<div dir="ltr"><br><div class="gmail_extra"><br><br><div class="gmail_quote">On Wed, Jul 17, 2013 at 2:36 AM, <span dir="ltr"><<a href="mailto:gopu@multicorewareinc.com" target="_blank">gopu@multicorewareinc.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"># HG changeset patch<br>
# User ggopu<br>
# Date 1374046566 -19800<br>
# Node ID c6c045e9272798ce5268643436759db22fd1950b<br>
# Parent 054d8c409569100c4aacb015ffb1b3281100d993<br>
TShortYUV : Implemented perfomance Primitives pixeladd (Clipadd)<br>
<br>
diff -r 054d8c409569 -r c6c045e92727 source/common/TShortYUV.cpp<br>
--- a/source/common/TShortYUV.cpp Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/common/TShortYUV.cpp Wed Jul 17 13:06:06 2013 +0530<br>
@@ -122,61 +122,38 @@<br>
#pragma warning (disable: 4244)<br>
#endif<br>
<br>
-void TShortYUV::addClipLuma(TShortYUV* pcYuvSrc0, TShortYUV* pcYuvSrc1, unsigned int uiTrUnitIdx, unsigned int uiPartSize)<br>
+void TShortYUV::addClipLuma(TShortYUV* pcYuvSrc0, TShortYUV* pcYuvSrc1, unsigned int trUnitIdx, unsigned int partSize)<br></blockquote><div><br></div><div>pcYuvSrc0 -> srcYuv0, pcYuvSrc1 -> srcYuv1</div><div><br></div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
{<br>
- int x, y;<br>
+ int x = partSize, y = partSize;<br></blockquote><div><br></div><div>only need one var here (in the other places too)</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
- short* pSrc0 = pcYuvSrc0->getLumaAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pSrc1 = pcYuvSrc1->getLumaAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pDst = getLumaAddr(uiTrUnitIdx, uiPartSize);<br>
+ short* src0 = pcYuvSrc0->getLumaAddr(trUnitIdx, partSize);<br>
+ short* src1 = pcYuvSrc1->getLumaAddr(trUnitIdx, partSize);<br>
+ short* dst = getLumaAddr(trUnitIdx, partSize);<br></blockquote><div><br></div><div>TShortYUV will probably need to be changed to UShort soon (globally) because I am going to change 16bpp Pel to UShort to remove the impedence mismatch with x264's pixel type.</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
- unsigned int iSrc0Stride = pcYuvSrc0->width;<br>
- unsigned int iSrc1Stride = pcYuvSrc1->width;<br>
- unsigned int iDstStride = width;<br>
+ unsigned int src0Stride = pcYuvSrc0->width;<br>
+ unsigned int src1Stride = pcYuvSrc1->width;<br>
+ unsigned int dstStride = width;<br></blockquote><div><br></div><div>To be pedantic, these should be intptr_t</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
- for (y = uiPartSize - 1; y >= 0; y--)<br>
- {<br>
- for (x = uiPartSize - 1; x >= 0; x--)<br>
- {<br>
- pDst[x] = ClipY(pSrc0[x] + pSrc1[x]);<br>
- }<br>
-<br>
- pSrc0 += iSrc0Stride;<br>
- pSrc1 += iSrc1Stride;<br>
- pDst += iDstStride;<br>
- }<br>
+ primitives.pixeladd(x, y, dst, dstStride, src0, src1, src0Stride, src1Stride);<br>
}<br>
<br>
-void TShortYUV::addClipChroma(TShortYUV* pcYuvSrc0, TShortYUV* pcYuvSrc1, unsigned int uiTrUnitIdx, unsigned int uiPartSize)<br>
+void TShortYUV::addClipChroma(TShortYUV* pcYuvSrc0, TShortYUV* pcYuvSrc1, unsigned int trUnitIdx, unsigned int partSize)<br>
{<br>
- int x, y;<br>
+ int x = partSize, y = partSize;<br>
<br>
- short* pSrcU0 = pcYuvSrc0->getCbAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pSrcU1 = pcYuvSrc1->getCbAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pSrcV0 = pcYuvSrc0->getCrAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pSrcV1 = pcYuvSrc1->getCrAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pDstU = getCbAddr(uiTrUnitIdx, uiPartSize);<br>
- short* pDstV = getCrAddr(uiTrUnitIdx, uiPartSize);<br>
+ short* srcU0 = pcYuvSrc0->getCbAddr(trUnitIdx, partSize);<br>
+ short* srcU1 = pcYuvSrc1->getCbAddr(trUnitIdx, partSize);<br>
+ short* srcV0 = pcYuvSrc0->getCrAddr(trUnitIdx, partSize);<br>
+ short* srcV1 = pcYuvSrc1->getCrAddr(trUnitIdx, partSize);<br>
+ short* dstU = getCbAddr(trUnitIdx, partSize);<br>
+ short* dstV = getCrAddr(trUnitIdx, partSize);<br>
<br>
- unsigned int iSrc0Stride = pcYuvSrc0->Cwidth;<br>
- unsigned int iSrc1Stride = pcYuvSrc1->Cwidth;<br>
- unsigned int iDstStride = Cwidth;<br>
+ unsigned int src0Stride = pcYuvSrc0->Cwidth;<br>
+ unsigned int src1Stride = pcYuvSrc1->Cwidth;<br>
+ unsigned int dstStride = Cwidth;<br>
<br>
- for (y = uiPartSize - 1; y >= 0; y--)<br>
- {<br>
- for (x = uiPartSize - 1; x >= 0; x--)<br>
- {<br>
- pDstU[x] = ClipC(pSrcU0[x] + pSrcU1[x]);<br>
- pDstV[x] = ClipC(pSrcV0[x] + pSrcV1[x]);<br>
- }<br>
-<br>
- pSrcU0 += iSrc0Stride;<br>
- pSrcU1 += iSrc1Stride;<br>
- pSrcV0 += iSrc0Stride;<br>
- pSrcV1 += iSrc1Stride;<br>
- pDstU += iDstStride;<br>
- pDstV += iDstStride;<br>
- }<br>
+ primitives.pixeladd(x, y, dstU, dstStride, srcU0, srcU1, src0Stride, src1Stride);<br>
+ primitives.pixeladd(x, y, dstV, dstStride, srcV0, srcV1, src0Stride, src1Stride);<br>
}<br>
<br>
#if _MSC_VER<br>
diff -r 054d8c409569 -r c6c045e92727 source/common/pixel.cpp<br>
--- a/source/common/pixel.cpp Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/common/pixel.cpp Wed Jul 17 13:06:06 2013 +0530<br>
@@ -30,7 +30,6 @@<br>
#include <algorithm><br>
#include <stdlib.h> // abs()<br></blockquote><div><br></div><div>unrelated white-space changes, should be in a separate patch</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
-<br>
#define SET_FUNC_PRIMITIVE_TABLE_C_SUBSET(WIDTH, FUNC_PREFIX, FUNC_PREFIX_DEF, FUNC_TYPE_CAST, DATA_TYPE1, DATA_TYPE2) \<br>
p.FUNC_PREFIX[PARTITION_ ## WIDTH ## x4] = (FUNC_TYPE_CAST)FUNC_PREFIX_DEF<WIDTH, 4, DATA_TYPE1, DATA_TYPE2>; \<br>
p.FUNC_PREFIX[PARTITION_ ## WIDTH ## x8] = (FUNC_TYPE_CAST)FUNC_PREFIX_DEF<WIDTH, 8, DATA_TYPE1, DATA_TYPE2>; \<br>
@@ -504,14 +503,15 @@<br>
void weightUnidir(short *src, pixel *dst, int srcStride, int dstStride, int width, int height, int w0, int round, int shift, int offset, int bitDepth)<br>
{<br>
int x, y;<br></blockquote><div><br></div><div>ditto</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+<br>
for (y = height - 1; y >= 0; y--)<br>
{<br>
for (x = width - 1; x >= 0; )<br>
{<br>
// note: luma min width is 4<br>
- dst[x] = (pixel) Clip3(0, ((1 << bitDepth) - 1), ((w0 * (src[x] + IF_INTERNAL_OFFS) + round) >> shift) + offset);<br>
+ dst[x] = (pixel)Clip3(0, ((1 << bitDepth) - 1), ((w0 * (src[x] + IF_INTERNAL_OFFS) + round) >> shift) + offset);<br>
x--;<br>
- dst[x] = (pixel) Clip3(0, ((1 << bitDepth) - 1), ((w0 * (src[x] + IF_INTERNAL_OFFS) + round) >> shift) + offset);<br>
+ dst[x] = (pixel)Clip3(0, ((1 << bitDepth) - 1), ((w0 * (src[x] + IF_INTERNAL_OFFS) + round) >> shift) + offset);<br>
x--;<br>
}<br>
<br>
@@ -535,6 +535,20 @@<br>
}<br>
}<br></blockquote><div><br></div><div>non-templated C primitives should end with _c</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+void pixeladd_ss(int bx, int by, short *a, intptr_t dstride, short *b0, short *b1, intptr_t sstride0, intptr_t sstride1)<br>
+{<br>
+ for (int y = 0; y < by; y++)<br>
+ {<br>
+ for (int x = 0; x < bx; x++)<br>
+ {<br>
+ a[x] = (short)ClipY(b0[x] + b1[x]);<br>
+ }<br>
+<br>
+ b0 += sstride0;<br>
+ b1 += sstride1;<br>
+ a += dstride;<br>
+ }<br>
+}<br>
} // end anonymous namespace<br>
<br>
namespace x265 {<br>
@@ -738,5 +752,6 @@<br>
p.weightpUni = weightUnidir;<br>
<br>
p.pixelsubsp = pixelsub_sp;<br>
+ p.pixeladd = pixeladd_ss;<br></blockquote><div><br></div><div>if the function requires shorts, presumably there is a TComPiicYuv which does the same operation using uint8_t, so we will want a version that supports 8bit pixels, so the primitive funcdef should be named p.pixeladd_pp, and this particular version should only be used for high-bit-depth compiles.</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
}<br>
}<br>
diff -r 054d8c409569 -r c6c045e92727 source/common/primitives.h<br>
--- a/source/common/primitives.h Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/common/primitives.h Wed Jul 17 13:06:06 2013 +0530<br>
@@ -215,7 +215,8 @@<br>
typedef void (*dequant_t)(int bitDepth, const int* src, int* dst, int width, int height, int mcqp_miper, int mcqp_mirem, bool useScalingList, unsigned int trSizeLog2, int *dequantCoef);<br>
typedef uint32_t (*quant_t)(int *coef, int *quantCoeff, int *deltaU, int *qCoef, int qBits, int add, int numCoeff);<br>
typedef void (*weightpUni_t)(short *src, pixel *dst, int srcStride, int dstStride, int width, int height, int w0, int round, int shift, int offset, int bitDepth);<br>
-typedef void (*pixelsub_sp_t)(int bx, int by, short *dst, intptr_t dstride, pixel *src0, pixel *src1, intptr_t sstride0, intptr_t sstride1);<br>
+typedef void (*pixelsub_sp_t)(int bx, int by, short *dst, intptr_t dstride, pixel *src0, pixel *src1, intptr_t sstride0, intptr_t sstride1);<br>
+typedef void (*pixeladd_t)(int bx, int by, short *dst, intptr_t dstride, short *src0, short *src1, intptr_t sstride0, intptr_t sstride1);<br>
<br>
/* Define a structure containing function pointers to optimized encoder<br>
* primitives. Each pointer can reference either an assembly routine,<br>
@@ -267,6 +268,7 @@<br>
<br>
weightpUni_t weightpUni;<br>
pixelsub_sp_t pixelsubsp;<br>
+ pixeladd_t pixeladd;<br>
};<br>
<br>
/* This copy of the table is what gets used by the encoder.<br>
diff -r 054d8c409569 -r c6c045e92727 source/common/vec/blockcopy.inc<br>
--- a/source/common/vec/blockcopy.inc Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/common/vec/blockcopy.inc Wed Jul 17 13:06:06 2013 +0530<br>
@@ -79,7 +79,7 @@<br>
}<br>
}<br>
else<br></blockquote><div><br></div><div>unrelated changes</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
-#endif<br>
+#endif /* if INSTRSET >= 8 */<br>
if (!(aligncheck & 15))<br>
{<br>
// fast path, multiples of 16 pixel wide blocks<br>
@@ -131,7 +131,7 @@<br>
}<br>
}<br>
else<br>
-#endif<br>
+#endif /* if INSTRSET >= 8 && 0 */<br>
if (!(aligncheck & 15))<br>
{<br>
// fast path, multiples of 16 pixel wide blocks<br>
@@ -170,6 +170,7 @@<br>
void blockcopy_s_p(int bx, int by, short *dst, intptr_t dstride, uint8_t *src, intptr_t sstride)<br>
{<br>
size_t aligncheck = (size_t)dst | (size_t)src | bx | sstride | dstride;<br>
+<br>
#if INSTRSET >= 8 && 0<br>
if (!(aligncheck & 31))<br>
{<br>
@@ -189,7 +190,7 @@<br>
}<br>
}<br>
else<br>
-#endif<br>
+#endif /* if INSTRSET >= 8 && 0 */<br>
if (!(aligncheck & 15))<br>
{<br>
// fast path, multiples of 16 pixel wide blocks<br>
@@ -292,6 +293,84 @@<br>
}<br>
}<br>
<br>
+void pixeladd_ss(int bx, int by, short *dst, intptr_t dstride, short *src0, short *src1, intptr_t sstride0, intptr_t sstride1)<br>
+{<br>
+ size_t aligncheck = (size_t)dst | (size_t)src0 | bx | sstride0 | dstride;<br></blockquote><div><br></div><div>alignment restriction for shorts are more complicated. the pointers and strides must be 32 byte aligned for AVX2 but the width only needs to be a multiple of 16</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+<br>
+#if INSTRSET >= 8 && 0<br>
+ if (!(aligncheck & 31))<br>
+ {<br>
+ // fast path, multiples of 32 pixel wide blocks<br>
+ // fast path, multiples of 16 pixel wide blocks<br></blockquote><div><br></div><div>?</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ for (int y = 0; y < by; y++)<br>
+ {<br>
+ for (int x = 0; x < bx; x += 8)<br>
+ {<br>
+ Vec8s vecsrc0, vecsrc1, vecsum;<br>
+ Vec8s zero(0), maxval((1 << 8) - 1); // Currently g_bitDepthY = 8 and g_bitDepthC = 8<br></blockquote><div><br></div><div>let's not add another bomb that awaits us when we want to use larger bit-depths. either use g_bitDepthY directly here or pass in the bit-d</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ vecsrc0.load_a(src0 + x);<br>
+ vecsrc1.load_a(src1 + x);<br>
+<br>
+ vecsum = vecsrc0 + vecsrc1;<br>
+ vecsum = max(vecsum, zero);<br>
+ vecsum = min(vecsum, maxval);<br>
+<br>
+ vecsum.store(dst + x);<br>
+ }<br>
+<br>
+ src0 += sstride0;<br>
+ src1 += sstride1;<br>
+ dst += dstride;<br>
+ }<br>
+ }<br>
+ else<br>
+#endif /* if INSTRSET >= 8 && 0 */<br>
+ if (!(aligncheck & 15))<br>
+ {<br>
+ // fast path, multiples of 16 pixel wide blocks<br>
+ for (int y = 0; y < by; y++)<br>
+ {<br>
+ for (int x = 0; x < bx; x += 8)<br>
+ {<br>
+ Vec8s vecsrc0, vecsrc1, vecsum;<br>
+ Vec8s zero(0), maxval((1 << 8) - 1); // Currently g_bitDepthY = 8 and g_bitDepthC = 8<br>
+ vecsrc0.load_a(src0 + x);<br>
+ vecsrc1.load_a(src1 + x);<br>
+<br>
+ vecsum = add_saturated(vecsrc0, vecsrc1);<br>
+ vecsum = max(vecsum, zero);<br>
+ vecsum = min(vecsum, maxval);<br>
+<br>
+ vecsum.store(dst + x);<br>
+ }<br>
+<br>
+ src0 += sstride0;<br>
+ src1 += sstride1;<br>
+ dst += dstride;<br>
+ }<br>
+ }<br>
+ else<br>
+ {<br>
+ int tmp;<br>
+ int max = (1 << 8) - 1;<br>
+ // slow path, irregular memory alignments or sizes<br>
+ for (int y = 0; y < by; y++)<br>
+ {<br>
+ for (int x = 0; x < bx; x++)<br>
+ {<br>
+ tmp = src0[x] + src1[x];<br>
+ tmp = tmp < 0 ? 0 : tmp;<br>
+ tmp = tmp > max ? max : tmp;<br>
+ dst[x] = (short)tmp;<br>
+ }<br>
+<br>
+ src0 += sstride0;<br>
+ src1 += sstride1;<br>
+ dst += dstride;<br>
+ }<br>
+ }<br>
+}<br>
+<br>
void Setup_Vec_BlockCopyPrimitives(EncoderPrimitives &p)<br>
{<br>
#if HIGH_BIT_DEPTH<br>
@@ -307,5 +386,6 @@<br>
p.blockcpy_sp = blockcopy_s_p;<br>
p.blockcpy_sc = blockcopy_s_p;<br>
p.pixelsubsp = pixelsub_sp;<br>
-#endif<br>
+ p.pixeladd = pixeladd_ss;<br>
+#endif /* if HIGH_BIT_DEPTH */<br>
}<br>
diff -r 054d8c409569 -r c6c045e92727 source/test/pixelharness.cpp<br>
--- a/source/test/pixelharness.cpp Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/test/pixelharness.cpp Wed Jul 17 13:06:06 2013 +0530<br>
@@ -376,6 +376,29 @@<br>
return true;<br>
}<br>
<br>
+bool PixelHarness::check_pixeladd_sp(x265::pixeladd_t ref, x265::pixeladd_t opt)<br>
+{<br>
+ ALIGN_VAR_16(short, ref_dest[64 * 64]);<br>
+ ALIGN_VAR_16(short, opt_dest[64 * 64]);<br>
+ int bx = 64;<br>
+ int by = 64;<br>
+ int j = 0;<br>
+ for (int i = 0; i <= 100; i++)<br>
+ {<br>
+ opt(bx, by, opt_dest, 64, (short*)pbuf2 + j, (short*)pbuf1 + j, 128, 128);<br>
+ ref(bx, by, ref_dest, 64, (short*)pbuf2 + j, (short*)pbuf1 + j, 128, 128);<br>
+<br>
+ if (memcmp(ref_dest, opt_dest, 64 * 64 * sizeof(short)))<br>
+ return false;<br>
+<br>
+ j += 4;<br>
+ bx = 4 * ((rand() & 15) + 1);<br>
+ by = 4 * ((rand() & 15) + 1);<br>
+ }<br>
+<br>
+ return true;<br>
+}<br>
+<br>
bool PixelHarness::testCorrectness(const EncoderPrimitives& ref, const EncoderPrimitives& opt)<br>
{<br>
for (uint16_t curpar = 0; curpar < NUM_PARTITIONS; curpar++)<br>
@@ -535,6 +558,14 @@<br>
}<br>
}<br>
<br>
+ if (opt.pixeladd)<br>
+ {<br>
+ if (!check_pixeladd_sp(ref.pixeladd, opt.pixeladd))<br>
+ {<br>
+ printf("Pixel Add failed!\n");<br></blockquote><div><br></div><div>upper case here not necessary</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ return false;<br>
+ }<br>
+ }<br>
return true;<br>
}<br>
<br>
@@ -649,4 +680,10 @@<br>
printf("Pixel Sub");<br></blockquote><div><br></div><div>upper case here not necessary</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
REPORT_SPEEDUP(opt.pixelsubsp, ref.pixelsubsp, 64, 64, (short*)pbuf1, FENC_STRIDE, pbuf2, pbuf1, STRIDE, STRIDE);<br>
}<br>
+<br>
+ if (opt.pixeladd)<br>
+ {<br>
+ printf("Pixel Add");<br>
+ REPORT_SPEEDUP(opt.pixeladd, ref.pixeladd, 64, 64, (short*)pbuf1, FENC_STRIDE, (short*)pbuf2, (short*)pbuf1, STRIDE, STRIDE);<br>
+ }<br>
}<br>
diff -r 054d8c409569 -r c6c045e92727 source/test/pixelharness.h<br>
--- a/source/test/pixelharness.h Wed Jul 17 11:33:20 2013 +0530<br>
+++ b/source/test/pixelharness.h Wed Jul 17 13:06:06 2013 +0530<br>
@@ -48,6 +48,7 @@<br>
bool check_calcrecon(x265::calcrecon_t ref, x265::calcrecon_t opt);<br>
bool check_weightpUni(x265::weightpUni_t ref, x265::weightpUni_t opt);<br>
bool check_pixelsub_sp(x265::pixelsub_sp_t ref, x265::pixelsub_sp_t opt);<br>
+ bool check_pixeladd_sp(x265::pixeladd_t ref, x265::pixeladd_t opt);<br>
<br>
public:<br>
<br>
_______________________________________________<br>
x265-devel mailing list<br>
<a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br>
<a href="http://mailman.videolan.org/listinfo/x265-devel" target="_blank">http://mailman.videolan.org/listinfo/x265-devel</a><br>
</blockquote></div><br><br clear="all"><div><br></div>-- <br>Steve Borho
</div></div>