shithub: openh264

Info  •  Files  •  Log  •  Branches

Use WELS_ROUND in the neon version of GeneralBilinearAccurateDownsampler

Also use it in the commented out sse2 version - in case it gets
taken into use at some point. (It only works on 32 bit x86
at the moment.)

This makes sure the scaling factors are rounded similarly to the
in GeneralBilinearAccurateDownsampler_c, making sure the
accelerated versions of these function return identical results
to the C version.

--- a/codec/processing/src/downsample/downsamplefuncs.cpp

+++ b/codec/processing/src/downsample/downsamplefuncs.cpp

@@ -208,8 +208,8 @@

 //  const int32_t kiScaleBitWidth = 16, kiScaleBitHeight = 15;

 //  const uint32_t kuiScaleWidth = (1 << kiScaleBitWidth), kuiScaleHeight = (1 << kiScaleBitHeight);

 //

-//  uint32_t uiScalex = (uint32_t) ((float)kiSrcWidth / (float)kiDstWidth * kuiScaleWidth);

-//  uint32_t uiScaley = (uint32_t) ((float)kiSrcHeight / (float)kiDstHeight * kuiScaleHeight);

+//  uint32_t uiScalex = WELS_ROUND ((float)kiSrcWidth / (float)kiDstWidth * kuiScaleWidth);

+//  uint32_t uiScaley = WELS_ROUND ((float)kiSrcHeight / (float)kiDstHeight * kuiScaleHeight);

 //

 //  GeneralBilinearFastDownsampler_sse2 (pDst, kiDstStride, kiDstWidth, kiDstHeight,

 //                                       pSrc, kiSrcStride, kiSrcWidth, kiSrcHeight, uiScalex, uiScaley);

@@ -221,8 +221,8 @@

 //  const int32_t kiScaleBit = 15;

 //  const uint32_t kuiScale = (1 << kiScaleBit);

 //

-//  uint32_t uiScalex = (uint32_t) ((float)kiSrcWidth / (float)kiDstWidth * kuiScale);

-//  uint32_t uiScaley = (uint32_t) ((float)kiSrcHeight / (float)kiDstHeight * kuiScale);

+//  uint32_t uiScalex = WELS_ROUND ((float)kiSrcWidth / (float)kiDstWidth * kuiScale);

+//  uint32_t uiScaley = WELS_ROUND ((float)kiSrcHeight / (float)kiDstHeight * kuiScale);

 //

 //  GeneralBilinearAccurateDownsampler_sse2 (pDst, kiDstStride, kiDstWidth, kiDstHeight,

 //      pSrc, kiSrcStride, kiSrcWidth, kiSrcHeight, uiScalex, uiScaley);

@@ -235,8 +235,8 @@

     uint8_t* pSrc, const int32_t kiSrcStride, const int32_t kiSrcWidth, const int32_t kiSrcHeight) {

   const int32_t kiScaleBit = 15;

   const uint32_t kuiScale = (1 << kiScaleBit);

-  uint32_t uiScalex = (uint32_t) ((float)kiSrcWidth / (float)kiDstWidth * kuiScale);

-  uint32_t uiScaley = (uint32_t) ((float)kiSrcHeight / (float)kiDstHeight * kuiScale);

+  uint32_t uiScalex = WELS_ROUND ((float)kiSrcWidth / (float)kiDstWidth * kuiScale);

+  uint32_t uiScaley = WELS_ROUND ((float)kiSrcHeight / (float)kiDstHeight * kuiScale);

   GeneralBilinearAccurateDownsampler_neon (pDst, kiDstStride, kiDstWidth, kiDstHeight, pSrc, kiSrcStride, uiScalex,

       uiScaley);

 }