updates

apps/LoopThroughWithOpenGLCompositing/gl/OpenGLComposite.cpp

@@ -105,7 +105,8 @@ bool OpenGLComposite::InitVideoIO()
 		MessageBoxA(NULL, initFailureReason.c_str(), title, MB_OK | MB_ICONERROR);
 		return false;
 	}
-	if (!mVideoIO->SelectPreferredFormats(videoModes, initFailureReason))
+	const bool outputAlphaRequired = mRuntimeHost && mRuntimeHost->ExternalKeyingEnabled();
+	if (!mVideoIO->SelectPreferredFormats(videoModes, outputAlphaRequired, initFailureReason))
 		goto error;
 
 	if (! CheckOpenGLExtensions())

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.cpp

@@ -34,8 +34,8 @@ bool OpenGLRenderPipeline::RenderFrame(const RenderPipelineFrameContext& context
 	glBindFramebuffer(GL_FRAMEBUFFER, mRenderer.OutputFramebuffer());
 	if (mOutputReady)
 		mOutputReady();
-	if (state.outputPixelFormat == VideoIOPixelFormat::V210)
-		PackOutputForV210(state);
+	if (state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10)
+		PackOutputFor10Bit(state);
 	glFlush();
 
 	const auto renderEndTime = std::chrono::steady_clock::now();
@@ -50,7 +50,7 @@ bool OpenGLRenderPipeline::RenderFrame(const RenderPipelineFrameContext& context
 	return true;
 }
 
-void OpenGLRenderPipeline::PackOutputForV210(const VideoIOState& state)
+void OpenGLRenderPipeline::PackOutputFor10Bit(const VideoIOState& state)
 {
 	glBindFramebuffer(GL_FRAMEBUFFER, mRenderer.OutputPackFramebuffer());
 	glViewport(0, 0, state.outputPackTextureWidth, state.outputFrameSize.height);
@@ -64,10 +64,13 @@ void OpenGLRenderPipeline::PackOutputForV210(const VideoIOState& state)
 
 	const GLint outputResolutionLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uOutputVideoResolution");
 	const GLint activeWordsLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uActiveV210Words");
+	const GLint packFormatLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uOutputPackFormat");
 	if (outputResolutionLocation >= 0)
 		glUniform2f(outputResolutionLocation, static_cast<float>(state.outputFrameSize.width), static_cast<float>(state.outputFrameSize.height));
 	if (activeWordsLocation >= 0)
 		glUniform1f(activeWordsLocation, static_cast<float>(ActiveV210WordsForWidth(state.outputFrameSize.width)));
+	if (packFormatLocation >= 0)
+		glUniform1i(packFormatLocation, state.outputPixelFormat == VideoIOPixelFormat::Yuva10 ? 2 : 1);
 
 	glDrawArrays(GL_TRIANGLES, 0, 3);
 	glUseProgram(0);
@@ -79,7 +82,7 @@ void OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOut
 {
 	glPixelStorei(GL_PACK_ALIGNMENT, 4);
 	glPixelStorei(GL_PACK_ROW_LENGTH, 0);
-	if (state.outputPixelFormat == VideoIOPixelFormat::V210)
+	if (state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10)
 	{
 		glBindFramebuffer(GL_READ_FRAMEBUFFER, mRenderer.OutputPackFramebuffer());
 		glReadPixels(0, 0, state.outputPackTextureWidth, state.outputFrameSize.height, GL_RGBA, GL_UNSIGNED_BYTE, outputFrame.bytes);

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.h

@@ -30,7 +30,7 @@ public:
 	bool RenderFrame(const RenderPipelineFrameContext& context, VideoIOOutputFrame& outputFrame);
 
 private:
-	void PackOutputForV210(const VideoIOState& state);
+	void PackOutputFor10Bit(const VideoIOState& state);
 	void ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame);
 
 	OpenGLRenderer& mRenderer;

apps/LoopThroughWithOpenGLCompositing/gl/shader/GlShaderSources.cpp

@@ -90,12 +90,17 @@ const char* kOutputPackFragmentShaderSource =
 	"layout(binding = 0) uniform sampler2D uOutputRgb;\n"
 	"uniform vec2 uOutputVideoResolution;\n"
 	"uniform float uActiveV210Words;\n"
+	"uniform int uOutputPackFormat;\n"
 	"in vec2 vTexCoord;\n"
 	"layout(location = 0) out vec4 fragColor;\n"
-	"vec3 rgbAt(int x, int y)\n"
+	"vec4 rgbaAt(int x, int y)\n"
 	"{\n"
 	"	ivec2 size = ivec2(max(uOutputVideoResolution, vec2(1.0, 1.0)));\n"
-	"	return clamp(texelFetch(uOutputRgb, ivec2(clamp(x, 0, size.x - 1), clamp(y, 0, size.y - 1)), 0).rgb, vec3(0.0), vec3(1.0));\n"
+	"	return clamp(texelFetch(uOutputRgb, ivec2(clamp(x, 0, size.x - 1), clamp(y, 0, size.y - 1)), 0), vec4(0.0), vec4(1.0));\n"
+	"}\n"
+	"vec3 rgbAt(int x, int y)\n"
+	"{\n"
+	"	return rgbaAt(x, y).rgb;\n"
 	"}\n"
 	"vec3 rgbToLegalYcbcr10(vec3 rgb)\n"
 	"{\n"
@@ -112,9 +117,35 @@ const char* kOutputPackFragmentShaderSource =
 	"{\n"
 	"	return vec4(float(word & 255u), float((word >> 8) & 255u), float((word >> 16) & 255u), float((word >> 24) & 255u)) / 255.0;\n"
 	"}\n"
+	"vec4 bigEndianWordToBytes(uint word)\n"
+	"{\n"
+	"	return vec4(float((word >> 24) & 255u), float((word >> 16) & 255u), float((word >> 8) & 255u), float(word & 255u)) / 255.0;\n"
+	"}\n"
+	"vec4 packAy10Word(ivec2 outCoord)\n"
+	"{\n"
+	"	ivec2 size = ivec2(max(uOutputVideoResolution, vec2(1.0, 1.0)));\n"
+	"	if (outCoord.x >= size.x)\n"
+	"		return vec4(0.0);\n"
+	"	int pixelBase = (outCoord.x / 2) * 2;\n"
+	"	int y = outCoord.y;\n"
+	"	vec4 rgba0 = rgbaAt(pixelBase + 0, y);\n"
+	"	vec4 rgba1 = rgbaAt(pixelBase + 1, y);\n"
+	"	vec3 c0 = rgbToLegalYcbcr10(rgba0.rgb);\n"
+	"	vec3 c1 = rgbToLegalYcbcr10(rgba1.rgb);\n"
+	"	float chroma = (outCoord.x & 1) == 0 ? round((c0.y + c1.y) * 0.5) : round((c0.z + c1.z) * 0.5);\n"
+	"	float alpha = round(clamp(((outCoord.x & 1) == 0 ? rgba0.a : rgba1.a), 0.0, 1.0) * 1023.0);\n"
+	"	float luma = (outCoord.x & 1) == 0 ? c0.x : c1.x;\n"
+	"	uint word = ((uint(luma) & 1023u) << 22) | ((uint(chroma) & 1023u) << 12) | ((uint(alpha) & 1023u) << 2);\n"
+	"	return bigEndianWordToBytes(word);\n"
+	"}\n"
 	"void main()\n"
 	"{\n"
 	"	ivec2 outCoord = ivec2(gl_FragCoord.xy);\n"
+	"	if (uOutputPackFormat == 2)\n"
+	"	{\n"
+	"		fragColor = packAy10Word(outCoord);\n"
+	"		return;\n"
+	"	}\n"
 	"	if (float(outCoord.x) >= uActiveV210Words)\n"
 	"	{\n"
 	"		fragColor = vec4(0.0);\n"