add new c and assembly functions to optimize downsampler when downscale equal 1:3/1:4

Guangwei Wang committed 11 years ago

Commit 64657d31 parent 69a62ea9 files · +1125 −8Browse files

Showing the changes to codec/processing/src/downsample (3 of 9 files). Show all

Unified · Split

3 files

codec/processing/src/downsample/downsample.cpp +33 −0
codec/processing/src/downsample/downsample.h +32 −5
codec/processing/src/downsample/downsamplefuncs.cpp +47 −0

codec/processing/src/downsample/downsample.cpp+33−0

@@ -53,6 +53,8 @@ void CDownsampling::InitDownsampleFuncs (SDownsampleFuncs& sDownsampleFunc, int · 52 unmodified lines
53	53	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsampler_c;
54	54	sDownsampleFunc.pfHalfAverage[2] = DyadicBilinearDownsampler_c;
55	55	sDownsampleFunc.pfHalfAverage[3] = DyadicBilinearDownsampler_c;
	56	+ sDownsampleFunc.pfOneThirdDownsampler = DyadicBilinearOneThirdDownsampler_c;
	57	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_c;
56	58	sDownsampleFunc.pfGeneralRatioChroma = GeneralBilinearAccurateDownsampler_c;
57	59	sDownsampleFunc.pfGeneralRatioLuma = GeneralBilinearFastDownsampler_c;
58	60	#if defined(X86_ASM)
@@ -60,6 +62,7 @@ void CDownsampling::InitDownsampleFuncs (SDownsampleFuncs& sDownsampleFunc, int · 1 unmodified line
60	62	sDownsampleFunc.pfHalfAverage[0] = DyadicBilinearDownsamplerWidthx32_sse;
61	63	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsamplerWidthx16_sse;
62	64	sDownsampleFunc.pfHalfAverage[2] = DyadicBilinearDownsamplerWidthx8_sse;
	65	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_sse;
63	66	}
64	67	if (iCpuFlag & WELS_CPU_SSE2) {
65	68	sDownsampleFunc.pfGeneralRatioChroma = GeneralBilinearAccurateDownsamplerWrap_sse2;
@@ -68,10 +71,14 @@ void CDownsampling::InitDownsampleFuncs (SDownsampleFuncs& sDownsampleFunc, int · 2 unmodified lines
68	71	if (iCpuFlag & WELS_CPU_SSSE3) {
69	72	sDownsampleFunc.pfHalfAverage[0] = DyadicBilinearDownsamplerWidthx32_ssse3;
70	73	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsamplerWidthx16_ssse3;
	74	+ sDownsampleFunc.pfOneThirdDownsampler = DyadicBilinearOneThirdDownsampler_ssse3;
	75	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_ssse3;
71	76	}
72	77	if (iCpuFlag & WELS_CPU_SSE41) {
73	78	sDownsampleFunc.pfHalfAverage[0] = DyadicBilinearDownsamplerWidthx32_sse4;
74	79	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsamplerWidthx16_sse4;
	80	+ sDownsampleFunc.pfOneThirdDownsampler = DyadicBilinearOneThirdDownsampler_sse4;
	81	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_sse4;
75	82	}
76	83	#endif//X86_ASM
77	84
@@ -81,6 +88,8 @@ void CDownsampling::InitDownsampleFuncs (SDownsampleFuncs& sDownsampleFunc, int · 3 unmodified lines
81	88	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsampler_neon;
82	89	sDownsampleFunc.pfHalfAverage[2] = DyadicBilinearDownsampler_neon;
83	90	sDownsampleFunc.pfHalfAverage[3] = DyadicBilinearDownsampler_neon;
	91	+ sDownsampleFunc.pfOneThirdDownsampler = DyadicBilinearOneThirdDownsampler_neon;
	92	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_neon;
84	93	sDownsampleFunc.pfGeneralRatioChroma = GeneralBilinearAccurateDownsamplerWrap_neon;
85	94	sDownsampleFunc.pfGeneralRatioLuma = GeneralBilinearAccurateDownsamplerWrap_neon;
86	95	}
@@ -92,6 +101,8 @@ void CDownsampling::InitDownsampleFuncs (SDownsampleFuncs& sDownsampleFunc, int · 5 unmodified lines
92	101	sDownsampleFunc.pfHalfAverage[1] = DyadicBilinearDownsampler_AArch64_neon;
93	102	sDownsampleFunc.pfHalfAverage[2] = DyadicBilinearDownsampler_AArch64_neon;
94	103	sDownsampleFunc.pfHalfAverage[3] = DyadicBilinearDownsampler_AArch64_neon;
	104	+ sDownsampleFunc.pfOneThirdDownsampler = DyadicBilinearOneThirdDownsampler_AArch64_neon;
	105	+ sDownsampleFunc.pfQuarterDownsampler = DyadicBilinearQuarterDownsampler_AArch64_neon;
95	106	sDownsampleFunc.pfGeneralRatioChroma = GeneralBilinearAccurateDownsamplerWrap_AArch64_neon;
96	107	sDownsampleFunc.pfGeneralRatioLuma = GeneralBilinearAccurateDownsamplerWrap_AArch64_neon;
97	108	}
@@ -124,6 +135,28 @@ EResult CDownsampling::Process (int32_t iType, SPixMap* pSrcPixMap, SPixMap* pDs · 26 unmodified lines
124	135	(uint8_t*)pSrcPixMap->pPixel[1], pSrcPixMap->iStride[1], iSrcWidthUV, iSrcHeightUV);
125	136	m_pfDownsample.pfHalfAverage[iAlignIndex] ((uint8_t*)pDstPixMap->pPixel[2], pDstPixMap->iStride[2],
126	137	(uint8_t*)pSrcPixMap->pPixel[2], pSrcPixMap->iStride[2], iSrcWidthUV, iSrcHeightUV);
	138	+ } else if ((iSrcWidthY >> 2) == iDstWidthY && (iSrcHeightY >> 2) == iDstHeightY) {
	139	+
	140	+ m_pfDownsample.pfQuarterDownsampler ((uint8_t*)pDstPixMap->pPixel[0], pDstPixMap->iStride[0],
	141	+ (uint8_t*)pSrcPixMap->pPixel[0], pSrcPixMap->iStride[0], iSrcWidthY, iSrcHeightY);
	142	+
	143	+ m_pfDownsample.pfQuarterDownsampler ((uint8_t*)pDstPixMap->pPixel[1], pDstPixMap->iStride[1],
	144	+ (uint8_t*)pSrcPixMap->pPixel[1], pSrcPixMap->iStride[1], iSrcWidthUV, iSrcHeightUV);
	145	+
	146	+ m_pfDownsample.pfQuarterDownsampler ((uint8_t*)pDstPixMap->pPixel[2], pDstPixMap->iStride[2],
	147	+ (uint8_t*)pSrcPixMap->pPixel[2], pSrcPixMap->iStride[2], iSrcWidthUV, iSrcHeightUV);
	148	+
	149	+ } else if ((iSrcWidthY / 3) == iDstWidthY && (iSrcHeightY / 3) == iDstHeightY) {
	150	+
	151	+ m_pfDownsample.pfOneThirdDownsampler ((uint8_t*)pDstPixMap->pPixel[0], pDstPixMap->iStride[0],
	152	+ (uint8_t*)pSrcPixMap->pPixel[0], pSrcPixMap->iStride[0], iSrcWidthY, iDstHeightY);
	153	+
	154	+ m_pfDownsample.pfOneThirdDownsampler ((uint8_t*)pDstPixMap->pPixel[1], pDstPixMap->iStride[1],
	155	+ (uint8_t*)pSrcPixMap->pPixel[1], pSrcPixMap->iStride[1], iSrcWidthUV, iDstHeightUV);
	156	+
	157	+ m_pfDownsample.pfOneThirdDownsampler ((uint8_t*)pDstPixMap->pPixel[2], pDstPixMap->iStride[2],
	158	+ (uint8_t*)pSrcPixMap->pPixel[2], pSrcPixMap->iStride[2], iSrcWidthUV, iDstHeightUV);
	159	+
127	160	} else {
128	161	m_pfDownsample.pfGeneralRatioLuma ((uint8_t*)pDstPixMap->pPixel[0], pDstPixMap->iStride[0], iDstWidthY, iDstHeightY,
129	162	(uint8_t*)pSrcPixMap->pPixel[0], pSrcPixMap->iStride[0], iSrcWidthY, iSrcHeightY);

codec/processing/src/downsample/downsample.h+32−5

@@ -54,20 +54,29 @@ typedef void (HalveDownsampleFunc) (uint8_t* pDst, const int32_t kiDstStride, · 53 unmodified lines
54	54	uint8_t* pSrc, const int32_t kiSrcStride,
55	55	const int32_t kiSrcWidth, const int32_t kiSrcHeight);
56	56
	57	+typedef void (SpecificDownsampleFunc) (uint8_t* pDst, const int32_t kiDstStride,
	58	+ uint8_t* pSrc, const int32_t kiSrcStride,
	59	+ const int32_t kiSrcWidth, const int32_t kiHeight);
	60	+
57	61	typedef void (GeneralDownsampleFunc) (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth,
58	62	const int32_t kiDstHeight,
59	63	uint8_t* pSrc, const int32_t kiSrcStride, const int32_t kiSrcWidth, const int32_t kiSrcHeight);
60	64
61	65	typedef HalveDownsampleFunc* PHalveDownsampleFunc;
	66	+typedef SpecificDownsampleFunc* PSpecificDownsampleFunc;
62	67	typedef GeneralDownsampleFunc* PGeneralDownsampleFunc;
63	68
64		−HalveDownsampleFunc DyadicBilinearDownsampler_c;
	69	+HalveDownsampleFunc DyadicBilinearDownsampler_c;
65	70	GeneralDownsampleFunc GeneralBilinearFastDownsampler_c;
66	71	GeneralDownsampleFunc GeneralBilinearAccurateDownsampler_c;
	72	+SpecificDownsampleFunc DyadicBilinearOneThirdDownsampler_c;
	73	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_c;
67	74
68	75	typedef struct {
69	76	// align_index: 0 = x32; 1 = x16; 2 = x8; 3 = common case left;
70	77	PHalveDownsampleFunc pfHalfAverage[4];
	78	+ PSpecificDownsampleFunc pfOneThirdDownsampler;
	79	+ PSpecificDownsampleFunc pfQuarterDownsampler;
71	80	PGeneralDownsampleFunc pfGeneralRatioLuma;
72	81	PGeneralDownsampleFunc pfGeneralRatioChroma;
73	82	} SDownsampleFuncs;
@@ -93,10 +102,19 @@ HalveDownsampleFunc DyadicBilinearDownsamplerWidthx32_sse4; · 19 unmodified lines
93	102	GeneralDownsampleFunc GeneralBilinearFastDownsamplerWrap_sse2;
94	103	GeneralDownsampleFunc GeneralBilinearAccurateDownsamplerWrap_sse2;
95	104
	105	+SpecificDownsampleFunc DyadicBilinearOneThirdDownsampler_ssse3;
	106	+SpecificDownsampleFunc DyadicBilinearOneThirdDownsampler_sse4;
	107	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_sse;
	108	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_ssse3;
	109	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_sse4;
	110	+
96	111	void GeneralBilinearFastDownsampler_sse2 (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth,
97		− const int32_t kiDstHeight, uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX, const uint32_t kuiScaleY);
	112	+ const int32_t kiDstHeight, uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX,
	113	+ const uint32_t kuiScaleY);
98	114	void GeneralBilinearAccurateDownsampler_sse2 (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth,
99		− const int32_t kiDstHeight, uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX, const uint32_t kuiScaleY);
	115	+ const int32_t kiDstHeight, uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX,
	116	+ const uint32_t kuiScaleY);
	117	+
100	118	WELSVP_EXTERN_C_END
101	119	#endif
102	120
@@ -109,6 +127,10 @@ HalveDownsampleFunc DyadicBilinearDownsamplerWidthx32_neon; · 6 unmodified lines
109	127
110	128	GeneralDownsampleFunc GeneralBilinearAccurateDownsamplerWrap_neon;
111	129
	130	+SpecificDownsampleFunc DyadicBilinearOneThirdDownsampler_neon;
	131	+
	132	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_neon;
	133	+
112	134	void GeneralBilinearAccurateDownsampler_neon (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth,
113	135	const int32_t kiDstHeight,
114	136	uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX, const uint32_t kuiScaleY);
@@ -125,8 +147,13 @@ HalveDownsampleFunc DyadicBilinearDownsamplerWidthx32_AArch64_neon; · 10 unmodified lines
125	147
126	148	GeneralDownsampleFunc GeneralBilinearAccurateDownsamplerWrap_AArch64_neon;
127	149
128		−void GeneralBilinearAccurateDownsampler_AArch64_neon (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth, const int32_t kiDstHeight,
129		− uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX, const uint32_t kuiScaleY);
	150	+SpecificDownsampleFunc DyadicBilinearOneThirdDownsampler_AArch64_neon;
	151	+
	152	+SpecificDownsampleFunc DyadicBilinearQuarterDownsampler_AArch64_neon;
	153	+
	154	+void GeneralBilinearAccurateDownsampler_AArch64_neon (uint8_t* pDst, const int32_t kiDstStride,
	155	+ const int32_t kiDstWidth, const int32_t kiDstHeight,
	156	+ uint8_t* pSrc, const int32_t kiSrcStride, const uint32_t kuiScaleX, const uint32_t kuiScaleY);
130	157
131	158	WELSVP_EXTERN_C_END
132	159	#endif

codec/processing/src/downsample/downsamplefuncs.cpp+47−0

@@ -68,6 +68,53 @@ void DyadicBilinearDownsampler_c (uint8_t* pDst, const int32_t kiDstStride, · 67 unmodified lines
68	68	}
69	69	}
70	70
	71	+void DyadicBilinearQuarterDownsampler_c (uint8_t* pDst, const int32_t kiDstStride,
	72	+ uint8_t* pSrc, const int32_t kiSrcStride,
	73	+ const int32_t kiSrcWidth, const int32_t kiSrcHeight)
	74	+
	75	+{
	76	+ uint8_t* pDstLine = pDst;
	77	+ uint8_t* pSrcLine = pSrc;
	78	+ const int32_t kiSrcStridex4 = kiSrcStride << 2;
	79	+ const int32_t kiDstWidth = kiSrcWidth >> 2;
	80	+ const int32_t kiDstHeight = kiSrcHeight >> 2;
	81	+
	82	+ for (int32_t j = 0; j < kiDstHeight; j ++) {
	83	+ for (int32_t i = 0; i < kiDstWidth; i ++) {
	84	+ const int32_t kiSrcX = i << 2;
	85	+ const int32_t kiTempRow1 = (pSrcLine[kiSrcX] + pSrcLine[kiSrcX + 1] + 1) >> 1;
	86	+ const int32_t kiTempRow2 = (pSrcLine[kiSrcX + kiSrcStride] + pSrcLine[kiSrcX + kiSrcStride + 1] + 1) >> 1;
	87	+
	88	+ pDstLine[i] = (uint8_t) ((kiTempRow1 + kiTempRow2 + 1) >> 1);
	89	+ }
	90	+ pDstLine += kiDstStride;
	91	+ pSrcLine += kiSrcStridex4;
	92	+ }
	93	+}
	94	+
	95	+void DyadicBilinearOneThirdDownsampler_c (uint8_t* pDst, const int32_t kiDstStride,
	96	+ uint8_t* pSrc, const int32_t kiSrcStride,
	97	+ const int32_t kiSrcWidth, const int32_t kiDstHeight)
	98	+
	99	+{
	100	+ uint8_t* pDstLine = pDst;
	101	+ uint8_t* pSrcLine = pSrc;
	102	+ const int32_t kiSrcStridex3 = kiSrcStride * 3;
	103	+ const int32_t kiDstWidth = kiSrcWidth / 3;
	104	+
	105	+ for (int32_t j = 0; j < kiDstHeight; j ++) {
	106	+ for (int32_t i = 0; i < kiDstWidth; i ++) {
	107	+ const int32_t kiSrcX = i * 3;
	108	+ const int32_t kiTempRow1 = (pSrcLine[kiSrcX] + pSrcLine[kiSrcX + 1] + 1) >> 1;
	109	+ const int32_t kiTempRow2 = (pSrcLine[kiSrcX + kiSrcStride] + pSrcLine[kiSrcX + kiSrcStride + 1] + 1) >> 1;
	110	+
	111	+ pDstLine[i] = (uint8_t) ((kiTempRow1 + kiTempRow2 + 1) >> 1);
	112	+ }
	113	+ pDstLine += kiDstStride;
	114	+ pSrcLine += kiSrcStridex3;
	115	+ }
	116	+}
	117	+
71	118	void GeneralBilinearFastDownsampler_c (uint8_t* pDst, const int32_t kiDstStride, const int32_t kiDstWidth,
72	119	const int32_t kiDstHeight,
73	120	uint8_t* pSrc, const int32_t kiSrcStride, const int32_t kiSrcWidth, const int32_t kiSrcHeight) {

add new c and assembly functions to optimize downsampler when downscale equal 1:3/1:4

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code