Fix Stabilize_Video test for platforms that doesn't use fast color space conversion

[kxxt]

The Stabilize_Video test fails on platforms that doesn't support accelerated colorspace conversion from yuv420p to rgba.

See the ffmpeg code for more info: https://github.com/FFmpeg/FFmpeg/blob/e8e486332571347dd55822c842ba67276ac308e2/libswscale/yuv2rgb.c#L678-L696

This PR fixes it by rounding the result that caused the test failure.

On x86_64, y = -0.001074643496478603
On riscv64, y = -0.0009904620157904118

The difference mainly comes from sws_scale function in src/FFmpegReader.cpp. (FMA instructions in OpenCV's library caused some insignificant difference that we can ignore here)

The following patch can be used to save a sample of the resulting buffers:

diff --git a/src/FFmpegReader.cpp b/src/FFmpegReader.cpp
index 291f585c..eba954f0 100644
--- a/src/FFmpegReader.cpp
+++ b/src/FFmpegReader.cpp
@@ -16,6 +16,9 @@
 #include <thread>	// for std::this_thread::sleep_for
 #include <chrono>	// for std::chrono::milliseconds
 #include <unistd.h>
+#include <iostream>
+#include <fstream>
+using namespace std;
 
 #include "FFmpegUtilities.h"
 
@@ -1485,10 +1488,18 @@ void FFmpegReader::ProcessVideoPacket(int64_t requested_frame) {
 	SwsContext *img_convert_ctx = sws_getContext(info.width, info.height, AV_GET_CODEC_PIXEL_FORMAT(pStream, pCodecCtx), width,
 												 height, PIX_FMT_RGBA, scale_mode, NULL, NULL, NULL);
 
+	cerr << "Before sws_scale" << endl;
+	auto file = std::fstream("1st.before.bin", std::ios::out | std::ios::binary);
+	file.write((const char *)pFrame->data[0], pFrame->linesize[0]);
+	file.close();
 	// Resize / Convert to RGB
 	sws_scale(img_convert_ctx, pFrame->data, pFrame->linesize, 0,
 			  original_height, pFrameRGB->data, pFrameRGB->linesize);
-
+	cerr << "After sws_scale" << endl;
+	file = std::fstream("1st.after.bin", std::ios::out | std::ios::binary);
+	file.write((const char *)pFrameRGB->data[0], pFrameRGB->linesize[0]);
+	file.close();
+	abort();
 	// Create or get the existing frame object
 	std::shared_ptr<Frame> f = CreateFrame(current_frame);
 

On x86_64:

$ sha256sum build/tests/1st.before.bin 
7685abd4a249adefa3fdcb5307cb3eb987d3e28294d7c2e4b82daee2580c7450  build/tests/1st.before.bin

$ sha256sum build/tests/1st.after.bin 
4e29bc67f3aabcfdb8f23228e4195efcea9afd778520989a8713d957d0358d55  build/tests/1st.after.bin

On riscv64:

[root@kxxt2 src]# sha256sum build/tests/1st.before.bin
7685abd4a249adefa3fdcb5307cb3eb987d3e28294d7c2e4b82daee2580c7450  build/tests/1st.before.bin
[root@kxxt2 src]# sha256sum build/tests/1st.after.bin
0d3a7149c8785610ad1e6ecaa1431bf589fac4ff6b6809e773cc6093735f73dc  build/tests/1st.after.bin

Close #893

[XieJiSS]

The mentioned FMA behavior is also reproducible on x86_64 if compiled with -mfma flag.

[codecov]

Codecov Report

Merging #929 (e6f6b64) into develop (e91bd82) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff            @@
##           develop     #929   +/-   ##
========================================
  Coverage    53.66%   53.66%           
========================================
  Files          182      182           
  Lines        16829    16829           
========================================
  Hits          9031     9031           
  Misses        7798     7798           

Impacted Files	Coverage Δ
tests/CVStabilizer.cpp	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[jonoomph]

Thanks! LGTM