Vulkan: Make scaling shaders compatible + fixes (#1392)

src/Cafe/GraphicPack/GraphicPack2.cpp

@@ -960,7 +960,7 @@ bool GraphicPack2::Activate()
 				auto option_upscale = rules.FindOption("upscaleMagFilter");
 				if(option_upscale && boost::iequals(*option_upscale, "NearestNeighbor"))
 					m_output_settings.upscale_filter = LatteTextureView::MagFilter::kNearestNeighbor;
-				auto option_downscale = rules.FindOption("NearestNeighbor");
+				auto option_downscale = rules.FindOption("downscaleMinFilter");
 				if (option_downscale && boost::iequals(*option_downscale, "NearestNeighbor"))
 					m_output_settings.downscale_filter = LatteTextureView::MagFilter::kNearestNeighbor;
 			}

src/Cafe/HW/Latte/Core/LatteRenderTarget.cpp

@@ -934,13 +934,6 @@ void LatteRenderTarget_copyToBackbuffer(LatteTextureView* textureView, bool isPa
 	{
 		sint32 scaling_filter = downscaling ? GetConfig().downscale_filter : GetConfig().upscale_filter;
 
-		if (g_renderer->GetType() == RendererAPI::Vulkan)
-		{
-			// force linear or nearest neighbor filter
-			if(scaling_filter != kLinearFilter && scaling_filter != kNearestNeighborFilter)
-				scaling_filter = kLinearFilter;
-		}
-
 		if (scaling_filter == kLinearFilter)
 		{
 			if(renderUpsideDown)
@@ -957,7 +950,7 @@ void LatteRenderTarget_copyToBackbuffer(LatteTextureView* textureView, bool isPa
 			else
 				shader = RendererOutputShader::s_bicubic_shader;
 
-			filter = LatteTextureView::MagFilter::kNearestNeighbor;
+			filter = LatteTextureView::MagFilter::kLinear;
 		}
 		else if (scaling_filter == kBicubicHermiteFilter)
 		{

src/Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.cpp

@@ -570,13 +570,10 @@ void OpenGLRenderer::DrawBackbufferQuad(LatteTextureView* texView, RendererOutpu
 		g_renderer->ClearColorbuffer(padView);
 	}
 
-	sint32 effectiveWidth, effectiveHeight;
-	texView->baseTexture->GetEffectiveSize(effectiveWidth, effectiveHeight, 0);
-
 	shader_unbind(RendererShader::ShaderType::kGeometry);
 	shader_bind(shader->GetVertexShader());
 	shader_bind(shader->GetFragmentShader());
-	shader->SetUniformParameters(*texView, { effectiveWidth, effectiveHeight }, { imageWidth, imageHeight });
+	shader->SetUniformParameters(*texView, {imageWidth, imageHeight});
 
 	// set viewport
 	glViewportIndexedf(0, imageX, imageY, imageWidth, imageHeight);
@@ -584,6 +581,12 @@ void OpenGLRenderer::DrawBackbufferQuad(LatteTextureView* texView, RendererOutpu
 	LatteTextureViewGL* texViewGL = (LatteTextureViewGL*)texView;
 	texture_bindAndActivate(texView, 0);
 
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+	texViewGL->samplerState.clampS = texViewGL->samplerState.clampT = 0xFF;
+
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, useLinearTexFilter ? GL_LINEAR : GL_NEAREST);
+	texViewGL->samplerState.filterMin = 0xFFFFFFFF;
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, useLinearTexFilter ? GL_LINEAR : GL_NEAREST);
 	texViewGL->samplerState.filterMag = 0xFFFFFFFF;
 

src/Cafe/HW/Latte/Renderer/RendererOuputShader.cpp

@@ -2,18 +2,7 @@
 #include "Cafe/HW/Latte/Renderer/OpenGL/OpenGLRenderer.h"
 
 const std::string RendererOutputShader::s_copy_shader_source =
-R"(#version 420
+R"(
-
-#ifdef VULKAN
-layout(location = 0) in vec2 passUV;
-layout(binding = 0) uniform sampler2D textureSrc;
-layout(location = 0) out vec4 colorOut0;
-#else
-in vec2 passUV;
-layout(binding=0) uniform sampler2D textureSrc;
-layout(location = 0) out vec4 colorOut0;
-#endif
-
 void main()
 {
 	colorOut0 = vec4(texture(textureSrc, passUV).rgb,1.0);
@@ -22,20 +11,6 @@ void main()
 
 const std::string RendererOutputShader::s_bicubic_shader_source =
 R"(
-#version 420
-
-#ifdef VULKAN
-layout(location = 0) in vec2 passUV;
-layout(binding = 0)  uniform sampler2D textureSrc;
-layout(binding = 1)  uniform vec2 textureSrcResolution;
-layout(location = 0) out vec4 colorOut0;
-#else
-in vec2 passUV;
-layout(binding=0) uniform sampler2D textureSrc;
-uniform vec2 textureSrcResolution;
-layout(location = 0) out vec4 colorOut0;
-#endif
-
 vec4 cubic(float x)
 {
 	float x2 = x * x;
@@ -48,24 +23,23 @@ vec4 cubic(float x)
 	return w / 6.0;
 }
 
-vec4 bcFilter(vec2 texcoord, vec2 texscale)
+vec4 bcFilter(vec2 uv, vec4 texelSize)
 {
-	float fx = fract(texcoord.x);
+	vec2 pixel = uv*texelSize.zw - 0.5;
-	float fy = fract(texcoord.y);
+	vec2 pixelFrac = fract(pixel);
-	texcoord.x -= fx;
+	vec2 pixelInt = pixel - pixelFrac;
-	texcoord.y -= fy;
 
-	vec4 xcubic = cubic(fx);
+	vec4 xcubic = cubic(pixelFrac.x);
-	vec4 ycubic = cubic(fy);
+	vec4 ycubic = cubic(pixelFrac.y);
 
-	vec4 c = vec4(texcoord.x - 0.5, texcoord.x + 1.5, texcoord.y - 0.5, texcoord.y + 1.5);
+	vec4 c = vec4(pixelInt.x - 0.5, pixelInt.x + 1.5, pixelInt.y - 0.5, pixelInt.y + 1.5);
 	vec4 s = vec4(xcubic.x + xcubic.y, xcubic.z + xcubic.w, ycubic.x + ycubic.y, ycubic.z + ycubic.w);
 	vec4 offset = c + vec4(xcubic.y, xcubic.w, ycubic.y, ycubic.w) / s;
 
-	vec4 sample0 = texture(textureSrc, vec2(offset.x, offset.z) * texscale);
+	vec4 sample0 = texture(textureSrc, vec2(offset.x, offset.z) * texelSize.xy);
-	vec4 sample1 = texture(textureSrc, vec2(offset.y, offset.z) * texscale);
+	vec4 sample1 = texture(textureSrc, vec2(offset.y, offset.z) * texelSize.xy);
-	vec4 sample2 = texture(textureSrc, vec2(offset.x, offset.w) * texscale);
+	vec4 sample2 = texture(textureSrc, vec2(offset.x, offset.w) * texelSize.xy);
-	vec4 sample3 = texture(textureSrc, vec2(offset.y, offset.w) * texscale);
+	vec4 sample3 = texture(textureSrc, vec2(offset.y, offset.w) * texelSize.xy);
 
 	float sx = s.x / (s.x + s.y);
 	float sy = s.z / (s.z + s.w);
@@ -76,20 +50,13 @@ vec4 bcFilter(vec2 texcoord, vec2 texscale)
 }
 
 void main(){
-	colorOut0 = vec4(bcFilter(passUV*textureSrcResolution, vec2(1.0,1.0)/textureSrcResolution).rgb,1.0);
+	vec4 texelSize = vec4( 1.0 / textureSrcResolution.xy, textureSrcResolution.xy);
+	colorOut0 = vec4(bcFilter(passUV, texelSize).rgb,1.0);
 }
 )";
 
 const std::string RendererOutputShader::s_hermite_shader_source =
-R"(#version 420
+R"(
-
-in vec4 gl_FragCoord;	
-in vec2 passUV;
-layout(binding=0) uniform sampler2D textureSrc;
-uniform vec2 textureSrcResolution;
-uniform vec2 outputResolution;
-layout(location = 0) out vec4 colorOut0;
-
 // https://www.shadertoy.com/view/MllSzX
 
 vec3 CubicHermite (vec3 A, vec3 B, vec3 C, vec3 D, float t)
@@ -111,7 +78,7 @@ vec3 BicubicHermiteTexture(vec2 uv, vec4 texelSize)
 	vec2 frac = fract(pixel);	
     pixel = floor(pixel) / texelSize.zw - vec2(texelSize.xy/2.0);
 	
-	vec4 doubleSize = texelSize*texelSize;
+	vec4 doubleSize = texelSize*2.0;
 
 	vec3 C00 = texture(textureSrc, pixel + vec2(-texelSize.x ,-texelSize.y)).rgb;
     vec3 C10 = texture(textureSrc, pixel + vec2( 0.0        ,-texelSize.y)).rgb;
@@ -142,15 +109,17 @@ vec3 BicubicHermiteTexture(vec2 uv, vec4 texelSize)
 }
 
 void main(){
-	vec4 texelSize = vec4( 1.0 / outputResolution.xy, outputResolution.xy);
+	vec4 texelSize = vec4( 1.0 / textureSrcResolution.xy, textureSrcResolution.xy);
 	colorOut0 = vec4(BicubicHermiteTexture(passUV, texelSize), 1.0);
 }
 )";
 
 RendererOutputShader::RendererOutputShader(const std::string& vertex_source, const std::string& fragment_source)
 {
+	auto finalFragmentSrc = PrependFragmentPreamble(fragment_source);
+
 	m_vertex_shader = g_renderer->shader_create(RendererShader::ShaderType::kVertex, 0, 0, vertex_source, false, false);
-	m_fragment_shader = g_renderer->shader_create(RendererShader::ShaderType::kFragment, 0, 0, fragment_source, false, false);
+	m_fragment_shader = g_renderer->shader_create(RendererShader::ShaderType::kFragment, 0, 0, finalFragmentSrc, false, false);
 
 	m_vertex_shader->PreponeCompilation(true);
 	m_fragment_shader->PreponeCompilation(true);
@@ -163,74 +132,45 @@ RendererOutputShader::RendererOutputShader(const std::string& vertex_source, con
 
 	if (g_renderer->GetType() == RendererAPI::OpenGL)
 	{
-		m_attributes[0].m_loc_texture_src_resolution = m_vertex_shader->GetUniformLocation("textureSrcResolution");
+		m_uniformLocations[0].m_loc_textureSrcResolution = m_vertex_shader->GetUniformLocation("textureSrcResolution");
-		m_attributes[0].m_loc_input_resolution = m_vertex_shader->GetUniformLocation("inputResolution");
+		m_uniformLocations[0].m_loc_nativeResolution = m_vertex_shader->GetUniformLocation("nativeResolution");
-		m_attributes[0].m_loc_output_resolution = m_vertex_shader->GetUniformLocation("outputResolution");
+		m_uniformLocations[0].m_loc_outputResolution = m_vertex_shader->GetUniformLocation("outputResolution");
 
-		m_attributes[1].m_loc_texture_src_resolution = m_fragment_shader->GetUniformLocation("textureSrcResolution");
+		m_uniformLocations[1].m_loc_textureSrcResolution = m_fragment_shader->GetUniformLocation("textureSrcResolution");
-		m_attributes[1].m_loc_input_resolution = m_fragment_shader->GetUniformLocation("inputResolution");
+		m_uniformLocations[1].m_loc_nativeResolution = m_fragment_shader->GetUniformLocation("nativeResolution");
-		m_attributes[1].m_loc_output_resolution = m_fragment_shader->GetUniformLocation("outputResolution");
+		m_uniformLocations[1].m_loc_outputResolution = m_fragment_shader->GetUniformLocation("outputResolution");
 	}
-	else
-	{
-		cemuLog_logDebug(LogType::Force, "RendererOutputShader() - todo for Vulkan");
-		m_attributes[0].m_loc_texture_src_resolution = -1;
-		m_attributes[0].m_loc_input_resolution = -1;
-		m_attributes[0].m_loc_output_resolution = -1;
-
-		m_attributes[1].m_loc_texture_src_resolution = -1;
-		m_attributes[1].m_loc_input_resolution = -1;
-		m_attributes[1].m_loc_output_resolution = -1;
-	}
-
 }
 
-void RendererOutputShader::SetUniformParameters(const LatteTextureView& texture_view, const Vector2i& input_res, const Vector2i& output_res) const
+void RendererOutputShader::SetUniformParameters(const LatteTextureView& texture_view, const Vector2i& output_res) const
 {
+	sint32 effectiveWidth, effectiveHeight;
+	texture_view.baseTexture->GetEffectiveSize(effectiveWidth, effectiveHeight, 0);
+	auto setUniforms = [&](RendererShader* shader, const UniformLocations& locations){
 	  float res[2];
-	// vertex shader
+	  if (locations.m_loc_textureSrcResolution != -1)
-	if (m_attributes[0].m_loc_texture_src_resolution != -1)
+	  {
+		  res[0] = (float)effectiveWidth;
+		  res[1] = (float)effectiveHeight;
+		  shader->SetUniform2fv(locations.m_loc_textureSrcResolution, res, 1);
+	  }
+
+	  if (locations.m_loc_nativeResolution != -1)
 	  {
 		  res[0] = (float)texture_view.baseTexture->width;
 		  res[1] = (float)texture_view.baseTexture->height;
-		m_vertex_shader->SetUniform2fv(m_attributes[0].m_loc_texture_src_resolution, res, 1);
+		  shader->SetUniform2fv(locations.m_loc_nativeResolution, res, 1);
 	  }
 
-	if (m_attributes[0].m_loc_input_resolution != -1)
+	  if (locations.m_loc_outputResolution != -1)
-	{
-		res[0] = (float)input_res.x;
-		res[1] = (float)input_res.y;
-		m_vertex_shader->SetUniform2fv(m_attributes[0].m_loc_input_resolution, res, 1);
-	}
-
-	if (m_attributes[0].m_loc_output_resolution != -1)
 	  {
 		  res[0] = (float)output_res.x;
 		  res[1] = (float)output_res.y;
-		m_vertex_shader->SetUniform2fv(m_attributes[0].m_loc_output_resolution, res, 1);
+		  shader->SetUniform2fv(locations.m_loc_outputResolution, res, 1);
-	}
-
-	// fragment shader
-	if (m_attributes[1].m_loc_texture_src_resolution != -1)
-	{
-		res[0] = (float)texture_view.baseTexture->width;
-		res[1] = (float)texture_view.baseTexture->height;
-		m_fragment_shader->SetUniform2fv(m_attributes[1].m_loc_texture_src_resolution, res, 1);
-	}
-
-	if (m_attributes[1].m_loc_input_resolution != -1)
-	{
-		res[0] = (float)input_res.x;
-		res[1] = (float)input_res.y;
-		m_fragment_shader->SetUniform2fv(m_attributes[1].m_loc_input_resolution, res, 1);
-	}
-
-	if (m_attributes[1].m_loc_output_resolution != -1)
-	{
-		res[0] = (float)output_res.x;
-		res[1] = (float)output_res.y;
-		m_fragment_shader->SetUniform2fv(m_attributes[1].m_loc_output_resolution, res, 1);
 	  }
+	};
+	setUniforms(m_vertex_shader, m_uniformLocations[0]);
+	setUniforms(m_fragment_shader, m_uniformLocations[1]);
 }
 
 RendererOutputShader* RendererOutputShader::s_copy_shader;
@@ -341,6 +281,27 @@ void main(){
 )";
 		return vertex_source.str();
 }
+
+std::string RendererOutputShader::PrependFragmentPreamble(const std::string& shaderSrc)
+{
+	return R"(#version 430
+#ifdef VULKAN
+layout(push_constant) uniform pc {
+	vec2 textureSrcResolution;
+	vec2 nativeResolution;
+	vec2 outputResolution;
+};
+#else
+uniform vec2 textureSrcResolution;
+uniform vec2 nativeResolution;
+uniform vec2 outputResolution;
+#endif
+
+layout(location = 0) in vec2 passUV;
+layout(binding = 0) uniform sampler2D textureSrc;
+layout(location = 0) out vec4 colorOut0;
+)" + shaderSrc;
+}
 void RendererOutputShader::InitializeStatic()
 {
 	std::string vertex_source, vertex_source_ud;
@@ -349,7 +310,12 @@ void RendererOutputShader::InitializeStatic()
 	{
 		vertex_source = GetOpenGlVertexSource(false);
 		vertex_source_ud = GetOpenGlVertexSource(true);
-
+	}
+	else
+	{
+		vertex_source = GetVulkanVertexSource(false);
+		vertex_source_ud = GetVulkanVertexSource(true);
+	}
 	s_copy_shader = new RendererOutputShader(vertex_source, s_copy_shader_source);
 	s_copy_shader_ud = new RendererOutputShader(vertex_source_ud, s_copy_shader_source);
 
@@ -358,19 +324,4 @@ void RendererOutputShader::InitializeStatic()
 
 	s_hermit_shader = new RendererOutputShader(vertex_source, s_hermite_shader_source);
 	s_hermit_shader_ud = new RendererOutputShader(vertex_source_ud, s_hermite_shader_source);
-	}
-	else
-	{
-		vertex_source = GetVulkanVertexSource(false);
-		vertex_source_ud = GetVulkanVertexSource(true);
-
-		s_copy_shader = new RendererOutputShader(vertex_source, s_copy_shader_source);
-		s_copy_shader_ud = new RendererOutputShader(vertex_source_ud, s_copy_shader_source);
-
-	/*	s_bicubic_shader = new RendererOutputShader(vertex_source, s_bicubic_shader_source); TODO
-		s_bicubic_shader_ud = new RendererOutputShader(vertex_source_ud, s_bicubic_shader_source);
-
-		s_hermit_shader = new RendererOutputShader(vertex_source, s_hermite_shader_source);
-		s_hermit_shader_ud = new RendererOutputShader(vertex_source_ud, s_hermite_shader_source);*/
-	}
 }

src/Cafe/HW/Latte/Renderer/RendererOuputShader.h

@@ -17,7 +17,7 @@ public:
 	RendererOutputShader(const std::string& vertex_source, const std::string& fragment_source);
 	virtual ~RendererOutputShader() = default;
 
-	void SetUniformParameters(const LatteTextureView& texture_view, const Vector2i& input_res, const Vector2i& output_res) const;
+	void SetUniformParameters(const LatteTextureView& texture_view, const Vector2i& output_res) const;
 
 	RendererShader* GetVertexShader() const
 	{
@@ -43,16 +43,18 @@ public:
 	static std::string GetVulkanVertexSource(bool render_upside_down);
 	static std::string GetOpenGlVertexSource(bool render_upside_down);
 
+	static std::string PrependFragmentPreamble(const std::string& shaderSrc);
+
 protected:
 	RendererShader* m_vertex_shader;
 	RendererShader* m_fragment_shader;
 
-	struct
+	struct UniformLocations
 	{
-		sint32 m_loc_texture_src_resolution = -1;
+		sint32 m_loc_textureSrcResolution = -1;
-		sint32 m_loc_input_resolution = -1;
+		sint32 m_loc_nativeResolution = -1;
-		sint32 m_loc_output_resolution = -1;
+		sint32 m_loc_outputResolution = -1;
-	} m_attributes[2]{};
+	} m_uniformLocations[2]{};
 
 private:
 	static const std::string s_copy_shader_source;

src/Cafe/HW/Latte/Renderer/Vulkan/LatteTextureViewVk.cpp

@@ -202,6 +202,13 @@ VkSampler LatteTextureViewVk::GetDefaultTextureSampler(bool useLinearTexFilter)
 	VkSamplerCreateInfo samplerInfo{};
 	samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
 
+	// emulate OpenGL minFilters
+	// see note under: https://docs.vulkan.org/spec/latest/chapters/samplers.html#VkSamplerCreateInfo
+	// if maxLod = 0 then magnification is always performed
+	samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
+	samplerInfo.minLod = 0.0f;
+	samplerInfo.maxLod = 0.25f;
+
 	if (useLinearTexFilter)
 	{
 		samplerInfo.magFilter = VK_FILTER_LINEAR;
@@ -212,6 +219,9 @@ VkSampler LatteTextureViewVk::GetDefaultTextureSampler(bool useLinearTexFilter)
 		samplerInfo.magFilter = VK_FILTER_NEAREST;
 		samplerInfo.minFilter = VK_FILTER_NEAREST;
 	}
+	samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
+	samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
+	samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
 
 	if (vkCreateSampler(m_device, &samplerInfo, nullptr, &sampler) != VK_SUCCESS)
 	{

src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.cpp

@@ -2581,10 +2581,18 @@ VkPipeline VulkanRenderer::backbufferBlit_createGraphicsPipeline(VkDescriptorSet
 	colorBlending.blendConstants[2] = 0.0f;
 	colorBlending.blendConstants[3] = 0.0f;
 
+	VkPushConstantRange pushConstantRange{
+		.stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT,
+		.offset = 0,
+		.size = 3 * sizeof(float) * 2 // 3 vec2's
+	};
+
 	VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
 	pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
 	pipelineLayoutInfo.setLayoutCount = 1;
 	pipelineLayoutInfo.pSetLayouts = &descriptorLayout;
+	pipelineLayoutInfo.pushConstantRangeCount = 1;
+	pipelineLayoutInfo.pPushConstantRanges = &pushConstantRange;
 
 	VkResult result = vkCreatePipelineLayout(m_logicalDevice, &pipelineLayoutInfo, nullptr, &m_pipelineLayout);
 	if (result != VK_SUCCESS)
@@ -2956,6 +2964,25 @@ void VulkanRenderer::DrawBackbufferQuad(LatteTextureView* texView, RendererOutpu
 
 	vkCmdBindDescriptorSets(m_state.currentCommandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, m_pipelineLayout, 0, 1, &descriptSet, 0, nullptr);
 
+	// update push constants
+	Vector2f pushData[3];
+
+	// textureSrcResolution
+	sint32 effectiveWidth, effectiveHeight;
+	texView->baseTexture->GetEffectiveSize(effectiveWidth, effectiveHeight, 0);
+	pushData[0] = {(float)effectiveWidth, (float)effectiveHeight};
+
+	// nativeResolution
+	pushData[1] = {
+		(float)texViewVk->baseTexture->width,
+		(float)texViewVk->baseTexture->height,
+	};
+
+	// outputResolution
+	pushData[2] = {(float)imageWidth,(float)imageHeight};
+
+	vkCmdPushConstants(m_state.currentCommandBuffer, m_pipelineLayout, VK_SHADER_STAGE_FRAGMENT_BIT, 0, sizeof(float) * 2 * 3, &pushData);
+
 	vkCmdDraw(m_state.currentCommandBuffer, 6, 1, 0, 0);
 
 	vkCmdEndRenderPass(m_state.currentCommandBuffer);