |
| 610 | for(int i=0; i<(m_iStrideA<<3); i++) | 622 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 611 | m_pPixSrcA[i]=rand()%256; | 623 | m_pPixSrcA[i] = rand() % 256; |
| 612 | for(int i=0; i<(m_iStrideB<<3); i++) | 624 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 613 | m_pPixSrcB[i]=rand()%256; | 625 | m_pPixSrcB[i] = rand() % 256; |
| 614 | 626 | ||
| 615 | EXPECT_EQ(WelsSampleSatd16x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x8_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 627 | EXPECT_EQ (WelsSampleSatd16x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x8_sse2 (m_pPixSrcA, |
| 628 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 616 | } | 629 | } |
| 617 | 630 | ||
| 618 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd16x16_sse2) { | 631 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd16x16_sse2) { |
| 619 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 632 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 620 | return; | 633 | return; |
| 621 | srand((uint32_t)time(NULL)); | 634 | srand ((uint32_t)time (NULL)); |
| 622 | for(int i=0; i<(m_iStrideA<<4); i++) | 635 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 623 | m_pPixSrcA[i]=rand()%256; | 636 | m_pPixSrcA[i] = rand() % 256; |
| 624 | for(int i=0; i<(m_iStrideB<<4); i++) | 637 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 625 | m_pPixSrcB[i]=rand()%256; | 638 | m_pPixSrcB[i] = rand() % 256; |
| 626 | 639 | ||
| 627 | EXPECT_EQ(WelsSampleSatd16x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 640 | EXPECT_EQ (WelsSampleSatd16x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x16_sse2 (m_pPixSrcA, |
| 641 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 628 | } | 642 | } |
| 629 | 643 | ||
| 630 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd4x4_sse41) { | 644 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd4x4_sse41) { |
| 631 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 645 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 632 | return; | 646 | return; |
| 633 | srand((uint32_t)time(NULL)); | 647 | srand ((uint32_t)time (NULL)); |
| 634 | for(int i=0; i<(m_iStrideA<<2); i++) | 648 | for (int i = 0; i < (m_iStrideA << 2); i++) |
| 635 | m_pPixSrcA[i]=rand()%256; | 649 | m_pPixSrcA[i] = rand() % 256; |
| 636 | for(int i=0; i<(m_iStrideB<<2); i++) | 650 | for (int i = 0; i < (m_iStrideB << 2); i++) |
| 637 | m_pPixSrcB[i]=rand()%256; | 651 | m_pPixSrcB[i] = rand() % 256; |
| 638 | 652 | ||
| 639 | EXPECT_EQ(WelsSampleSatd4x4_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd4x4_sse41(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 653 | EXPECT_EQ (WelsSampleSatd4x4_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd4x4_sse41 (m_pPixSrcA, |
| 654 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 640 | } | 655 | } |
| 641 | 656 | ||
| 642 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd8x8_sse41) { | 657 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd8x8_sse41) { |
| 643 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 658 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 644 | return; | 659 | return; |
| 645 | srand((uint32_t)time(NULL)); | 660 | srand ((uint32_t)time (NULL)); |
| 646 | for(int i=0; i<(m_iStrideA<<3); i++) | 661 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 647 | m_pPixSrcA[i]=rand()%256; | 662 | m_pPixSrcA[i] = rand() % 256; |
| 648 | for(int i=0; i<(m_iStrideB<<3); i++) | 663 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 649 | m_pPixSrcB[i]=rand()%256; | 664 | m_pPixSrcB[i] = rand() % 256; |
| 650 | 665 | ||
| 651 | EXPECT_EQ(WelsSampleSatd8x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x8_sse41(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB )); | 666 | EXPECT_EQ (WelsSampleSatd8x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x8_sse41 (m_pPixSrcA, |
| 667 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 652 | } | 668 | } |
| 653 | 669 | ||
| 654 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd8x16_sse41) { | 670 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd8x16_sse41) { |
| 655 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 671 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 656 | return; | 672 | return; |
| 657 | srand((uint32_t)time(NULL)); | 673 | srand ((uint32_t)time (NULL)); |
| 658 | for(int i=0; i<(m_iStrideA<<4); i++) | 674 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 659 | m_pPixSrcA[i]=rand()%256; | 675 | m_pPixSrcA[i] = rand() % 256; |
| 660 | for(int i=0; i<(m_iStrideB<<4); i++) | 676 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 661 | m_pPixSrcB[i]=rand()%256; | 677 | m_pPixSrcB[i] = rand() % 256; |
| 662 | 678 | ||
| 663 | EXPECT_EQ(WelsSampleSatd8x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x16_sse41(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 679 | EXPECT_EQ (WelsSampleSatd8x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd8x16_sse41 (m_pPixSrcA, |
| 680 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 664 | } | 681 | } |
| 665 | 682 | ||
| 666 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd16x8_sse41) { | 683 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd16x8_sse41) { |
| 667 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 684 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 668 | return; | 685 | return; |
| 669 | srand((uint32_t)time(NULL)); | 686 | srand ((uint32_t)time (NULL)); |
| 670 | for(int i=0; i<(m_iStrideA<<3); i++) | 687 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 671 | m_pPixSrcA[i]=rand()%256; | 688 | m_pPixSrcA[i] = rand() % 256; |
| 672 | for(int i=0; i<(m_iStrideB<<3); i++) | 689 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 673 | m_pPixSrcB[i]=rand()%256; | 690 | m_pPixSrcB[i] = rand() % 256; |
| 674 | 691 | ||
| 675 | EXPECT_EQ(WelsSampleSatd16x8_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x8_sse41(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 692 | EXPECT_EQ (WelsSampleSatd16x8_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x8_sse41 (m_pPixSrcA, |
| 693 | m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 676 | } | 694 | } |
| 677 | 695 | ||
| 678 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSatd16x16_sse41) { | 696 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSatd16x16_sse41) { |
| 679 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) | 697 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE41)) |
| 680 | return; | 698 | return; |
| 681 | srand((uint32_t)time(NULL)); | 699 | srand ((uint32_t)time (NULL)); |
| 682 | for(int i=0; i<(m_iStrideA<<4); i++) | 700 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 683 | m_pPixSrcA[i]=rand()%256; | 701 | m_pPixSrcA[i] = rand() % 256; |
| 684 | for(int i=0; i<(m_iStrideB<<4); i++) | 702 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 685 | m_pPixSrcB[i]=rand()%256; | 703 | m_pPixSrcB[i] = rand() % 256; |
| 686 | 704 | ||
| 687 | EXPECT_EQ(WelsSampleSatd16x16_c(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), WelsSampleSatd16x16_sse41(m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | 705 | EXPECT_EQ (WelsSampleSatd16x16_c (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB), |
| 706 | WelsSampleSatd16x16_sse41 (m_pPixSrcA, m_iStrideA, m_pPixSrcB, m_iStrideB)); | ||
| 688 | } | 707 | } |
| 689 | 708 | ||
| 690 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSadFour16x16_sse2) { | 709 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSadFour16x16_sse2) { |
| 691 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 710 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 692 | return; | 711 | return; |
| 693 | 712 | ||
| 694 | srand((uint32_t)time(NULL)); | 713 | srand ((uint32_t)time (NULL)); |
| 695 | for(int i=0; i<(m_iStrideA<<5); i++) | 714 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 696 | m_pPixSrcA[i]=rand()%256; | 715 | m_pPixSrcA[i] = rand() % 256; |
| 697 | for(int i=0; i<(m_iStrideB<<5); i++) | 716 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 698 | m_pPixSrcB[i]=rand()%256; | 717 | m_pPixSrcB[i] = rand() % 256; |
| 699 | uint8_t *pPixA=m_pPixSrcA; | 718 | uint8_t* pPixA = m_pPixSrcA; |
| 700 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 719 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 701 | 720 | ||
| 702 | int32_t iSumSad = 0; | 721 | int32_t iSumSad = 0; |
| 703 | for (int i = 0; i <16; i++ ) { | 722 | for (int i = 0; i < 16; i++) { |
| 704 | for(int j=0; j<16; j++) { | 723 | for (int j = 0; j < 16; j++) { |
| 705 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 724 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 706 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 725 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 707 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 726 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 708 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 727 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 709 | } | 728 | } |
| 710 | 729 | ||
| 711 | pPixA += m_iStrideA; | 730 | pPixA += m_iStrideA; |
| 712 | pPixB += m_iStrideB; | 731 | pPixB += m_iStrideB; |
| 713 | } | 732 | } |
| 714 | 733 | ||
| 715 | WelsSampleSadFour16x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 734 | WelsSampleSadFour16x16_sse2 (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 716 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 735 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 717 | } | 736 | } |
| 718 | 737 | ||
| 719 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSadFour16x8_sse2) { | 738 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSadFour16x8_sse2) { |
| 720 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 739 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 721 | return; | 740 | return; |
| 722 | 741 | ||
| 723 | srand((uint32_t)time(NULL)); | 742 | srand ((uint32_t)time (NULL)); |
| 724 | for(int i=0; i<(m_iStrideA<<5); i++) | 743 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 725 | m_pPixSrcA[i]=rand()%256; | 744 | m_pPixSrcA[i] = rand() % 256; |
| 726 | for(int i=0; i<(m_iStrideB<<5); i++) | 745 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 727 | m_pPixSrcB[i]=rand()%256; | 746 | m_pPixSrcB[i] = rand() % 256; |
| 728 | uint8_t *pPixA=m_pPixSrcA; | 747 | uint8_t* pPixA = m_pPixSrcA; |
| 729 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 748 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 730 | 749 | ||
| 731 | int32_t iSumSad = 0; | 750 | int32_t iSumSad = 0; |
| 732 | for (int i = 0; i <8; i++ ) { | 751 | for (int i = 0; i < 8; i++) { |
| 733 | for(int j=0; j<16; j++) { | 752 | for (int j = 0; j < 16; j++) { |
| 734 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 753 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 735 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 754 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 736 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 755 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 737 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 756 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 738 | } | 757 | } |
| 739 | 758 | ||
| 740 | pPixA += m_iStrideA; | 759 | pPixA += m_iStrideA; |
| 741 | pPixB += m_iStrideB; | 760 | pPixB += m_iStrideB; |
| 742 | } | 761 | } |
| 743 | 762 | ||
| 744 | WelsSampleSadFour16x8_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 763 | WelsSampleSadFour16x8_sse2 (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 745 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 764 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 746 | } | 765 | } |
| 747 | 766 | ||
| 748 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSadFour8x16_sse2) { | 767 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSadFour8x16_sse2) { |
| 749 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 768 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 750 | return; | 769 | return; |
| 751 | 770 | ||
| 752 | srand((uint32_t)time(NULL)); | 771 | srand ((uint32_t)time (NULL)); |
| 753 | for(int i=0; i<(m_iStrideA<<5); i++) | 772 | for (int i = 0; i < (m_iStrideA << 5); i++) |
| 754 | m_pPixSrcA[i]=rand()%256; | 773 | m_pPixSrcA[i] = rand() % 256; |
| 755 | for(int i=0; i<(m_iStrideB<<5); i++) | 774 | for (int i = 0; i < (m_iStrideB << 5); i++) |
| 756 | m_pPixSrcB[i]=rand()%256; | 775 | m_pPixSrcB[i] = rand() % 256; |
| 757 | uint8_t *pPixA=m_pPixSrcA; | 776 | uint8_t* pPixA = m_pPixSrcA; |
| 758 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 777 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 759 | 778 | ||
| 760 | int32_t iSumSad = 0; | 779 | int32_t iSumSad = 0; |
| 761 | for (int i = 0; i <16; i++ ) { | 780 | for (int i = 0; i < 16; i++) { |
| 762 | for(int j=0; j<8; j++) { | 781 | for (int j = 0; j < 8; j++) { |
| 763 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 782 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 764 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 783 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 765 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 784 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 766 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 785 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 767 | } | 786 | } |
| 768 | 787 | ||
| 769 | pPixA += m_iStrideA; | 788 | pPixA += m_iStrideA; |
| 770 | pPixB += m_iStrideB; | 789 | pPixB += m_iStrideB; |
| 771 | } | 790 | } |
| 772 | 791 | ||
| 773 | WelsSampleSadFour8x16_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 792 | WelsSampleSadFour8x16_sse2 (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 774 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 793 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 775 | } | 794 | } |
| 776 | 795 | ||
| 777 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSadFour8x8_sse2) { | 796 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSadFour8x8_sse2) { |
| 778 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 797 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 779 | return; | 798 | return; |
| 780 | 799 | ||
| 781 | srand((uint32_t)time(NULL)); | 800 | srand ((uint32_t)time (NULL)); |
| 782 | for(int i=0; i<(m_iStrideA<<4); i++) | 801 | for (int i = 0; i < (m_iStrideA << 4); i++) |
| 783 | m_pPixSrcA[i]=rand()%256; | 802 | m_pPixSrcA[i] = rand() % 256; |
| 784 | for(int i=0; i<(m_iStrideB<<4); i++) | 803 | for (int i = 0; i < (m_iStrideB << 4); i++) |
| 785 | m_pPixSrcB[i]=rand()%256; | 804 | m_pPixSrcB[i] = rand() % 256; |
| 786 | uint8_t *pPixA=m_pPixSrcA; | 805 | uint8_t* pPixA = m_pPixSrcA; |
| 787 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 806 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 788 | 807 | ||
| 789 | int32_t iSumSad = 0; | 808 | int32_t iSumSad = 0; |
| 790 | for (int i = 0; i <8; i++ ) { | 809 | for (int i = 0; i < 8; i++) { |
| 791 | for(int j=0; j<8; j++) { | 810 | for (int j = 0; j < 8; j++) { |
| 792 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 811 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 793 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 812 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 794 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 813 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 795 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 814 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 796 | } | 815 | } |
| 797 | 816 | ||
| 798 | pPixA += m_iStrideA; | 817 | pPixA += m_iStrideA; |
| 799 | pPixB += m_iStrideB; | 818 | pPixB += m_iStrideB; |
| 800 | } | 819 | } |
| 801 | 820 | ||
| 802 | WelsSampleSadFour8x8_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 821 | WelsSampleSadFour8x8_sse2 (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 803 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 822 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 804 | } | 823 | } |
| 805 | 824 | ||
| 806 | 825 | ||
| 807 | TEST_F(SadSatdAssemblyFuncTest, WelsSampleSadFour4x4_sse2) { | 826 | TEST_F (SadSatdAssemblyFuncTest, WelsSampleSadFour4x4_sse2) { |
| 808 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) | 827 | if (0 == (m_uiCpuFeatureFlag & WELS_CPU_SSE2)) |
| 809 | return; | 828 | return; |
| 810 | 829 | ||
| 811 | srand((uint32_t)time(NULL)); | 830 | srand ((uint32_t)time (NULL)); |
| 812 | for(int i=0; i<(m_iStrideA<<3); i++) | 831 | for (int i = 0; i < (m_iStrideA << 3); i++) |
| 813 | m_pPixSrcA[i]=rand()%256; | 832 | m_pPixSrcA[i] = rand() % 256; |
| 814 | for(int i=0; i<(m_iStrideB<<3); i++) | 833 | for (int i = 0; i < (m_iStrideB << 3); i++) |
| 815 | m_pPixSrcB[i]=rand()%256; | 834 | m_pPixSrcB[i] = rand() % 256; |
| 816 | uint8_t *pPixA=m_pPixSrcA; | 835 | uint8_t* pPixA = m_pPixSrcA; |
| 817 | uint8_t *pPixB=m_pPixSrcB+m_iStrideB; | 836 | uint8_t* pPixB = m_pPixSrcB + m_iStrideB; |
| 818 | 837 | ||
| 819 | int32_t iSumSad = 0; | 838 | int32_t iSumSad = 0; |
| 820 | for (int i = 0; i <4; i++ ) { | 839 | for (int i = 0; i < 4; i++) { |
| 821 | for(int j=0; j<4; j++) { | 840 | for (int j = 0; j < 4; j++) { |
| 822 | iSumSad+=abs(pPixA[j]-pPixB[j-1]); | 841 | iSumSad += abs (pPixA[j] - pPixB[j - 1]); |
| 823 | iSumSad+=abs(pPixA[j]-pPixB[j+1]); | 842 | iSumSad += abs (pPixA[j] - pPixB[j + 1]); |
| 824 | iSumSad+=abs(pPixA[j]-pPixB[j-m_iStrideB]); | 843 | iSumSad += abs (pPixA[j] - pPixB[j - m_iStrideB]); |
| 825 | iSumSad+=abs(pPixA[j]-pPixB[j+m_iStrideB]); | 844 | iSumSad += abs (pPixA[j] - pPixB[j + m_iStrideB]); |
| 826 | } | 845 | } |
| 827 | 846 | ||
| 828 | pPixA += m_iStrideA; | 847 | pPixA += m_iStrideA; |
| 829 | pPixB += m_iStrideB; | 848 | pPixB += m_iStrideB; |
| 830 | } | 849 | } |
| 831 | 850 | ||
| 832 | WelsSampleSadFour4x4_sse2(m_pPixSrcA, m_iStrideA, m_pPixSrcB+m_iStrideB, m_iStrideB, m_pSad); | 851 | WelsSampleSadFour4x4_sse2 (m_pPixSrcA, m_iStrideA, m_pPixSrcB + m_iStrideB, m_iStrideB, m_pSad); |
| 833 | EXPECT_EQ(m_pSad[0]+m_pSad[1]+m_pSad[2]+m_pSad[3],iSumSad); | 852 | EXPECT_EQ (m_pSad[0] + m_pSad[1] + m_pSad[2] + m_pSad[3], iSumSad); |
| 834 | } | 853 | } |
| 835 | #endif | 854 | #endif |