// SPDX-FileCopyrightText: 2002-2026 PCSX2 Dev Team
// SPDX-License-Identifier: GPL-3.0+

#include "GS/Renderers/OpenGL/GLContext.h"
#include "GS/Renderers/OpenGL/GSDeviceOGL.h"
#include "GS/Renderers/OpenGL/GLState.h"
#include "GS/Renderers/Common/GSGPUProfile.h"
#include "GS/GSState.h"
#include "GS/Renderers/Common/GSRenderer.h"
#include "GS/GSGL.h"
#include "GS/GSPerfMon.h"
#include "GS/GSUtil.h"
#include "Host.h"

#include "common/Console.h"
#include "common/Error.h"
#include "common/ScopedGuard.h"
#include "common/StringUtil.h"

#include "imgui.h"
#include "IconsFontAwesome.h"

#ifdef ARMSX2_HAS_LIBRASHADER
// librashader only declares its OpenGL entry points when the consumer opts in. Defined by
// CMake only when the Rust toolchain actually produced the library.
#define LIBRA_RUNTIME_OPENGL
#include "librashader.h"
#endif

#include <cinttypes>
#include <fstream>
#include <sstream>

static constexpr u32 g_vs_pc_index        = 4;
static constexpr u32 g_vs_ib_index        = 3;
static constexpr u32 g_vs_vb_index        = 2;
static constexpr u32 g_vs_cb_index        = 1;
static constexpr u32 g_ps_cb_index        = 0;

static constexpr u32 VERTEX_BUFFER_SIZE = 32 * 1024 * 1024;
static constexpr u32 INDEX_BUFFER_SIZE = 16 * 1024 * 1024;
static constexpr u32 VERTEX_UNIFORM_BUFFER_SIZE = 8 * 1024 * 1024;
static constexpr u32 FRAGMENT_UNIFORM_BUFFER_SIZE = 8 * 1024 * 1024;
static constexpr u32 VERTEX_PUSH_CONSTANT_BUFFER_SIZE = 1024 * 1024;
static constexpr u32 TEXTURE_UPLOAD_BUFFER_SIZE = 128 * 1024 * 1024;

namespace ReplaceGL
{
	static void GLAPIENTRY ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height)
	{
		glScissor(left, bottom, width, height);
	}

	static void GLAPIENTRY ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h)
	{
		glViewport(GLint(x), GLint(y), GLsizei(w), GLsizei(h));
	}

	static void GLAPIENTRY MemoryBarrierAsTextureBarrier()
	{
		glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
	}

	static void GLAPIENTRY TextureBarrier()
	{

	}

} // namespace ReplaceGL

namespace Emulate_DSA
{
	// Texture entry point
	static void GLAPIENTRY BindTextureUnit(GLuint unit, GLuint texture)
	{
		glActiveTexture(GL_TEXTURE0 + unit);
		glBindTexture(GL_TEXTURE_2D, texture);
	}

	static void GLAPIENTRY CreateTexture(GLenum target, GLsizei n, GLuint* textures)
	{
		glGenTextures(1, textures);
	}

	static void GLAPIENTRY TextureStorage(
		GLuint texture, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height)
	{
		BindTextureUnit(7, texture);
		glTexStorage2D(GL_TEXTURE_2D, levels, internalformat, width, height);
	}

	static void GLAPIENTRY TextureSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
		GLsizei height, GLenum format, GLenum type, const void* pixels)
	{
		BindTextureUnit(7, texture);
		glTexSubImage2D(GL_TEXTURE_2D, level, xoffset, yoffset, width, height, format, type, pixels);
	}

	static void GLAPIENTRY CopyTextureSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width, GLsizei height)
	{
		BindTextureUnit(7, texture);
		glCopyTexSubImage2D(GL_TEXTURE_2D, level, xoffset, yoffset, x, y, width, height);
	}

	static void GLAPIENTRY CompressedTextureSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset,
		GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data)
	{
		BindTextureUnit(7, texture);
		glCompressedTexSubImage2D(GL_TEXTURE_2D, level, xoffset, yoffset, width, height, format, imageSize, data);
	}

	static void GLAPIENTRY GetTexureImage(
		GLuint texture, GLint level, GLenum format, GLenum type, GLsizei bufSize, void* pixels)
	{
		BindTextureUnit(7, texture);
		glGetTexImage(GL_TEXTURE_2D, level, format, type, pixels);
	}

	static void GLAPIENTRY TextureParameteri(GLuint texture, GLenum pname, GLint param)
	{
		BindTextureUnit(7, texture);
		glTexParameteri(GL_TEXTURE_2D, pname, param);
	}

	static void GLAPIENTRY GenerateTextureMipmap(GLuint texture)
	{
		BindTextureUnit(7, texture);
		glGenerateMipmap(GL_TEXTURE_2D);
	}

	// Misc entry point
	static void GLAPIENTRY CreateSamplers(GLsizei n, GLuint* samplers)
	{
		glGenSamplers(n, samplers);
	}

	// Replace function pointer to emulate DSA behavior
	static void Init()
	{
		glBindTextureUnit = BindTextureUnit;
		glCreateTextures = CreateTexture;
		glTextureStorage2D = TextureStorage;
		glTextureSubImage2D = TextureSubImage;
		glCopyTextureSubImage2D = CopyTextureSubImage;
		glCompressedTextureSubImage2D = CompressedTextureSubImage;
		glGetTextureImage = GetTexureImage;
		glTextureParameteri = TextureParameteri;
		glGenerateTextureMipmap = GenerateTextureMipmap;
		glCreateSamplers = CreateSamplers;
	}
} // namespace Emulate_DSA

GSDeviceOGL::GSDeviceOGL() = default;

GSDeviceOGL::~GSDeviceOGL()
{
	pxAssert(!m_gl_context);
}

std::vector<GSAdapterInfo> GSDeviceOGL::GetAdapterInfo()
{
	std::vector<GSAdapterInfo> ret;

	// If OGL is currently active, query the live device directly.
	if (g_gs_device && g_gs_device->GetRenderAPI() == RenderAPI::OpenGL)
	{
		GSAdapterInfo ai;
		const std::string& live_name = static_cast<GSDeviceOGL*>(g_gs_device.get())->m_name;
		ai.name = live_name.empty() ? "OpenGL" : live_name;
		ai.max_texture_size = g_gs_device->GetMaxTextureSize();
		ai.max_upscale_multiplier = GSGetMaxUpscaleMultiplier(ai.max_texture_size);
		ret.emplace_back(std::move(ai));
		return ret;
	}

	static std::optional<GSAdapterInfo> s_cached_adapter;
	if (s_cached_adapter.has_value())
	{
		ret.emplace_back(*s_cached_adapter);
		return ret;
	}

	static bool s_logged_probe_failure = false;

	WindowInfo wi = {};
	wi.type = WindowInfo::Type::Surfaceless;

	Error error;
	std::unique_ptr<GLContext> context = GLContext::Create(wi, &error);
	if (!context)
	{
		if (!s_logged_probe_failure)
		{
			Console.WarningFmt("GL: Failed to query GL_MAX_TEXTURE_SIZE for adapter info: {}", error.GetDescription());
			s_logged_probe_failure = true;
		}
		return ret;
	}

	if (!context->MakeCurrent())
	{
		if (!s_logged_probe_failure)
		{
			Console.Warning("GL: Failed to make temporary context current while querying GL_MAX_TEXTURE_SIZE for adapter info.");
			s_logged_probe_failure = true;
		}
		return ret;
	}

	const ScopedGuard done_current([&context]() { context->DoneCurrent(); });

	GLint max_texture_size = 1024;
	glGetIntegerv(GL_MAX_TEXTURE_SIZE, &max_texture_size);

	GSAdapterInfo ai;
	const GLubyte* renderer_name = glGetString(GL_RENDERER);
	ai.name = renderer_name ? reinterpret_cast<const char*>(renderer_name) : "OpenGL";
	ai.max_texture_size = std::max(static_cast<u32>((max_texture_size > 0) ? max_texture_size : 1024), 1024u);
	ai.max_upscale_multiplier = GSGetMaxUpscaleMultiplier(ai.max_texture_size);
	s_cached_adapter = ai;
	ret.emplace_back(std::move(ai));
	return ret;
}

GSTexture* GSDeviceOGL::CreateSurface(GSTexture::Usage usage, int width, int height, int levels, GSTexture::Format format)
{
	GL_PUSH("Create surface");
	pxAssert(GLAD_GL_ARB_shader_image_load_store || !GSTexture::IsShaderWrite(usage));
	return new GSTextureOGL(usage, width, height, levels, format);
}

RenderAPI GSDeviceOGL::GetRenderAPI() const
{
	return RenderAPI::OpenGL;
}

bool GSDeviceOGL::HasSurface() const
{
	return m_window_info.type != WindowInfo::Type::Surfaceless;
}

void GSDeviceOGL::SetVSyncMode(GSVSyncMode mode, bool allow_present_throttle)
{
	m_allow_present_throttle = allow_present_throttle;

	if (m_vsync_mode == mode)
		return;

	m_vsync_mode = mode;
	SetSwapInterval();
}

bool GSDeviceOGL::Create(GSVSyncMode vsync_mode, bool allow_present_throttle)
{
	if (!GSDevice::Create(vsync_mode, allow_present_throttle))
		return false;

	// GL is a pain and needs the window super early to create the context.
	if (!AcquireWindow(true))
		return false;

	Error error;
	m_gl_context = GLContext::Create(m_window_info, &error);
	if (!m_gl_context)
	{
		Console.ErrorFmt("GL: Failed to create any context: {}", error.GetDescription());
		return false;
	}

	if (!m_gl_context->MakeCurrent())
	{
		Console.Error("GL: Failed to make context current");
		return false;
	}

	m_is_gles = m_gl_context->IsGLES();

	if (!CheckFeatures())
		return false;

	// Store adapter name currently in use
	m_name = reinterpret_cast<const char*>(glGetString(GL_RENDERER));

	SetSwapInterval();

	// Render a frame as soon as possible to clear out whatever was previously being displayed.
	if (m_window_info.type != WindowInfo::Type::Surfaceless)
		RenderBlankFrame();

	if (!GSConfig.DisableShaderCache)
	{
		if (!m_shader_cache.Open(m_is_gles))
			Console.Warning("GL: Shader cache failed to open.");
	}
	else
	{
		Console.WriteLn("GL: Not using shader cache.");
	}

	// GL-ES init bisect markers (Adreno 650/740 boot crash). The crash is a
	// silent driver fault with no bad-shader dump, so the LAST stage printed in
	// the emulog before it cuts off pinpoints the faulting phase.
	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=objects");
	// because of fbo bindings below...
	GLState::Clear();

	// ****************************************************************
	// Debug helper
	// ****************************************************************
	if (GSConfig.UseDebugDevice)
	{
		if (!m_is_gles) {
			glDebugMessageCallback(DebugMessageCallback, NULL);

			glDebugMessageControl(GL_DONT_CARE, GL_DONT_CARE, GL_DONT_CARE, 0, NULL, true);
			// Useless info message on Nvidia driver
			static constexpr const GLuint ids[] = {0x20004};
			glDebugMessageControl(GL_DEBUG_SOURCE_API_ARB, GL_DEBUG_TYPE_OTHER_ARB, GL_DONT_CARE,
								  std::size(ids), ids, false);
		}
		else if (GLAD_GL_KHR_debug)
		{
			glDebugMessageCallback(DebugMessageCallback, nullptr);

			glDebugMessageControl(GL_DONT_CARE, GL_DONT_CARE, GL_DONT_CARE, 0, nullptr, true);

			// Useless info message on Nvidia driver
			static constexpr const GLuint ids[] = { 0x20004 };
			glDebugMessageControl(GL_DEBUG_SOURCE_API_ARB, GL_DEBUG_TYPE_OTHER_ARB, GL_DONT_CARE, std::size(ids), ids, false);
		}

		// Uncomment synchronous if you want callstacks which match where the error occurred.
		glEnable(GL_DEBUG_OUTPUT);
		//glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
	}

	// WARNING it must be done after the control setup (at least on MESA)
	GL_PUSH("GSDeviceOGL::Create");

	// ****************************************************************
	// Various object
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Various");

		glGenFramebuffers(1, &m_fbo);
		glGenFramebuffers(1, &m_fbo_read);
		glGenFramebuffers(1, &m_fbo_write);

		OMSetFBO(m_fbo);

		// Always read from the first buffer
		glBindFramebuffer(GL_READ_FRAMEBUFFER, m_fbo_read);
		glReadBuffer(GL_COLOR_ATTACHMENT0);
		glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
	}

	// ****************************************************************
	// Vertex buffer state
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Vertex Buffer");

		glGenVertexArrays(1, &m_vao);
		IASetVAO(m_vao);

		m_vertex_stream_buffer = GLStreamBuffer::Create(GL_ARRAY_BUFFER, VERTEX_BUFFER_SIZE);
		m_index_stream_buffer = GLStreamBuffer::Create(GL_ELEMENT_ARRAY_BUFFER, INDEX_BUFFER_SIZE);
		m_expand_index_stream_buffer = GLStreamBuffer::Create(GL_ARRAY_BUFFER, INDEX_BUFFER_SIZE);
		m_vertex_uniform_stream_buffer = GLStreamBuffer::Create(GL_UNIFORM_BUFFER, VERTEX_UNIFORM_BUFFER_SIZE);
		m_fragment_uniform_stream_buffer = GLStreamBuffer::Create(GL_UNIFORM_BUFFER, FRAGMENT_UNIFORM_BUFFER_SIZE);
		m_vertex_push_constants_stream_buffer = GLStreamBuffer::Create(GL_UNIFORM_BUFFER, VERTEX_PUSH_CONSTANT_BUFFER_SIZE);
		glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &m_uniform_buffer_alignment);
		if (!m_vertex_stream_buffer || !m_index_stream_buffer || !m_expand_index_stream_buffer ||
			!m_vertex_uniform_stream_buffer || !m_fragment_uniform_stream_buffer || !m_vertex_push_constants_stream_buffer)
		{
			Host::ReportErrorAsync("GS", "Failed to create vertex/index/uniform streaming buffers");
			return false;
		}

		m_vertex_stream_buffer->Bind();
		m_index_stream_buffer->Bind();

		// Force UBOs to be uploaded on first use.
		std::memset(static_cast<void*>(&m_vs_cb_cache), 0xFF, sizeof(m_vs_cb_cache));
		std::memset(static_cast<void*>(&m_ps_cb_cache), 0xFF, sizeof(m_ps_cb_cache));

		static_assert(sizeof(GSVertexPT1) == sizeof(GSVertex), "wrong GSVertex size");
		for (u32 i = 0; i < 8; i++)
			glEnableVertexAttribArray(i);

		glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(GSVertexPT1), (const GLvoid*)(0));
		glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(GSVertexPT1), (const GLvoid*)(16));
		glVertexAttribPointer(2, 4, GL_UNSIGNED_BYTE, GL_FALSE, sizeof(GSVertex), (const GLvoid*)(8));
		glVertexAttribPointer(3, 1, GL_FLOAT, GL_FALSE, sizeof(GSVertex), (const GLvoid*)(12));
		glVertexAttribIPointer(4, 2, GL_UNSIGNED_SHORT, sizeof(GSVertex), (const GLvoid*)(16));
		glVertexAttribIPointer(5, 1, GL_UNSIGNED_INT, sizeof(GSVertex), (const GLvoid*)(20));
		glVertexAttribIPointer(6, 2, GL_UNSIGNED_SHORT, sizeof(GSVertex), (const GLvoid*)(24));
		glVertexAttribPointer(7, 4, GL_UNSIGNED_BYTE, GL_TRUE, sizeof(GSVertex), (const GLvoid*)(28));

		if (m_features.vs_expand)
		{
			glGenVertexArrays(1, &m_expand_vao);
			glBindVertexArray(m_expand_vao);
			IASetVAO(m_expand_vao);

			// Still need the vertex buffer bound, because uploads happen to GL_ARRAY_BUFFER.
			m_vertex_stream_buffer->Bind();

			std::unique_ptr<u8[]> expand_data = std::make_unique<u8[]>(EXPAND_BUFFER_SIZE);
			GenerateExpansionIndexBuffer(expand_data.get());
			glGenBuffers(1, &m_expand_ibo);
			glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, m_expand_ibo);
			glBufferData(GL_ELEMENT_ARRAY_BUFFER, EXPAND_BUFFER_SIZE, expand_data.get(), GL_STATIC_DRAW);
			glBindBufferRange(GL_SHADER_STORAGE_BUFFER, g_vs_vb_index, m_vertex_stream_buffer->GetGLBufferId(), 0, VERTEX_BUFFER_SIZE);
		}

		if (m_features.aa1)
		{
			glGenVertexArrays(1, &m_dummy_vao);
			glBindBufferRange(GL_SHADER_STORAGE_BUFFER, g_vs_ib_index, m_expand_index_stream_buffer->GetGLBufferId(), 0, INDEX_BUFFER_SIZE);
		}

		VSSetPushConstants(0, 0, true); // Avoid undefined data.
	}

	// ****************************************************************
	// Pre Generate the different sampler object
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Sampler");

		for (u32 key = 0; key < std::size(m_ps_ss); key++)
		{
			m_ps_ss[key] = CreateSampler(PSSamplerSelector(key));
		}
	}

	// these all share the same vertex shader
	const std::optional<std::string> convert_glsl = ReadShaderSource("shaders/opengl/convert.glsl");
	if (!convert_glsl.has_value())
	{
		Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/convert.glsl.");
		return false;
	}

	// ****************************************************************
	// convert
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Convert");

		m_convert.vs = GetShaderSource("vs_main", GL_VERTEX_SHADER, *convert_glsl);

		m_convert.ps.resize(ShaderConvertSelector::NUM_TOTAL_SHADERS);
		for (u32 i = 0; i < ShaderConvertSelector::NUM_TOTAL_SHADERS; i++)
		{
			const ShaderConvertSelector shader = ShaderConvertSelector::Get(i);

			const char* name = shader.EntryPoint();

			std::string macro;
			macro += fmt::format("#define HAS_BILN {}\n", static_cast<int>(shader.Biln()));
			macro += fmt::format("#define HAS_STENCIL_OUTPUT {}\n", static_cast<int>(shader.StencilOutput()));
			macro += fmt::format("#define HAS_INTEGER_OUTPUT {}\n", static_cast<int>(shader.IntegerOutputBpp() != 0));
			macro += fmt::format("#define HAS_DEPTH_OUTPUT {}\n", static_cast<int>(shader.DepthOutput()));
			macro += fmt::format("#define HAS_FLOAT32_INPUT {}\n", static_cast<int>(shader.Float32Input()));
			macro += fmt::format("#define HAS_FLOAT32_OUTPUT {}\n", static_cast<int>(shader.Float32Output()));

			const std::string ps(GetShaderSource(name, GL_FRAGMENT_SHADER, *convert_glsl, macro));

			GLProgram& prog = m_convert.ps[i];

			if (!m_shader_cache.GetProgram(&prog, m_convert.vs, ps))
				return false;

			prog.SetFormattedName("Convert pipeline (%s, mask=%x, depth=%d, biln=%d)",
				shader.Name(), shader.Mask(), static_cast<int>(shader.DepthOutput()),
				static_cast<int>(shader.Biln()));

			if (shader.Shader() == ShaderConvert::RGBA_TO_8I || shader.Shader() == ShaderConvert::RGB5A1_TO_8I)
			{
				prog.RegisterUniform("SBW");
				prog.RegisterUniform("DBW");
				prog.RegisterUniform("PSM");
				prog.RegisterUniform("ScaleFactor");
			}
			else if (shader.Shader() == ShaderConvert::YUV)
			{
				prog.RegisterUniform("EMOD");
			}
			else if (shader.Shader() == ShaderConvert::CLUT_4 || shader.Shader() == ShaderConvert::CLUT_8)
			{
				prog.RegisterUniform("offset");
				prog.RegisterUniform("scale");
			}
			else if (shader.Shader() == ShaderConvert::DOWNSAMPLE_COPY)
			{
				prog.RegisterUniform("ClampMin");
				prog.RegisterUniform("DownsampleFactor");
				prog.RegisterUniform("Weight");
				prog.RegisterUniform("StepMultiplier");
			}
		}

		const PSSamplerSelector point;
		m_convert.pt = GetSamplerID(point);

		PSSamplerSelector bilinear;
		bilinear.biln = true;
		m_convert.ln = GetSamplerID(bilinear);

		m_convert.dss = new GSDepthStencilOGL();
		m_convert.dss_write = new GSDepthStencilOGL();
		m_convert.dss_write->EnableDepth();
		m_convert.dss_write->SetDepth(GL_ALWAYS, true);
	}

	// ****************************************************************
	// present
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Present");

		// these all share the same vertex shader
		const std::optional<std::string> shader = ReadShaderSource("shaders/opengl/present.glsl");
		if (!shader.has_value())
		{
			Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/present.glsl.");
			return false;
		}

		std::string present_vs(GetShaderSource("vs_main", GL_VERTEX_SHADER, *shader));

		for (size_t i = 0; i < std::size(m_present); i++)
		{
			const char* name = ShaderEntryPoint(static_cast<PresentShader>(i));
			const std::string ps(GetShaderSource(name, GL_FRAGMENT_SHADER, *shader));
			if (!m_shader_cache.GetProgram(&m_present[i], present_vs, ps))
				return false;
			m_present[i].SetFormattedName("Present pipe %s", name);

			// This is a bit disgusting, but it saves allocating a UBO when no shaders currently need it.
			m_present[i].RegisterUniform("u_source_rect");
			m_present[i].RegisterUniform("u_target_rect");
			m_present[i].RegisterUniform("u_source_size");
			m_present[i].RegisterUniform("u_target_size");
			m_present[i].RegisterUniform("u_target_resolution");
			m_present[i].RegisterUniform("u_rcp_target_resolution");
			m_present[i].RegisterUniform("u_source_resolution");
			m_present[i].RegisterUniform("u_rcp_source_resolution");
			m_present[i].RegisterUniform("u_time");
		}
	}

	// ****************************************************************
	// merge
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Merge");

		const std::optional<std::string> shader = ReadShaderSource("shaders/opengl/merge.glsl");
		if (!shader.has_value())
		{
			Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/merge.glsl.");
			return false;
		}

		for (size_t i = 0; i < std::size(m_merge_obj.ps); i++)
		{
			const std::string ps(GetShaderSource(fmt::format("ps_main{}", i), GL_FRAGMENT_SHADER, *shader));
			if (!m_shader_cache.GetProgram(&m_merge_obj.ps[i], m_convert.vs, ps))
				return false;
			m_merge_obj.ps[i].SetFormattedName("Merge pipe %zu", i);
			m_merge_obj.ps[i].RegisterUniform("BGColor");
		}
	}

	// ****************************************************************
	// interlace
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Interlace");

		const std::optional<std::string> shader = ReadShaderSource("shaders/opengl/interlace.glsl");
		if (!shader.has_value())
		{
			Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/interlace.glsl.");
			return false;
		}

		for (size_t i = 0; i < std::size(m_interlace.ps); i++)
		{
			const std::string ps(GetShaderSource(fmt::format("ps_main{}", i), GL_FRAGMENT_SHADER, *shader));
			if (!m_shader_cache.GetProgram(&m_interlace.ps[i], m_convert.vs, ps))
				return false;
			m_interlace.ps[i].SetFormattedName("Merge pipe %zu", i);
			m_interlace.ps[i].RegisterUniform("ZrH");
		}
	}

	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=convert_present_merge_interlace_done");
	// ****************************************************************
	// Post processing
	// ****************************************************************
	if (!CompileShadeBoostProgram() || !CompileFXAAProgram())
		return false;

	// Image load store and GLSL 420pack is core in GL4.2, no need to check.
	// NOTE: CAS uses a desktop-GL compute shader (cas.glsl is "#version 420" +
	// "#extension GL_ARB_compute_shader") — that source is invalid GLSL ES. On
	// devices whose driver only gives a GL ES 3.2 context (e.g. Adreno 650 on the
	// Retroid Pocket Mini, where the desktop 4.2 context request fails and we fall
	// back to ES), feeding the ES compiler that shader hard-crashes the driver
	// during GS init. So gate CAS to a real desktop-GL 4.2 context only; ES devices
	// just go without the sharpening filter (TV/CRT present shaders are unaffected —
	// they use the ES-aware "#version 320 es" header and compile fine).
	m_features.cas_sharpening = (GLAD_GL_VERSION_4_2 && GLAD_GL_ARB_compute_shader) && CreateCASPrograms();
	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=postproc_done");

	// ****************************************************************
	// rasterization configuration
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Rasterization");

		if (!m_is_gles) {
			glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
			glDisable(GL_MULTISAMPLE);
		}

		glDisable(GL_CULL_FACE);
		glEnable(GL_SCISSOR_TEST);

		glDisable(GL_DITHER); // Honestly I don't know!

		// Initialise stencil ref/mask.
		glStencilFunc(GLState::stencil_func, 1, 1);
	}

	// ****************************************************************
	// DATE
	// ****************************************************************
	{
		GL_PUSH("GSDeviceOGL::Date");

		m_date.dss = new GSDepthStencilOGL();
		m_date.dss->EnableStencil();
		m_date.dss->SetStencil(GL_ALWAYS, GL_REPLACE);

		for (size_t i = 0; i < std::size(m_date.primid_ps); i++)
		{
			const std::string ps(GetShaderSource(
				fmt::format("ps_primid_image_init_{}", i),
				GL_FRAGMENT_SHADER, *convert_glsl));
			m_shader_cache.GetProgram(&m_date.primid_ps[i], m_convert.vs, ps);
			m_date.primid_ps[i].SetFormattedName("PrimID Destination Alpha Init %d", i);
		}
	}

	// ****************************************************************
	// Use DX coordinate convention
	// ****************************************************************

	// VS gl_position.z => [-1,-1]
	// FS depth => [0, 1]
	// because of -1 we loose lot of precision for small GS value
	// This extension allow FS depth to range from -1 to 1. So
	// gl_position.z could range from [0, 1]
	// Change depth convention
	if (GLAD_GL_ARB_clip_control)
		glClipControl(GL_LOWER_LEFT, GL_ZERO_TO_ONE);
	else if (m_is_gles && GLAD_GL_EXT_clip_control)
		// GLES has no ARB_clip_control; GL_EXT_clip_control (advertised by Adreno, and
		// some Mali) is the same API/enums. Without it, GLES uses the legacy z-remap that
		// CLAMPS PS2 Z >= 2^24 to the far plane (tfx_vgs.glsl), collapsing far depth so
		// large-Z world geometry z-fights/vanishes — e.g. God of War II's transparent
		// walls. Pairs with HAS_CLIP_CONTROL below (same condition) so the shader matches.
		glClipControlEXT(GL_LOWER_LEFT_EXT, GL_ZERO_TO_ONE_EXT);

	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=date_raster_done");
	// ****************************************************************
	// HW renderer shader
	// ****************************************************************
	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=texturefx_begin");
	if (!CreateTextureFX())
		return false;
	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=texturefx_done");

	// ****************************************************************
	// Pbo Pool allocation
	// ****************************************************************
	if (!m_bugs.buggy_pbo)
	{
		m_texture_upload_buffer = GLStreamBuffer::Create(GL_PIXEL_UNPACK_BUFFER, TEXTURE_UPLOAD_BUFFER_SIZE);
		if (m_texture_upload_buffer)
		{
			// Don't keep it bound, we'll re-bind when we need it.
			// Otherwise non-PBO texture uploads break. Yay for global state.
			m_texture_upload_buffer->Unbind();
		}
		else
		{
			Console.Error("GL: Failed to create texture upload buffer. Using slow path.");
		}
	}

	if (!CreateImGuiProgram())
		return false;

	// GLES has no pipeline-statistics queries; this extension is desktop-GL only,
	// so on Android this stays false and the OSD line shows 0 / degrades gracefully.
	m_gpu_pipeline_statistics_supported = (GLAD_GL_ARB_pipeline_statistics_query != 0);

	// Basic to ensure structures are correctly packed
	static_assert(sizeof(VSSelector) == 1, "Wrong VSSelector size");
	static_assert(sizeof(PSSelector) == 16, "Wrong PSSelector size");
	static_assert(sizeof(PSSamplerSelector) == 1, "Wrong PSSamplerSelector size");
	static_assert(sizeof(OMDepthStencilSelector) == 1, "Wrong OMDepthStencilSelector size");
	static_assert(sizeof(OMColorMaskSelector) == 1, "Wrong OMColorMaskSelector size");

	Console.WriteLn("@@ANDROID_GL_INIT@@ stage=create_done");
	return true;
}

void GSDeviceOGL::Destroy()
{
	// Frees GL objects, so it has to run before the context is dropped below.
	DestroyShaderChain();

	GSDevice::Destroy();

	if (m_gl_context)
	{
		DestroyTimestampQueries();
		DestroyPipelineStatisticsQueries();
		DestroyResources();

		m_gl_context->DoneCurrent();
		m_gl_context.reset();
	}
}

bool GSDeviceOGL::CreateTextureFX()
{
	GL_PUSH("GSDeviceOGL::CreateTextureFX");

	std::optional<std::string> vertex_shader = ReadShaderSource("shaders/opengl/tfx_vgs.glsl");
	std::optional<std::string> fragment_shader = ReadShaderSource("shaders/opengl/tfx_fs.glsl");
	if (!vertex_shader.has_value() || !fragment_shader.has_value())
	{
		Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/tfx_{vgs,fs}.glsl.");
		return false;
	}

	m_shader_tfx_vgs = std::move(*vertex_shader);
	m_shader_tfx_fs = std::move(*fragment_shader);

	// warning 1 sampler by image unit. So you cannot reuse m_ps_ss...
	m_palette_ss = CreateSampler(PSSamplerSelector(0));
	glBindSampler(1, m_palette_ss);

	// Enable all bits for stencil operations. Technically 1 bit is
	// enough but buffer is polluted with noise. Clear will be limited
	// to the mask.
	glStencilMask(0xFF);
	for (u32 key = 0; key < std::size(m_om_dss); key++)
	{
		m_om_dss[key] = CreateDepthStencil(OMDepthStencilSelector(key));
	}

	GLProgram::ResetLastProgram();
	return true;
}

bool GSDeviceOGL::CheckFeatures()
{
	//bool vendor_id_amd = false;
	//bool vendor_id_nvidia = false;
	//bool vendor_id_intel = false;

	memset(&m_bugs, 0, sizeof(m_bugs));

	bool vendor_id_mali = false;
	bool vendor_id_adreno = false;
	bool vendor_id_apple = false;

	const char* vendor_raw = (const char*)glGetString(GL_VENDOR);
	const char* renderer_raw = (const char*)glGetString(GL_RENDERER);
	const char* gl_version_raw = (const char*)glGetString(GL_VERSION);
	const char* vendor_str = vendor_raw ? vendor_raw : "";
	const char* renderer_str = renderer_raw ? renderer_raw : "";
	// GL_VERSION is the only place the GLES stack names its own build ("OpenGL ES 3.2 v1.r44p1-...",
	// "... V@0502", "... build 1.9@4850625"), so it is what the driver-profile resolver parses.
	const char* gl_version_str = gl_version_raw ? gl_version_raw : "";

	if (std::strstr(vendor_str, "Advanced Micro Devices") || std::strstr(vendor_str, "ATI Technologies Inc.") ||
		std::strstr(vendor_str, "ATI"))
	{
		Console.WriteLn(Color_StrongRed, "GL: AMD GPU detected.");
		//vendor_id_amd = true;
	}
	else if (std::strstr(vendor_str, "NVIDIA Corporation"))
	{
		Console.WriteLn(Color_StrongGreen, "GL: NVIDIA GPU detected.");
		//vendor_id_nvidia = true;
		m_bugs.broken_blend_coherency = true;
	}
	else if (std::strstr(vendor_str, "Intel"))
	{
		Console.WriteLn(Color_StrongBlue, "GL: Intel GPU detected.");
		//vendor_id_intel = true;
	}
	else if (std::strstr(vendor_str, "ARM") || std::strstr(renderer_str, "Mali"))
	{
		Console.WriteLn(Color_Yellow, "GL: ARM Mali GPU detected.");
		vendor_id_mali = true;
	}
	else if (std::strstr(vendor_str, "Qualcomm") || std::strstr(renderer_str, "Adreno"))
	{
		Console.WriteLn(Color_Cyan, "GL: Qualcomm Adreno GPU detected.");
		vendor_id_adreno = true;
	}
	// Matched on the renderer, not the vendor: Apple silicon reports the vendor of whoever
	// wrote the driver ("Mesa" under Asahi, "Apple Inc." on macOS), while an Intel Mac reports
	// vendor "Apple Inc." with an AMD or Intel GPU. The renderer names the actual GPU.
	//
	// Apple silicon is a TBDR, but it is not a mobile-vendor part and must not inherit their
	// workarounds — detected explicitly so it resolves to its own profile instead of falling
	// through to the old not-Mali-therefore-Adreno guess.
	else if (std::strstr(renderer_str, "Apple"))
	{
		Console.WriteLn(Color_StrongCyan, "GL: Apple GPU detected.");
		vendor_id_apple = true;
	}

#if defined(__ANDROID__)
	// ANGLE (GLES-on-Vulkan) reports the underlying GPU in GL_RENDERER, e.g.
	// "ANGLE (ARM, Vulkan 1.1.177 (Mali-G77 MC9 ...))", so the native Mali profile and its vendor
	// extensions (GL_ARM_shader_framebuffer_fetch, tile optimisations) engage through ANGLE too —
	// and that is desirable: they are a large performance win on Mali (dropping them measurably
	// regressed FPS on Mali-G615). The Mali-G77 crash under ANGLE was NOT these hacks but stale
	// program binaries from the native driver being fed to ANGLE's glProgramBinary(); that is
	// fixed at the source in GLShaderCache (driver-keyed cache), so no per-GPU profile gating here.
	//
	// The driver context below feeds the driver-bug database (ported from EmuCoreX/sashkinbro with
	// his approval). GL has no equivalent of VkPhysicalDeviceDriverProperties, so the renderer and
	// version strings are all the identity there is; the resolver parses the vendor-specific build
	// tag out of them. Nothing here changes behaviour on its own — every workaround it can turn on
	// is off unless a rule matches this exact driver.
	MobileDriverContext driver_context;
	driver_context.api = MobileGpuApi::OpenGL;
	driver_context.driver_name = renderer_str;
	driver_context.api_version_string = gl_version_str;
	const GpuProfileSelection profile_selection = GpuProfileDetector::Resolve(
		GSConfig.AndroidGpuProfileOverride, vendor_str, renderer_str, driver_context);
	SetRuntimeGPUProfile(profile_selection.runtime_profile);
	SetMobileGPUIdentity(profile_selection.gpu);
	SetMobileGSTuning(profile_selection.gs_tuning);
	SetMobileDriverProfile(profile_selection.driver);
	SetMediaTekSoC(profile_selection.is_mediatek_soc);
	Console.WriteLn("GL: GPU profile override='%s' resolved='%s' driver='%s' version=%u.%u.%u.%u "
					"rules=%u bugs=%016llx workarounds=%016llx.",
		GpuProfileDetector::OverrideToConfigString(profile_selection.override_mode),
		GpuProfileDetector::RuntimeProfileToString(profile_selection.runtime_profile),
		GpuProfileDetector::DriverToString(profile_selection.driver.driver),
		static_cast<unsigned>(profile_selection.driver.version.major),
		static_cast<unsigned>(profile_selection.driver.version.minor),
		static_cast<unsigned>(profile_selection.driver.version.patch),
		static_cast<unsigned>(profile_selection.driver.version.build),
		static_cast<unsigned>(profile_selection.driver.matched_rule_count),
		static_cast<unsigned long long>(profile_selection.driver.bugs),
		static_cast<unsigned long long>(profile_selection.driver.workarounds));
	DevCon.WriteLn("GL: GPU profile hints: %s", profile_selection.hints.c_str());
	bool use_mali_profile = IsMaliGPUProfile();
	bool use_adreno_profile = IsAdrenoGPUProfile();
	bool use_powervr_profile = IsPowerVRGPUProfile();
#else
	// ★ Was `vendor_id_mali ? Mali : Adreno`, which claimed ADRENO for every non-Mali desktop GPU —
	// NVIDIA, AMD, Intel and Apple Silicon all identified as Adreno. The locals below were already
	// correct (real per-vendor detection), so only the member misfired, which is why it hid: it
	// surfaced as Adreno-only workarounds engaging on an M2 (reported by bmd: "GL: Adreno - routing
	// depth feedback through the depth sampler"). Mirror the locals instead of guessing, and fall
	// back to Unknown — desktop GPUs are not tilers and want none of the mobile vendor paths.
	SetRuntimeGPUProfile(vendor_id_mali    ? RuntimeGpuProfile::Mali :
						 vendor_id_adreno  ? RuntimeGpuProfile::Adreno :
						 vendor_id_apple   ? RuntimeGpuProfile::Apple :
											 RuntimeGpuProfile::Unknown);
	bool use_mali_profile = vendor_id_mali;
	bool use_adreno_profile = vendor_id_adreno;
	bool use_powervr_profile = false;
#endif

	GLint major_gl = 0;
	GLint minor_gl = 0;
	glGetIntegerv(GL_MAJOR_VERSION, &major_gl);
	glGetIntegerv(GL_MINOR_VERSION, &minor_gl);
	if (!m_is_gles && !GLAD_GL_VERSION_3_3)
	{
		Host::ReportErrorAsync(
			"GS", fmt::format(TRANSLATE_FS("GSDeviceOGL", "OpenGL renderer is not supported. Only OpenGL {}.{}\n was found"), major_gl, minor_gl));
		return false;
	}

	// Log extension string for debugging purposes.
	Console.WriteLn(fmt::format("GL_VENDOR: {}", reinterpret_cast<const char*>(glGetString(GL_VENDOR))));
	Console.WriteLn(fmt::format("GL_VERSION: {}", gl_version_str));
	Console.WriteLn(fmt::format("GL_RENDERER: {}", reinterpret_cast<const char*>(glGetString(GL_RENDERER))));
	Console.WriteLn(fmt::format(
		"GL_SHADING_LANGUAGE_VERSION: {}", reinterpret_cast<const char*>(glGetString(GL_SHADING_LANGUAGE_VERSION))));
	std::string extensions = "GL_EXTENSIONS:";
	GLint num_extensions = 0;
	glGetIntegerv(GL_NUM_EXTENSIONS, &num_extensions);
	for (GLint i = 0; i < num_extensions; i++)
	{
		const char* ext = reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, i));
		if (ext)
		{
			extensions += ' ';
			extensions.append(ext);
		}
	}
	DevCon.WriteLn(std::move(extensions));

	if (!m_is_gles) {
		if (!GLAD_GL_ARB_shading_language_420pack)
		{
			Host::ReportFormattedErrorAsync(
					"GS", "GL_ARB_shading_language_420pack is not supported, this is required for the OpenGL renderer.");
			return false;
		}

		if (!GLAD_GL_VERSION_4_3 && !GLAD_GL_ARB_copy_image && !GLAD_GL_EXT_copy_image && !GLAD_GL_NV_copy_image)
		{
			Host::AddOSDMessage(
				"GL_ARB_copy_image is not supported, copies will be slower.", Host::OSD_ERROR_DURATION);
		}

		if (!GLAD_GL_VERSION_4_5 && !GLAD_GL_ARB_clip_control &&
			!(m_is_gles && GLAD_GL_EXT_clip_control))
		{
			Host::AddOSDMessage(
				"GL_ARB_clip_control is not supported, depth will be less accurate.", Host::OSD_ERROR_DURATION);
		}
	}


	if (!GLAD_GL_ARB_viewport_array)
	{
		glScissorIndexed = ReplaceGL::ScissorIndexed;
		glViewportIndexedf = ReplaceGL::ViewportIndexedf;
		Console.Warning("GL_ARB_viewport_array is not supported! Function pointer will be replaced.");
	}

	if (!GLAD_GL_ARB_texture_barrier)
	{
		// First try NV only barrier.
		// If that doesn't work then switch to multidraw fb copy.
		if (GLAD_GL_NV_texture_barrier)
			glTextureBarrier = glTextureBarrierNV;
		else
		{
			glTextureBarrier = ReplaceGL::TextureBarrier;
			m_features.multidraw_fb_copy = true;
/*			Host::AddOSDMessage(
				"GL_ARB_texture_barrier is not supported, blending will be slower.", Host::OSD_ERROR_DURATION);*/
		}
	}

	if (!GLAD_GL_ARB_direct_state_access)
	{
		Console.Warning("GL_ARB_direct_state_access is not supported, this will reduce performance.");
		Emulate_DSA::Init();
	}

	// glDrawElementsBaseVertex entered core in GL ES 3.2. ANGLE-over-Vulkan caps its context at
	// ES 3.1, so GLAD leaves the UNSUFFIXED core pointer NULL even though ANGLE advertises
	// GL_OES/EXT_draw_elements_base_vertex (whose suffixed entry points ARE loaded). Every draw
	// path calls the unsuffixed glDrawElementsBaseVertex directly (RenderImGui + DrawIndexedPrimitive),
	// so on ANGLE the first overlay draw jumped to a null pointer and aborted via the fastmem SIGSEGV
	// handler (Mali-G77 / Retroid Pocket 4 Pro). Native Mali/Adreno report ES 3.2 so the core pointer
	// is non-null there — which is why this only bit ANGLE. Alias it to the OES/EXT variant, matching
	// the glTextureBarrier = glTextureBarrierNV replacement above. Fixes all call sites at once.
	if (!glDrawElementsBaseVertex)
	{
		if (GLAD_GL_OES_draw_elements_base_vertex && glDrawElementsBaseVertexOES)
			glDrawElementsBaseVertex = glDrawElementsBaseVertexOES;
		else if (GLAD_GL_EXT_draw_elements_base_vertex && glDrawElementsBaseVertexEXT)
			glDrawElementsBaseVertex = glDrawElementsBaseVertexEXT;
		else
			Console.Error("GL: glDrawElementsBaseVertex is unavailable (no core/OES/EXT) — draws will fail.");
	}

	// glColorMaski (indexed color mask) is core in GLSL ES 3.2 but only extension-provided on
	// an ES 3.1 context; GLAD leaves the unsuffixed pointer null on ANGLE (which caps at ES 3.1),
	// so GSDeviceOGL::OMSetColorMaskState calls a null pointer → SIGSEGV mid-render (VSync→Merge→
	// GetOutput→StretchRect). ANGLE exposes the suffixed variant via GL_OES/EXT_draw_buffers_indexed,
	// so alias to it. Same class of fix as glDrawElementsBaseVertex above.
	if (!glColorMaski)
	{
		if (GLAD_GL_OES_draw_buffers_indexed && glColorMaskiOES)
			glColorMaski = glColorMaskiOES;
		else if (GLAD_GL_EXT_draw_buffers_indexed && glColorMaskiEXT)
			glColorMaski = glColorMaskiEXT;
		else
			Console.Error("GL: glColorMaski is unavailable (no core/OES/EXT) — draws will fail.");
	}

	// Don't use PBOs when we don't have ARB_buffer_storage, orphaning buffers probably ends up worse than just
	// using the normal texture update routines and letting the driver take care of it.
	if (!m_is_gles) {
		m_bugs.buggy_pbo = !GLAD_GL_VERSION_4_4 && !GLAD_GL_ARB_buffer_storage && !GLAD_GL_EXT_buffer_storage;
	} else {
		// Mirrors the desktop check: PBOs are useful only when EXT_buffer_storage
		// (the GLES port of buffer_storage) is available so we can pin a
		// persistent-mapped staging region. Without it the orphaning fallback
		// is slower than letting the driver handle the upload directly. The
		// previous form here had the test inverted (set buggy=TRUE when the
		// extension WAS supported), disabling PBOs on every modern Adreno/Mali
		// device. Aligning with the desktop branch's polarity.
		m_bugs.buggy_pbo = !GLAD_GL_EXT_buffer_storage;
	}

	if (m_bugs.buggy_pbo)
		Console.Warning("GL: Not using PBOs for texture uploads because buffer_storage is unavailable.");

	// Give the user the option to disable PBO usage for downloads.
	// Most drivers seem to be faster with PBO.
	m_disable_download_pbo = Host::GetBoolSettingValue("EmuCore/GS", "DisableGLDownloadPBO", false);
	if (m_disable_download_pbo)
		Console.Warning("GL: Not using PBOs for texture downloads, this may reduce performance.");

	// optional features based on context
	m_features.broken_point_sampler = false;
	m_features.primitive_id = true;

	// Apple GPUs miscompare depth written from the shader (the PS2 32-bit Z floor) against the
	// fixed-function interpolation a later read-only pass tests with, so a GEQUAL retest of the
	// same geometry drops out along shared triangle edges and the layer underneath shows through
	// as pinpoints -- God of War II's Athena statue, and dark walls in Black. Reproduces here
	// identically under GL and Vulkan (748 stray pixels either way), so it is the GPU, not the
	// API. See the matching gate in GSDeviceVK::CheckFeatures for the measurements. Mali is
	// deliberately not included: the Vulkan path opts it out for early-ZS, but that has not been
	// tested on a Mali GL driver.
	m_features.no_ps2_z_quantization = GSConfig.DisablePS2DepthQuantization || vendor_id_apple;

	// GLES may omit dual-source blending (GL_EXT/ARB_blend_func_extended); desktop GL always has it.
	// When absent, GSRendererHW emulates SRC1 blend equations in-shader per-draw rather than forcing
	// a global high blending-accuracy level (Mali no longer needs Blending=Max). From sashkinbro/EmuCoreX.
	m_features.dual_source_blend =
		!m_is_gles || GLAD_GL_EXT_blend_func_extended || GLAD_GL_ARB_blend_func_extended;

	m_features.framebuffer_fetch = (GLAD_GL_ARM_shader_framebuffer_fetch || GLAD_GL_EXT_shader_framebuffer_fetch);

	// The Mali r44p1 blob loses the rendering context under the in-tile framebuffer-fetch blend path,
	// exactly as it loses the Vulkan device under attachment-feedback-loop (VK_ERROR_DEVICE_LOST on
	// effectively every game -- Mali-G615 r44p1). Mirror the Vulkan-side r44p1 gate (see GSDeviceVK.cpp)
	// and drop this one blob to the non-fetch (copy) blend path. Narrow by driver version, not vendor,
	// so other (working) Mali blobs keep the fast path. GL_VERSION reads e.g. "OpenGL ES 3.2 v1.r44p1-...".
	if (m_features.framebuffer_fetch)
	{
		const char* gl_version = reinterpret_cast<const char*>(glGetString(GL_VERSION));
		if (gl_version && std::strstr(gl_version, "r44p1"))
		{
			Console.WriteLn("Mali r44p1: disabling framebuffer fetch (GL context-lost workaround; matches the Vulkan gate).");
			m_features.framebuffer_fetch = false;
		}
	}

	if (m_features.framebuffer_fetch && GSConfig.DisableFramebufferFetch)
	{
		Host::AddOSDMessage(
			"Framebuffer fetch was found but is disabled. This will reduce performance.", Host::OSD_ERROR_DURATION);
		m_features.framebuffer_fetch = false;
	}

	if (GSConfig.OverrideTextureBarriers == 0)
	{
		m_features.texture_barrier = m_features.framebuffer_fetch; // Force Disabled
		m_features.multidraw_fb_copy = false;
		Host::AddOSDMessage(
			"Texture Barrier is disabled, blending will not be accurate.", Host::OSD_ERROR_DURATION);
	}
	else if (GSConfig.OverrideTextureBarriers == 1)
	{
		// No texture barriers supported so try to use memory barrier.
		// Not guaranteed to work so only enable it in Force Enabled mode.
		if (!GLAD_GL_ARB_texture_barrier && !GLAD_GL_NV_texture_barrier && GLAD_GL_ARB_shader_image_load_store)
			glTextureBarrier = ReplaceGL::MemoryBarrierAsTextureBarrier;

		m_features.texture_barrier = true; // Force Enabled
		m_features.multidraw_fb_copy = false;
	}
	else
		m_features.texture_barrier = m_features.framebuffer_fetch || GLAD_GL_ARB_texture_barrier || GLAD_GL_NV_texture_barrier;

	m_features.provoking_vertex_last = true;
	m_features.dxt_textures = GLAD_GL_EXT_texture_compression_s3tc;
	m_features.bptc_textures =
		GLAD_GL_VERSION_4_2 || GLAD_GL_ARB_texture_compression_bptc || GLAD_GL_EXT_texture_compression_bptc;
	m_features.prefer_new_textures = false;
	m_features.stencil_buffer = true;
	m_features.test_and_sample_depth = true;
	// Auto select chooses depth-as-rt as it appears to be more compatible across hardware.
	m_features.depth_feedback = GSConfig.DepthFeedbackMode == GSDepthFeedbackMode::Depth;
	if (!m_features.texture_barrier && m_features.multidraw_fb_copy)
	{
		// Multidraw fb copy can do depth feedback just fine
		m_features.depth_feedback |= GSConfig.DepthFeedbackMode == GSDepthFeedbackMode::Auto;
	}

	// ARMSX2 (Adreno GLES only; inert on every other GPU/profile). Adreno's driver
	// rejects a fragment shader declaring TWO framebuffer-fetch `inout` outputs (o_col0
	// colour + o_col1 depth), which the depth-as-colour SW-Z path emits for accurate-
	// alpha-test draws -> link failure -> garbage (Everybody's Golf 4 / Minna no Golf 4).
	// Route depth feedback through the depth path (a single fetch output) so it links, and
	// read prior depth via the coherent ARM depth-stencil fetch (gl_LastFragDepthARM) when
	// available -- the mode-1 depth sampler read is incoherent on GLES (no barrier on a
	// sampled depth attachment) and makes occluded triangles poke through as white shards.
	// Only overrides Auto; an explicit DepthFeedbackMode choice is honoured. The GPU
	// profile is already resolved above (SetRuntimeGPUProfile), so IsAdrenoGPUProfile()
	// is valid here.
	if (m_features.framebuffer_fetch && IsAdrenoGPUProfile() &&
		GSConfig.DepthFeedbackMode == GSDepthFeedbackMode::Auto)
	{
		m_features.depth_feedback = true;
		m_arm_depth_fetch = GLAD_GL_ARM_shader_framebuffer_fetch_depth_stencil;
		Console.WriteLn(m_arm_depth_fetch
			? "GL: Adreno - depth feedback via coherent ARM depth-stencil fetch (gl_LastFragDepthARM)."
			: "GL: Adreno - routing depth feedback through the depth sampler "
			  "(avoids the dual framebuffer-fetch output link failure).");
	}

	// Mobile tile-based GPU profiles. Both Mali and Adreno prefer fresh
	// textures over reused ones (avoids tile-flush stalls on partial
	// writes), so the texture-pool hint is shared. Mali additionally
	// pins framebuffer_fetch to the ARM extension and (on Auto) reuses
	// it as the texture-barrier substitute — matched by the
	// `#if GPU_PROFILE_MALI` branch in tfx_fs.glsl which picks
	// gl_LastFragColorARM. Adreno + Generic fall through to the EXT/PLS
	// inout path in the shader's `#else` arm; GPU_PROFILE_ADRENO is
	// emitted but not currently consumed by any shader.
	if (use_mali_profile || use_adreno_profile || use_powervr_profile)
		m_features.prefer_new_textures = true;

#if defined(__ANDROID__)
	// Narrows the profile-driven choice above with the per-model tuning; deliberately an
	// AND rather than an assignment, so a part whose table entry says "reuse" cannot turn
	// preference back on for a profile that did not want it. sashkinbro/EmuCoreX.
	m_features.prefer_new_textures &= GetMobileGSTuning().prefer_new_textures;
	if (GetMobileGSTuning().force_partial_texture_preloading &&
		GSConfig.TexturePreloading == TexturePreloadingLevel::Full)
	{
		GSConfig.TexturePreloading = TexturePreloadingLevel::Partial;
		Console.Warning("GL: Mobile GS %s/%s profile lowered texture preloading to partial.",
			GpuProfileDetector::RuntimeProfileToString(GetRuntimeGPUProfile()),
			GetMobileGPUIdentity().name.c_str());
	}
#endif

	if (use_mali_profile)
	{
		// Mali path prefers ARM_shader_framebuffer_fetch (gl_LastFragColorARM) because
		// the EXT inout path has been broken across every tested Mali driver. If a
		// device was force-overridden to Mali but lacks ARM fbfetch (rare but
		// possible), demote to PowerVR profile which uses the same EXT/PLS path the
		// catch-all default uses.
		if (GLAD_GL_ARM_shader_framebuffer_fetch)
		{
			Console.WriteLn(Color_Yellow, "GL: Applying Mali-specific optimizations for tile-based rendering.");
			m_features.framebuffer_fetch = true;
			if (GSConfig.OverrideTextureBarriers == -1)
			{
				m_features.texture_barrier = m_features.framebuffer_fetch;
				Console.WriteLn("GL: Mali optimization - using ARM framebuffer fetch over texture barriers.");
			}
		}
		else
		{
			Console.Warning("GL: Mali profile selected but ARM framebuffer fetch is unavailable; demoting to PowerVR/EXT profile.");
			SetRuntimeGPUProfile(RuntimeGpuProfile::PowerVR);
			use_mali_profile = false;
			use_powervr_profile = true;
		}
	}

	if (use_powervr_profile)
	{
		// PowerVR (Imagination) is tile-based like Mali but ships EXT/PLS fbfetch
		// (PLS originated on PowerVR). framebuffer_fetch + texture_barrier values
		// from line ~882/908 already reflect the EXT path, so no override needed —
		// just confirm the path is wired up.
		Console.WriteLn(Color_Yellow, "GL: PowerVR profile active (EXT/PLS framebuffer fetch).");
	}
	else if (use_adreno_profile)
	{
		Console.WriteLn(Color_Cyan, "GL: Adreno profile active (EXT/PLS framebuffer fetch).");

		// Adreno's GLES driver rejects a fragment shader that declares TWO
		// framebuffer-fetch `inout` outputs. The depth-as-colour feedback path
		// (DEPTH_FEEDBACK_SUPPORT 2) emits exactly that whenever the colour output
		// already needs fetch AND a SW-Z depth draw is in flight -- o_col0 (colour
		// fetch) at location 0 and o_col1 (depth fetch) at location 1 both become
		// `inout`. That combination is produced by the accurate-alpha-test RGB-only
		// + depth-write path, so any game carrying accurateAlphaTest (e.g. Everybody's
		// Golf 4 / Minna no Golf 4, SCKA-20057 / SCPS-15059) fails to link those draws
		// -> "Output o_col1 location or component exceeds max allowed" -> garbage
		// (black-boxed faces, a floating RT rectangle, blue bars). Vulkan is unaffected
		// (real depth attachment, no second fetch output). Route depth feedback through
		// the real depth sampler (DEPTH_FEEDBACK_SUPPORT 1) so only o_col0 is a fetch
		// output and the program links. test_and_sample_depth is already true above,
		// and texture_barrier==true here keeps the DS-clone path (bind at ~3402) inert.
		// Only override Auto -- an explicit DepthFeedbackMode choice is honoured.
		if (m_features.framebuffer_fetch && GSConfig.DepthFeedbackMode == GSDepthFeedbackMode::Auto)
		{
			m_features.depth_feedback = true;
			// The mode-1 depth SAMPLER read is incoherent on GLES (no texture_barrier
			// for a sampled depth attachment) -> stale reads make occluded/interior
			// triangles poke through as white shards. When the coherent ARM depth-
			// stencil fetch extension is present, read prior depth via gl_LastFragDepthARM
			// instead (tile-local, one output, no sampler, no feedback-loop bind).
			m_arm_depth_fetch = GLAD_GL_ARM_shader_framebuffer_fetch_depth_stencil;
			Console.WriteLn(m_arm_depth_fetch
				? "GL: Adreno - depth feedback via coherent ARM depth-stencil fetch (gl_LastFragDepthARM)."
				: "GL: Adreno - routing depth feedback through the depth sampler "
				  "(avoids the dual framebuffer-fetch output link failure).");
		}
	}

	{
		const bool has_arm_fetch = GLAD_GL_ARM_shader_framebuffer_fetch;
		const bool has_ext_fetch = GLAD_GL_EXT_shader_framebuffer_fetch;
		const bool has_pls_fetch = GLAD_GL_EXT_shader_pixel_local_storage;
		Console.WriteLn("GL: Framebuffer fetch extension caps: arm=%d ext=%d pls=%d.",
			has_arm_fetch ? 1 : 0, has_ext_fetch ? 1 : 0, has_pls_fetch ? 1 : 0);

		const char* active_profile_name = use_mali_profile ? "Mali" :
			(use_powervr_profile ? "PowerVR" :
			(use_adreno_profile ? "Adreno" : "Generic"));
		const char* active_fetch_backend = "None";
		if (m_features.framebuffer_fetch)
		{
			if (use_mali_profile)
				active_fetch_backend = "ARM";
			else if (has_ext_fetch || has_pls_fetch)
				active_fetch_backend = "EXT/PLS";
			else if (has_arm_fetch)
				active_fetch_backend = "ARM";
		}
		Console.WriteLn("GL: Active framebuffer fetch backend (%s profile): %s.", active_profile_name, active_fetch_backend);

		if (use_mali_profile && !has_arm_fetch)
			Console.Warning("GL: Mali profile selected but ARM framebuffer fetch is unavailable; using non-fetch fallback.");
	}

	if (GLAD_GL_ARB_shader_storage_buffer_object)
	{
		GLint max_vertex_ssbos = 0;
		glGetIntegerv(GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS, &max_vertex_ssbos);
		DevCon.WriteLn("GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS: %d", max_vertex_ssbos);
		m_features.vs_expand = (!GSConfig.DisableVertexShaderExpand && max_vertex_ssbos > 0 && GLAD_GL_ARB_gpu_shader5);
	}
	if (!m_features.vs_expand)
		Console.Warning("GL: Vertex expansion is not supported. This will reduce performance.");

	GLint point_range[2] = {};
	glGetIntegerv(GL_ALIASED_POINT_SIZE_RANGE, point_range);
	m_features.point_expand =
		(point_range[0] <= GSConfig.UpscaleMultiplier && point_range[1] >= GSConfig.UpscaleMultiplier);
	m_features.line_expand = false;

	GLint max_texture_size = 1024;
	glGetIntegerv(GL_MAX_TEXTURE_SIZE, &max_texture_size);
	m_max_texture_size = std::max(1024u, static_cast<u32>(max_texture_size));

	Console.WriteLn("GL: Using %s for point expansion, %s for line expansion and %s for sprite expansion.",
		m_features.point_expand ? "hardware" : (m_features.vs_expand ? "vertex expanding" : "UNSUPPORTED"),
		m_features.line_expand ? "hardware" : (m_features.vs_expand ? "vertex expanding" : "UNSUPPORTED"),
		m_features.vs_expand ? "vertex expanding" : "CPU");

	if (!GLAD_GL_ARB_conservative_depth)
	{
		Console.Warning("GLAD_GL_ARB_conservative_depth is not supported. This will reduce performance.");
	}
	
	m_features.aa1 = GSConfig.HWAA1 && m_features.vs_expand && m_features.feedback_loops();
	
	return true;
}

void GSDeviceOGL::SetSwapInterval()
{
	if (m_window_info.type == WindowInfo::Type::Surfaceless)
		return;

	// OpenGL does not support mailbox, only effectively FIFO.
	// Fall back to manual throttling in this case.
	m_vsync_mode = (m_vsync_mode == GSVSyncMode::Mailbox) ? GSVSyncMode::FIFO : m_vsync_mode;

	// Window framebuffer has to be bound to call SetSwapInterval.
	s32 interval = static_cast<s32>(m_vsync_mode == GSVSyncMode::FIFO);
	// ARM Mali GLES breaks with eglSwapInterval(0): after a handful of swaps the surface
	// stops presenting entirely (frozen screen). Never pass 0 on Mali — force interval 1.
	// A forced-on vsync beats a frozen display, and the FIFO present pacer still lets
	// fast-forward exceed the panel rate via dropped presents. (cf. Dolphin BUG_BROKEN_VSYNC)
	if (interval == 0 && IsMaliGPUProfile())
		interval = 1;
	GLint current_fbo = 0;
	glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &current_fbo);
	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);

	if (!m_gl_context->SetSwapInterval(interval))
		WARNING_LOG("GL: Failed to set swap interval to {}", interval);

	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, current_fbo);
}

void GSDeviceOGL::DestroyResources()
{
	m_shader_cache.Close();

	if (m_palette_ss != 0)
		glDeleteSamplers(1, &m_palette_ss);

	m_programs.clear();

	for (GSDepthStencilOGL* ds : m_om_dss)
		delete ds;

	if (m_ps_ss[0] != 0)
		glDeleteSamplers(std::size(m_ps_ss), m_ps_ss);

	m_imgui.ps.Destroy();
	if (m_imgui.vao != 0)
		glDeleteVertexArrays(1, &m_imgui.vao);

	m_cas.upscale_ps.Destroy();
	m_cas.sharpen_ps.Destroy();

	m_shadeboost.ps.Destroy();

	for (GLProgram& prog : m_date.primid_ps)
		prog.Destroy();
	delete m_date.dss;

	m_fxaa.ps.Destroy();

	for (GLProgram& prog : m_present)
		prog.Destroy();

	for (auto& prog : m_convert.ps)
		prog.Destroy();

	delete m_convert.dss;
	delete m_convert.dss_write;

	for (GLProgram& prog : m_interlace.ps)
		prog.Destroy();

	for (GLProgram& prog : m_merge_obj.ps)
		prog.Destroy();

	m_fragment_uniform_stream_buffer.reset();
	m_vertex_uniform_stream_buffer.reset();
	m_vertex_push_constants_stream_buffer.reset();

	glBindVertexArray(0);
	// Delete the expand VAO here (not m_expand_ibo, which is a buffer object and is
	// correctly freed with glDeleteBuffers below). The old code deleted m_expand_ibo
	// as a VAO — a no-op — so m_expand_vao leaked on every device teardown/recreate.
	if (m_expand_vao != 0)
		glDeleteVertexArrays(1, &m_expand_vao);
	if (m_vao != 0)
		glDeleteVertexArrays(1, &m_vao);
	if (m_dummy_vao != 0)
		glDeleteVertexArrays(1, &m_dummy_vao);

	m_index_stream_buffer.reset();
	m_expand_index_stream_buffer.reset();
	m_vertex_stream_buffer.reset();
	m_texture_upload_buffer.reset();
	if (m_expand_ibo)
		glDeleteBuffers(1, &m_expand_ibo);

	if (m_fbo != 0)
	{
		glDeleteFramebuffers(1, &m_fbo);
		m_fbo = 0;
	}
	if (m_fbo_read != 0)
	{
		glDeleteFramebuffers(1, &m_fbo_read);
		m_fbo_read = 0;
	}
	if (m_fbo_write != 0)
	{
		glDeleteFramebuffers(1, &m_fbo_write);
		m_fbo_write = 0;
	}

	GLState::fbo = 0;
}

bool GSDeviceOGL::UpdateWindow()
{
	pxAssert(m_gl_context);

	DestroySurface();

	if (!AcquireWindow(false))
		return false;

	if (!m_gl_context->ChangeSurface(m_window_info))
	{
		Console.Error("GL: Failed to change surface");
		return false;
	}

	m_window_info = m_gl_context->GetWindowInfo();

	if (m_window_info.type != WindowInfo::Type::Surfaceless)
	{
		// reset vsync rate, since it (usually) gets lost
		SetSwapInterval();
		RenderBlankFrame();
	}

	return true;
}

void GSDeviceOGL::ResizeWindow(u32 new_window_width, u32 new_window_height, float new_window_scale)
{
	m_window_info.surface_scale = new_window_scale;
	if (m_window_info.type == WindowInfo::Type::Surfaceless ||
		(m_window_info.surface_width == new_window_width &&
			m_window_info.surface_height == new_window_height))
	{
		return;
	}

	m_gl_context->ResizeSurface(new_window_width, new_window_height);
	m_window_info = m_gl_context->GetWindowInfo();
}

bool GSDeviceOGL::SupportsExclusiveFullscreen() const
{
	return false;
}

void GSDeviceOGL::DestroySurface()
{
	m_window_info = {};
	if (!m_gl_context->ChangeSurface(m_window_info))
		Console.Error("GL: Failed to switch to surfaceless");
}

std::string GSDeviceOGL::GetDriverInfo() const
{
	const char* gl_vendor = reinterpret_cast<const char*>(glGetString(GL_VENDOR));
	const char* gl_renderer = reinterpret_cast<const char*>(glGetString(GL_RENDERER));
	const char* gl_version = reinterpret_cast<const char*>(glGetString(GL_VERSION));
	const char* gl_shading_language_version = reinterpret_cast<const char*>(glGetString(GL_SHADING_LANGUAGE_VERSION));
	return fmt::format(
		"OpenGL Context:\n{}\n{} {}\nGLSL: {}", gl_version, gl_vendor, gl_renderer, gl_shading_language_version);
}

GSDevice::PresentResult GSDeviceOGL::DoBeginPresent(bool frame_skip)
{
	if (frame_skip || m_window_info.type == WindowInfo::Type::Surfaceless)
		return PresentResult::FrameSkipped;

	// Get the pipeline statistics for this frame before postprocessing.
	if (m_gpu_pipeline_statistics_enabled)
		PopPipelineStatisticsQuery();

	OMSetFBO(0);
	OMSetColorMaskState();

	// On TBDR, hint that the default framebuffer's prior content is throwaway
	// before the tile is loaded for the present quad. The color attachment is
	// fully overwritten by the clear+blit below, and depth/stencil are never
	// used at all on the system framebuffer. Default-FBO uses GL_COLOR / DEPTH
	// / STENCIL (not GL_*_ATTACHMENT). Pure TBDR tile-bandwidth win and inert on
	// desktop immediate renderers, so gated to GLES to keep the desktop path canonical.
	if (m_is_gles)
	{
		const GLenum attachments[] = {GL_COLOR, GL_DEPTH, GL_STENCIL};
		glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(attachments), attachments);
	}

	glDisable(GL_SCISSOR_TEST);
	glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
	glClear(GL_COLOR_BUFFER_BIT);
	glEnable(GL_SCISSOR_TEST);

	const GSVector2i size = GetWindowSize();
	SetViewport(size);
	SetScissor(GSVector4i::loadh(size));

	return PresentResult::OK;
}

void GSDeviceOGL::EndPresent()
{
	RenderImGui();

	if (m_gpu_timing_enabled)
		PopTimestampQuery();

	// Discard the default framebuffer's depth/stencil before the swap. We
	// never wrote anything meaningful to them, so on TBDR drivers writing
	// the tile back to system memory at SwapBuffers is wasted bandwidth.
	// Color is preserved (it's what gets presented). GLES/TBDR-only (inert on desktop).
	if (m_is_gles)
	{
		const GLenum attachments[] = {GL_DEPTH, GL_STENCIL};
		glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(attachments), attachments);
	}

	m_gl_context->SwapBuffers();

	if (m_gpu_timing_enabled)
		KickTimestampQuery();

	if (m_gpu_pipeline_statistics_enabled)
		KickPipelineStatisticsQuery();
}

void GSDeviceOGL::CreateTimestampQueries()
{
	glGenQueries(static_cast<u32>(m_timestamp_queries.size()), m_timestamp_queries.data());
	KickTimestampQuery();
}

void GSDeviceOGL::DestroyTimestampQueries()
{
	if (m_timestamp_queries[0] == 0)
		return;

	if (m_timestamp_query_started)
		glEndQuery(GL_TIME_ELAPSED);

	glDeleteQueries(static_cast<u32>(m_timestamp_queries.size()), m_timestamp_queries.data());
	m_timestamp_queries.fill(0);
	m_read_timestamp_query = 0;
	m_write_timestamp_query = 0;
	m_waiting_timestamp_queries = 0;
	m_timestamp_query_started = false;
}

void GSDeviceOGL::PopTimestampQuery()
{
	while (m_waiting_timestamp_queries > 0)
	{
#if defined(__ANDROID__)
		// GLES doesn't expose glGetQueryObjectiv / glGetQueryObjectui64v; both
		// availability and result use the u32 form. Caps at ~4.29s of
		// nanoseconds — fine for per-frame timing. Provided by the
		// EXT_disjoint_timer_query extension (GL_TIME_ELAPSED_EXT === 0x88BF
		// === GL_TIME_ELAPSED here).
		//
		// Prior version of this branch was broken: it called glBeginQuery on
		// the read slot then immediately tried to read its result (always 0,
		// query never ended) and incremented m_waiting_timestamp_queries
		// instead of decrementing — accumulator stayed stuck at 0 in HW
		// renderer OSD ("GPU: 0%" symptom).
		GLuint available = 0;
		glGetQueryObjectuiv(m_timestamp_queries[m_read_timestamp_query], GL_QUERY_RESULT_AVAILABLE, &available);
		if (!available)
			break;

		GLuint result = 0;
		glGetQueryObjectuiv(m_timestamp_queries[m_read_timestamp_query], GL_QUERY_RESULT, &result);
		m_accumulated_gpu_time += static_cast<float>(static_cast<double>(result) / 1000000.0);
#else
		GLint available = 0;
		glGetQueryObjectiv(m_timestamp_queries[m_read_timestamp_query], GL_QUERY_RESULT_AVAILABLE, &available);

		if (!available)
			break;

		u64 result = 0;
		glGetQueryObjectui64v(m_timestamp_queries[m_read_timestamp_query], GL_QUERY_RESULT, &result);
		m_accumulated_gpu_time += static_cast<float>(static_cast<double>(result) / 1000000.0);
#endif
		m_read_timestamp_query = (m_read_timestamp_query + 1) % NUM_TIMESTAMP_QUERIES;
		m_waiting_timestamp_queries--;
	}

	if (m_timestamp_query_started)
	{
		glEndQuery(GL_TIME_ELAPSED);
		m_write_timestamp_query = (m_write_timestamp_query + 1) % NUM_TIMESTAMP_QUERIES;
		m_timestamp_query_started = false;
		m_waiting_timestamp_queries++;
	}
}

void GSDeviceOGL::KickTimestampQuery()
{
	if (m_timestamp_query_started || m_waiting_timestamp_queries == NUM_TIMESTAMP_QUERIES)
		return;

	glBeginQuery(GL_TIME_ELAPSED, m_timestamp_queries[m_write_timestamp_query]);
	m_timestamp_query_started = true;
}

bool GSDeviceOGL::SetGPUTimingEnabled(bool enabled)
{
	if (m_gpu_timing_enabled == enabled)
		return true;

	m_gpu_timing_enabled = enabled;
	if (m_gpu_timing_enabled)
		CreateTimestampQueries();
	else
		DestroyTimestampQueries();

	return true;
}

float GSDeviceOGL::GetAndResetAccumulatedGPUTime()
{
	const float value = m_accumulated_gpu_time;
	m_accumulated_gpu_time = 0.0f;
	return value;
}

// NOTE: These GL pipeline-statistics queries are desktop-GL only. GLES (Android)
// has neither GL_ARB_pipeline_statistics_query nor the glGetQueryObjectiv /
// glGetQueryObjectui64v result readers (see PopTimestampQuery for the same
// GLES gap), so the whole path is compiled out under __ANDROID__. On Android
// m_gpu_pipeline_statistics_supported stays false and these are never invoked;
// the OSD line just shows 0 / degrades to n/a. Real stats come from Vulkan.
void GSDeviceOGL::PopPipelineStatisticsQuery()
{
#if !defined(__ANDROID__)
	while (m_waiting_pipeline_statistics_queries > 0)
	{
		GLint available[2] = {};
		glGetQueryObjectiv(m_pipeline_statistics_queries[m_read_pipeline_statistics_query][0], GL_QUERY_RESULT_AVAILABLE, &available[0]);
		glGetQueryObjectiv(m_pipeline_statistics_queries[m_read_pipeline_statistics_query][1], GL_QUERY_RESULT_AVAILABLE, &available[1]);

		if (!(available[0] && available[1]))
			break;

		GPUPipelineStatistics stats = {};
		glGetQueryObjectui64v(m_pipeline_statistics_queries[m_read_pipeline_statistics_query][0], GL_QUERY_RESULT, &stats.vs_invocations);
		glGetQueryObjectui64v(m_pipeline_statistics_queries[m_read_pipeline_statistics_query][1], GL_QUERY_RESULT, &stats.ps_invocations);
		m_accumulated_gpu_pipeline_statistics.vs_invocations += stats.vs_invocations;
		m_accumulated_gpu_pipeline_statistics.ps_invocations += stats.ps_invocations;
		m_read_pipeline_statistics_query = (m_read_pipeline_statistics_query + 1) % NUM_PIPELINE_STATISTICS_QUERIES;
		m_waiting_pipeline_statistics_queries--;
	}

	if (m_pipeline_statistics_query_started)
	{
		glEndQuery(GL_VERTEX_SHADER_INVOCATIONS_ARB);
		glEndQuery(GL_FRAGMENT_SHADER_INVOCATIONS_ARB);

		m_write_pipeline_statistics_query = (m_write_pipeline_statistics_query + 1) % NUM_PIPELINE_STATISTICS_QUERIES;
		m_pipeline_statistics_query_started = false;
		m_waiting_pipeline_statistics_queries++;
	}
#endif
}

void GSDeviceOGL::KickPipelineStatisticsQuery()
{
#if !defined(__ANDROID__)
	if (m_pipeline_statistics_query_started || m_waiting_pipeline_statistics_queries == NUM_PIPELINE_STATISTICS_QUERIES)
		return;

	glBeginQuery(GL_VERTEX_SHADER_INVOCATIONS_ARB, m_pipeline_statistics_queries[m_write_pipeline_statistics_query][0]);
	glBeginQuery(GL_FRAGMENT_SHADER_INVOCATIONS_ARB, m_pipeline_statistics_queries[m_write_pipeline_statistics_query][1]);
	m_pipeline_statistics_query_started = true;
#endif
}

void GSDeviceOGL::CreatePipelineStatisticsQueries()
{
#if !defined(__ANDROID__)
	for (int i = 0; i < NUM_PIPELINE_STATISTICS_QUERIES; i++)
	{
		glGenQueries(2, m_pipeline_statistics_queries[i].data());
	}
	KickPipelineStatisticsQuery();
#endif
}

void GSDeviceOGL::DestroyPipelineStatisticsQueries()
{
#if !defined(__ANDROID__)
	if (m_pipeline_statistics_queries[0][0] == 0)
		return;

	if (m_pipeline_statistics_query_started)
	{
		glEndQuery(GL_VERTEX_SHADER_INVOCATIONS_ARB);
		glEndQuery(GL_FRAGMENT_SHADER_INVOCATIONS_ARB);
	}

	for (size_t i = 0; i < m_pipeline_statistics_queries.size(); i++)
	{
		glDeleteQueries(2, m_pipeline_statistics_queries[i].data());
		m_pipeline_statistics_queries[i].fill(0);
	}
	m_read_pipeline_statistics_query = 0;
	m_write_pipeline_statistics_query = 0;
	m_waiting_pipeline_statistics_queries = 0;
	m_pipeline_statistics_query_started = false;
#endif
}

GPUPipelineStatistics GSDeviceOGL::GetAndResetAccumulatedGPUPipelineStatistics()
{
	GPUPipelineStatistics stats = m_accumulated_gpu_pipeline_statistics;
	m_accumulated_gpu_pipeline_statistics = {};
	return stats;
}

bool GSDeviceOGL::SetGPUPipelineStatisticsEnabled(bool enabled)
{
	if (m_gpu_pipeline_statistics_enabled == enabled)
		return true;

	m_gpu_pipeline_statistics_enabled = enabled && m_gpu_pipeline_statistics_supported;

	if (m_gpu_pipeline_statistics_enabled)
		CreatePipelineStatisticsQueries();
	else
		DestroyPipelineStatisticsQueries();

	return true;
}

void GSDeviceOGL::DrawPrimitive()
{
	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
	glDrawArrays(m_draw_topology, m_vertex.start, m_vertex.count);
}

void GSDeviceOGL::DrawIndexedPrimitive()
{
	DrawIndexedPrimitive(0, m_index.count);
}

void GSDeviceOGL::DrawIndexedPrimitive(int offset, int count)
{
	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
	glDrawElementsBaseVertex(m_draw_topology, count, GL_UNSIGNED_SHORT,
		reinterpret_cast<void*>((static_cast<u32>(m_index.start) + static_cast<u32>(offset)) * sizeof(u16)),
		static_cast<GLint>(m_vertex.start));
}

void GSDeviceOGL::DrawIndexedPrimitiveVSExpand(int offset, int count, bool vs_indexing, int vs_indexing_expansion)
{
	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
	if (vs_indexing)
	{
		VSSetPushConstants(m_vertex.start, m_index.start + offset);
		glDrawArrays(m_draw_topology, 0, count * vs_indexing_expansion);
	}
	else
	{
		VSSetPushConstants(m_vertex.start);
		glDrawElementsBaseVertex(m_draw_topology, count, GL_UNSIGNED_SHORT,
			reinterpret_cast<void*>((static_cast<u32>(m_index.start) + static_cast<u32>(offset)) * sizeof(u16)), 0);
	}
}

void GSDeviceOGL::Draw(const GSHWDrawConfig& config, int offset, int count)
{
	if (config.vs.expand != GSHWDrawConfig::VSExpand::None)
	{
		const bool vs_indexing = config.vs.UseVSExpandIndexBuffer();
		const u32 vs_indexing_expansion = GetExpansionFactor(config.vs.expand);
		DrawIndexedPrimitiveVSExpand(offset, count, vs_indexing, vs_indexing_expansion);
	}
	else
	{
		DrawIndexedPrimitive(offset, count);
	}
}

void GSDeviceOGL::Draw(const GSHWDrawConfig& config)
{
	Draw(config, 0, m_index.count);
}

void GSDeviceOGL::CommitClear(GSTexture* t, bool use_write_fbo)
{
	GSTextureOGL* T = static_cast<GSTextureOGL*>(t);
	if (!T->IsRenderTargetOrDepthStencil() || T->GetState() == GSTexture::State::Dirty)
		return;

	if (use_write_fbo)
	{
		glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_fbo_write);
		glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D,
			t->IsRenderTarget() ? static_cast<GSTextureOGL*>(t)->GetID() : 0, 0);
		glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, m_features.framebuffer_fetch ? GL_DEPTH_ATTACHMENT : GL_DEPTH_STENCIL_ATTACHMENT,
			GL_TEXTURE_2D, t->IsDepthStencil() ? static_cast<GSTextureOGL*>(t)->GetID() : 0, 0);
	}
	else
	{
		OMSetFBO(m_fbo);
		if (T->IsDepthStencil())
		{
			if (GLState::rt && GLState::rt->GetSize() != T->GetSize())
				OMAttachRt(nullptr);
			OMAttachDs(T);
		}
		else
		{
			if (GLState::ds && GLState::ds->GetSize() != T->GetSize())
				OMAttachDs(nullptr);
			OMAttachRt(T);
		}
	}

	if (T->GetState() == GSTexture::State::Invalidated)
	{
		// glInvalidateFramebuffer is core in GL 4.3 and GLES 3.0. The original
		// gate skipped GLES, so on Adreno/Mali every "this content is dead"
		// hint from the texture cache fell through to a no-op — the tile got
		// written back to system memory anyway. On TBDR that's pure waste.
		if (GLAD_GL_VERSION_4_3 || m_is_gles)
		{
			if (T->IsDepthStencil())
			{
				const GLenum attachments[] = {GL_DEPTH_STENCIL_ATTACHMENT};
				glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(attachments), attachments);
			}
			else
			{
				const GLenum attachments[] = {GL_COLOR_ATTACHMENT0};
				glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(attachments), attachments);
			}
		}
	}
	else
	{
		glDisable(GL_SCISSOR_TEST);

		if (T->IsDepthStencil())
		{
			const float d = T->GetClearDepth();
			if (GLState::depth_mask)
			{
				glClearBufferfv(GL_DEPTH, 0, &d);
			}
			else
			{
				glDepthMask(true);
				glClearBufferfv(GL_DEPTH, 0, &d);
				glDepthMask(false);
			}
		}
		else
		{
			const u32 old_color_mask = GLState::wrgba;
			OMSetColorMaskState();

			const GSVector4 c_unorm = T->GetClearForFormat();

			if (T->IsIntegerFormat())
			{
				if (T->IsUnsignedFormat())
					glClearBufferuiv(GL_COLOR, 0, c_unorm.U32);
				else
					glClearBufferiv(GL_COLOR, 0, c_unorm.I32);
			}
			else
			{
				glClearBufferfv(GL_COLOR, 0, c_unorm.v);
			}

			OMSetColorMaskState(OMColorMaskSelector(old_color_mask));
		}

		glEnable(GL_SCISSOR_TEST);
	}

	T->SetState(GSTexture::State::Dirty);

	if (use_write_fbo)
	{
		glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER,
			t->IsRenderTarget() ?
				GL_COLOR_ATTACHMENT0 :
				(m_features.framebuffer_fetch ? GL_DEPTH_ATTACHMENT : GL_DEPTH_STENCIL_ATTACHMENT),
			GL_TEXTURE_2D, 0, 0);
		glBindFramebuffer(GL_DRAW_FRAMEBUFFER, GLState::fbo);
	}
}

std::unique_ptr<GSDownloadTexture> GSDeviceOGL::CreateDownloadTexture(u32 width, u32 height, GSTexture::Format format)
{
	return GSDownloadTextureOGL::Create(width, height, format);
}

GLuint GSDeviceOGL::CreateSampler(PSSamplerSelector sel)
{
	GL_PUSH("Create Sampler");

	GLuint sampler;
	glCreateSamplers(1, &sampler);

	glSamplerParameteri(sampler, GL_TEXTURE_MAG_FILTER, sel.IsMagFilterLinear() ? GL_LINEAR : GL_NEAREST);
	if (!sel.UseMipmapFiltering())
	{
		glSamplerParameteri(sampler, GL_TEXTURE_MIN_FILTER, sel.IsMinFilterLinear() ? GL_LINEAR : GL_NEAREST);
	}
	else
	{
		if (sel.IsMipFilterLinear())
			glSamplerParameteri(sampler, GL_TEXTURE_MIN_FILTER, sel.IsMinFilterLinear() ? GL_LINEAR_MIPMAP_LINEAR : GL_NEAREST_MIPMAP_LINEAR);
		else
			glSamplerParameteri(sampler, GL_TEXTURE_MIN_FILTER, sel.IsMinFilterLinear() ? GL_LINEAR_MIPMAP_NEAREST : GL_NEAREST_MIPMAP_NEAREST);
	}

	glSamplerParameterf(sampler, GL_TEXTURE_MIN_LOD, -1000.0f);
	glSamplerParameterf(sampler, GL_TEXTURE_MAX_LOD, sel.lodclamp ? 0.25f : 1000.0f);

	if (sel.tau)
		glSamplerParameteri(sampler, GL_TEXTURE_WRAP_S, GL_REPEAT);
	else
		glSamplerParameteri(sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
	if (sel.tav)
		glSamplerParameteri(sampler, GL_TEXTURE_WRAP_T, GL_REPEAT);
	else
		glSamplerParameteri(sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);

	glSamplerParameteri(sampler, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);

	return sampler;
}

GLuint GSDeviceOGL::GetSamplerID(PSSamplerSelector ssel)
{
	return m_ps_ss[ssel.key];
}

GSDepthStencilOGL* GSDeviceOGL::CreateDepthStencil(OMDepthStencilSelector dssel)
{
	GSDepthStencilOGL* dss = new GSDepthStencilOGL();

	if (dssel.date)
	{
		dss->EnableStencil();
		if (dssel.date_one)
			dss->SetStencil(GL_EQUAL, GL_ZERO);
		else
			dss->SetStencil(GL_EQUAL, GL_KEEP);
	}

	if (dssel.ztst != ZTST_ALWAYS || dssel.zwe)
	{
		static const GLenum ztst[] =
		{
			GL_NEVER,
			GL_ALWAYS,
			GL_GEQUAL,
			GL_GREATER
		};
		dss->EnableDepth();
		dss->SetDepth(ztst[dssel.ztst], dssel.zwe);
	}

	return dss;
}

GSTexture* GSDeviceOGL::InitPrimDateTexture(GSTexture* rt, const GSVector4i& area, SetDATM datm)
{
	const GSVector2i& rtsize = rt->GetSize();

	GSTexture* tex = CreateRenderTarget(rtsize.x, rtsize.y, GSTexture::Format::PrimID, false);
	if (!tex)
		return nullptr;

	GL_PUSH("PrimID Destination Alpha Clear");
	DoStretchRect(rt, GSVector4(area) / GSVector4(rtsize).xyxy(), tex, GSVector4(area), m_date.primid_ps[static_cast<u8>(datm)], Nearest);
	return tex;
}

std::string GSDeviceOGL::GetShaderSource(const std::string_view entry, GLenum type, const std::string_view glsl_h_code, const std::string_view macro_sel)
{
	std::string src = GenGlslHeader(entry, type, macro_sel);
	src += glsl_h_code;
	return src;
}

std::string GSDeviceOGL::GenGlslHeader(const std::string_view entry, GLenum type, const std::string_view macro)
{
	std::string header;
    if (m_is_gles)
    {
        if (GLAD_GL_ES_VERSION_3_2)
            header = "#version 320 es\n";
        else if (GLAD_GL_ES_VERSION_3_1)
            header = "#version 310 es\n";

        // Vertex↔fragment interface blocks (out/in SHADER {}) are core only in GLSL ES
        // 3.20. On an ES 3.1 context — notably ANGLE, which reports ES 3.1 and validates
        // shaders strictly — they must be enabled explicitly or every TFX shader fails with
        // "'out' : invalid qualifier: shader IO blocks need shader io block extension",
        // leaving no valid programs (black screen, and downstream crashes when a null
        // program is later bound). Native Mali is lenient and compiled these fine without
        // it; ANGLE does not. Harmless where already core.
        if (!GLAD_GL_ES_VERSION_3_2)
        {
            if (GLAD_GL_EXT_shader_io_blocks)
                header += "#extension GL_EXT_shader_io_blocks : enable\n";
            else if (GLAD_GL_OES_shader_io_blocks)
                header += "#extension GL_OES_shader_io_blocks : enable\n";
        }

        if (GLAD_GL_EXT_blend_func_extended)
            header += "#extension GL_EXT_blend_func_extended : require\n";
        if (GLAD_GL_ARB_blend_func_extended)
            header += "#extension GL_ARB_blend_func_extended : require\n";

        if (m_features.framebuffer_fetch)
        {
            if (GLAD_GL_ARM_shader_framebuffer_fetch)
                header += "#extension GL_ARM_shader_framebuffer_fetch : require\n";
            else if (GLAD_GL_EXT_shader_framebuffer_fetch)
                header += "#extension GL_EXT_shader_framebuffer_fetch : require\n";
        }

        // Coherent prior-depth read for SW-Z feedback (gl_LastFragDepthARM).
        if (m_arm_depth_fetch)
            header += "#extension GL_ARM_shader_framebuffer_fetch_depth_stencil : require\n";

        header += "precision highp float;\n";
        header += "precision highp int;\n";
        header += "precision highp sampler2D;\n";
        if (GLAD_GL_ES_VERSION_3_1)
            header += "precision highp sampler2DMS;\n";
        if (GLAD_GL_ES_VERSION_3_2)
            header += "precision highp usamplerBuffer;\n";

        if (!GLAD_GL_EXT_blend_func_extended && !GLAD_GL_ARB_blend_func_extended)
        {
            if (!GLAD_GL_ARM_shader_framebuffer_fetch)
                fprintf(stderr, "Dual source blending is not supported\n");

            header += "#define DISABLE_DUAL_SOURCE\n";
        }
    }
    else {
		// Intel's GL driver doesn't like the readonly qualifier with 3.3 GLSL.
		if (m_features.vs_expand && GLAD_GL_VERSION_4_3)
		{
			header = "#version 430 core\n";
		}
		else
		{
			header = "#version 330 core\n";
			header += "#extension GL_ARB_shading_language_420pack : require\n";
			if (GLAD_GL_ARB_gpu_shader5)
				header += "#extension GL_ARB_gpu_shader5 : require\n";
			if (m_features.vs_expand)
				header += "#extension GL_ARB_shader_storage_buffer_object: require\n";
		}

		if (m_features.framebuffer_fetch && GLAD_GL_EXT_shader_framebuffer_fetch)
			header += "#extension GL_EXT_shader_framebuffer_fetch : require\n";

	}

	if (m_features.framebuffer_fetch)
		header += "#define HAS_FRAMEBUFFER_FETCH 1\n";
	else
		header += "#define HAS_FRAMEBUFFER_FETCH 0\n";

	header += fmt::format("#define HAS_EXT_SHADER_FRAMEBUFFER_FETCH {}\n", GLAD_GL_EXT_shader_framebuffer_fetch ? 1 : 0);
	header += fmt::format("#define HAS_ARM_SHADER_FRAMEBUFFER_FETCH {}\n", GLAD_GL_ARM_shader_framebuffer_fetch ? 1 : 0);
	header += fmt::format("#define HAS_ARM_DEPTH_FETCH {}\n", m_arm_depth_fetch ? 1 : 0);
	header += fmt::format("#define HAS_EXT_SHADER_PIXEL_LOCAL_STORAGE {}\n", GLAD_GL_EXT_shader_pixel_local_storage ? 1 : 0);
	header += fmt::format("#define GPU_PROFILE_MALI {}\n", IsMaliGPUProfile() ? 1 : 0);
	header += fmt::format("#define GPU_PROFILE_ADRENO {}\n", IsAdrenoGPUProfile() ? 1 : 0);
	header += fmt::format("#define GPU_PROFILE_POWERVR {}\n", IsPowerVRGPUProfile() ? 1 : 0);
	header += fmt::format("#define HAS_ARM_DEPTH_FETCH {}\n", m_arm_depth_fetch ? 1 : 0);

	// Shader-compiler workarounds from the driver-bug database (ported from EmuCoreX/sashkinbro
	// with his approval). Each one is off unless a rule matched this exact driver, so the emitted
	// GLSL is byte-identical to before on anything the database does not know about.
	//
	// ScalarizeVectorBitwiseAnd additionally keeps the pre-existing IsMaliGPUProfile() gate: this
	// tree has scalarized vector ANDs on every Mali GL profile since the original fix, including
	// Mali reached through ANGLE or Panfrost where the database resolves a non-ARM driver and would
	// otherwise match nothing. Widening only, never narrowing — nobody loses a fix they had.
	header += fmt::format("#define DRIVER_SCALARIZE_VECTOR_BITWISE_AND {}\n",
		(UsesMobileDriverWorkaround(DriverWorkaround::ScalarizeVectorBitwiseAnd) || IsMaliGPUProfile()) ? 1 : 0);
	header += fmt::format("#define DRIVER_REWRITE_BOOLEAN_NEGATION {}\n",
		UsesMobileDriverWorkaround(DriverWorkaround::RewriteBooleanNegation) ? 1 : 0);
	header += fmt::format("#define DRIVER_STORE_BITWISE_NEGATION_IN_TEMPORARY {}\n",
		UsesMobileDriverWorkaround(DriverWorkaround::StoreBitwiseNegationInTemporary) ? 1 : 0);

	if (GLAD_GL_ARB_conservative_depth)
	{
		header += "#extension GL_ARB_conservative_depth : enable\n";
		header += "#define PS_HAS_CONSERVATIVE_DEPTH 1\n";
	}
	else
	{
		header += "#define PS_HAS_CONSERVATIVE_DEPTH 0\n";
	}

	if (!m_features.texture_barrier && !m_features.multidraw_fb_copy)
	{
		header += "#define DEPTH_FEEDBACK_SUPPORT 0\n"; // None
	}
	else if (m_features.depth_feedback)
	{
		header += "#define DEPTH_FEEDBACK_SUPPORT 1\n"; // Depth
	}
	else
	{
		header += "#define DEPTH_FEEDBACK_SUPPORT 2\n"; // Depth as RT
	}

	// Must match the glClipControl(EXT) enable above: desktop ARB, or GLES with EXT.
	if (GLAD_GL_ARB_clip_control || (m_is_gles && GLAD_GL_EXT_clip_control))
		header += "#define HAS_CLIP_CONTROL 1\n";
	else
		header += "#define HAS_CLIP_CONTROL 0\n";

	// Allow to puts several shader in 1 files
	switch (type)
	{
		case GL_VERTEX_SHADER:
			header += "#define VERTEX_SHADER 1\n";
			break;
		case GL_GEOMETRY_SHADER:
			header += "#define GEOMETRY_SHADER 1\n";
			break;
		case GL_FRAGMENT_SHADER:
			header += "#define FRAGMENT_SHADER 1\n";
			break;
		default:
			pxAssert(0);
	}

	// Don't remove this, the recursive macro breaks some Intel drivers.
	if (entry != "main")
	{
		// Select the entry point ie the main function
		header += "#define ";
		header += entry;
		header += " main\n";
	}

	header += macro;

	// Emitted last, after every #extension directive above: GLSL requires those to precede any
	// non-preprocessor token, and these are real function definitions. The bodies come straight
	// from EmuCoreX so the .glsl call sites stay identical between the two trees.
	header += R"(
bool gpu_boolean_not(bool value)
{
#if DRIVER_REWRITE_BOOLEAN_NEGATION
	return value == false;
#else
	return !value;
#endif
}

uvec2 gpu_bitwise_and(uvec2 a, uvec2 b)
{
#if DRIVER_SCALARIZE_VECTOR_BITWISE_AND
	return uvec2(a.x & b.x, a.y & b.y);
#else
	return a & b;
#endif
}

uvec3 gpu_bitwise_and(uvec3 a, uvec3 b)
{
#if DRIVER_SCALARIZE_VECTOR_BITWISE_AND
	return uvec3(a.x & b.x, a.y & b.y, a.z & b.z);
#else
	return a & b;
#endif
}

uvec4 gpu_bitwise_and(uvec4 a, uvec4 b)
{
#if DRIVER_SCALARIZE_VECTOR_BITWISE_AND
	return uvec4(a.x & b.x, a.y & b.y, a.z & b.z, a.w & b.w);
#else
	return a & b;
#endif
}

ivec3 gpu_bitwise_and(ivec3 a, ivec3 b)
{
#if DRIVER_SCALARIZE_VECTOR_BITWISE_AND
	return ivec3(a.x & b.x, a.y & b.y, a.z & b.z);
#else
	return a & b;
#endif
}

uvec4 gpu_bitwise_not(uvec4 value)
{
#if DRIVER_STORE_BITWISE_NEGATION_IN_TEMPORARY
	uvec4 result = ~value;
	return result;
#else
	return ~value;
#endif
}
)";

	return header;
}

std::string GSDeviceOGL::GetVSSource(VSSelector sel)
{
	DevCon.WriteLn("GL: Compiling new vertex shader with selector 0x%" PRIX64, sel.key);

	std::string macro = fmt::format("#define VS_FST {}\n", static_cast<u32>(sel.fst))
		+ fmt::format("#define VS_IIP {}\n", static_cast<u32>(sel.iip))
		+ fmt::format("#define VS_POINT_SIZE {}\n", static_cast<u32>(sel.point_size))
		+ fmt::format("#define VS_EXPAND {}\n", static_cast<int>(sel.expand));

	std::string src = GenGlslHeader("vs_main", GL_VERTEX_SHADER, macro);
	src += m_shader_tfx_vgs;
	return src;
}

std::string GSDeviceOGL::GetPSSource(const PSSelector& sel)
{
	DevCon.WriteLn("GL: Compiling new pixel shader with selector 0x%016" PRIX64 "_%016" PRIX64, sel.key_hi, sel.key_lo);

	std::string macro = fmt::format("#define PS_FST {}\n", sel.fst)
		+ fmt::format("#define PS_WMS {}\n", sel.wms)
		+ fmt::format("#define PS_WMT {}\n", sel.wmt)
		+ fmt::format("#define PS_ADJS {}\n", sel.adjs)
		+ fmt::format("#define PS_ADJT {}\n", sel.adjt)
		+ fmt::format("#define PS_AEM_FMT {}\n", sel.aem_fmt)
		+ fmt::format("#define PS_PAL_FMT {}\n", sel.pal_fmt)
		+ fmt::format("#define PS_DST_FMT {}\n", sel.dst_fmt)
		+ fmt::format("#define PS_DEPTH_FMT {}\n", sel.depth_fmt)
		+ fmt::format("#define PS_CHANNEL_FETCH {}\n", sel.channel)
		+ fmt::format("#define PS_URBAN_CHAOS_HLE {}\n", sel.urban_chaos_hle)
		+ fmt::format("#define PS_TALES_OF_ABYSS_HLE {}\n", sel.tales_of_abyss_hle)
		+ fmt::format("#define PS_TEX_IS_FB {}\n", sel.tex_is_fb)
		+ fmt::format("#define PS_AEM {}\n", sel.aem)
		+ fmt::format("#define PS_TFX {}\n", sel.tfx)
		+ fmt::format("#define PS_TCC {}\n", sel.tcc)
		+ fmt::format("#define PS_ATST {}\n", static_cast<u32>(sel.atst))
		+ fmt::format("#define PS_AFAIL {}\n", static_cast<u32>(sel.afail))
		+ fmt::format("#define PS_FOG {}\n", sel.fog)
		+ fmt::format("#define PS_BLEND_HW {}\n", sel.blend_hw)
		+ fmt::format("#define PS_A_MASKED {}\n", sel.a_masked)
		+ fmt::format("#define PS_FBA {}\n", sel.fba)
		+ fmt::format("#define PS_LTF {}\n", sel.ltf)
		+ fmt::format("#define PS_AUTOMATIC_LOD {}\n", sel.automatic_lod)
		+ fmt::format("#define PS_MANUAL_LOD {}\n", sel.manual_lod)
		+ fmt::format("#define PS_COLCLIP {}\n", sel.colclip)
		+ fmt::format("#define PS_DATE {}\n", sel.date)
		+ fmt::format("#define PS_TCOFFSETHACK {}\n", sel.tcoffsethack)
		+ fmt::format("#define PS_REGION_RECT {}\n", sel.region_rect)
		+ fmt::format("#define PS_BLEND_A {}\n", sel.blend_a)
		+ fmt::format("#define PS_BLEND_B {}\n", sel.blend_b)
		+ fmt::format("#define PS_BLEND_C {}\n", sel.blend_c)
		+ fmt::format("#define PS_BLEND_D {}\n", sel.blend_d)
		+ fmt::format("#define PS_IIP {}\n", sel.iip)
		+ fmt::format("#define PS_SHUFFLE {}\n", sel.shuffle)
		+ fmt::format("#define PS_SHUFFLE_SAME {}\n", sel.shuffle_same)
		+ fmt::format("#define PS_PROCESS_BA {}\n", sel.process_ba)
		+ fmt::format("#define PS_PROCESS_RG {}\n", sel.process_rg)
		+ fmt::format("#define PS_SHUFFLE_ACROSS {}\n", sel.shuffle_across)
		+ fmt::format("#define PS_READ16_SRC {}\n", sel.real16src)
		+ fmt::format("#define PS_WRITE_RG {}\n", sel.write_rg)
		+ fmt::format("#define PS_FBMASK {}\n", sel.fbmask)
		+ fmt::format("#define PS_COLCLIP_HW {}\n", sel.colclip_hw)
		+ fmt::format("#define PS_RTA_CORRECTION {}\n", sel.rta_correction)
		+ fmt::format("#define PS_RTA_SRC_CORRECTION {}\n", sel.rta_source_correction)
		+ fmt::format("#define PS_DITHER {}\n", sel.dither)
		+ fmt::format("#define PS_DITHER_ADJUST {}\n", sel.dither_adjust)
		+ fmt::format("#define PS_ZCLAMP {}\n", sel.zclamp)
		+ fmt::format("#define PS_ZFLOOR {}\n", sel.zfloor)
		+ fmt::format("#define PS_BLEND_MIX {}\n", sel.blend_mix)
		+ fmt::format("#define PS_ROUND_INV {}\n", sel.round_inv)
		+ fmt::format("#define PS_FIXED_ONE_A {}\n", sel.fixed_one_a)
		+ fmt::format("#define PS_PABE {}\n", sel.pabe)
		+ fmt::format("#define PS_SCANMSK {}\n", sel.scanmsk)
		+ fmt::format("#define PS_NO_COLOR {}\n", sel.no_color)
		+ fmt::format("#define PS_NO_COLOR1 {}\n", sel.no_color1)
		+ fmt::format("#define PS_ZTST {}\n", sel.ztst)
		+ fmt::format("#define PS_AA1 {}\n", static_cast<u32>(sel.aa1))
		+ fmt::format("#define PS_ABE {}\n", sel.abe)
		+ fmt::format("#define PS_ANISOTROPIC_FILTERING {}\n", sel.sw_aniso)
		+ fmt::format("#define PS_ROV_COLOR {}\n", 0)
		+ fmt::format("#define PS_ROV_DEPTH {}\n", 0)
	;

	std::string src = GenGlslHeader("ps_main", GL_FRAGMENT_SHADER, macro);
	src += m_shader_tfx_fs;
	return src;
}

// Copy a sub part of texture (same as below but force a conversion)
void GSDeviceOGL::BlitRect(GSTexture* sTex, const GSVector4i& r, const GSVector2i& dsize, bool at_origin, Filter filter)
{
	g_perfmon.Put(GSPerfMon::TextureCopies, 1);

	CommitClear(sTex, true);

	GL_PUSH(fmt::format("CopyRectConv from {}", static_cast<GSTextureOGL*>(sTex)->GetID()).c_str());

	// NOTE: This previously used glCopyTextureSubImage2D(), but this appears to leak memory in
	// the loading screens of Evolution Snowboarding in Intel/NVIDIA drivers.
	glDisable(GL_SCISSOR_TEST);

	const GSVector4 float_r(r);

	GetConvertProgram(ShaderConvert::COPY).Bind();
	OMSetDepthStencilState(m_convert.dss);
	OMSetBlendState();
	OMSetColorMaskState();
	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(filter == Biln ? m_convert.ln : m_convert.pt);
	DrawStretchRect(float_r / (GSVector4(sTex->GetSize()).xyxy()), float_r, dsize);

	glEnable(GL_SCISSOR_TEST);
}

// Copy a sub part of a texture into another
void GSDeviceOGL::DoCopyRect(GSTexture* sTex, GSTexture* dTex, const GSVector4i& r, u32 destX, u32 destY)
{
	// Empty rect, abort copy.
	if (r.rempty())
	{
		GL_INS("GL: DoCopyRect rect empty.");
		return;
	}

	const GLuint& sid = static_cast<GSTextureOGL*>(sTex)->GetID();
	const GLuint& did = static_cast<GSTextureOGL*>(dTex)->GetID();
	const GSVector4i dst_rect(0, 0, dTex->GetWidth(), dTex->GetHeight());
	const bool full_draw_copy = dst_rect.eq(r);

	// Source is cleared, if destination is a render target, we can carry the clear forward.
	if (sTex->GetState() == GSTexture::State::Cleared)
	{
		if (dTex->IsRenderTargetOrDepthStencil() && ProcessClearsBeforeCopy(sTex, dTex, full_draw_copy))
			return;

		// Commit clear for the source texture.
		CommitClear(sTex, false);
	}

	g_perfmon.Put(GSPerfMon::TextureCopies, 1);
	GL_PUSH("DoCopyRect from %d to %d", sid, did);

	// Commit destination clear if partially overwritten (color only).
	if (dTex->GetState() == GSTexture::State::Cleared && !full_draw_copy)
		CommitClear(dTex, false);

	if (GLAD_GL_VERSION_4_3 || GLAD_GL_ARB_copy_image)
	{
		glCopyImageSubData(sid, GL_TEXTURE_2D, 0, r.x, r.y, 0, did, GL_TEXTURE_2D,
			0, destX, destY, 0, r.width(), r.height(), 1);
	}
	else if (GLAD_GL_EXT_copy_image)
	{
		glCopyImageSubDataEXT(sid, GL_TEXTURE_2D, 0, r.x, r.y, 0, did, GL_TEXTURE_2D,
			0, destX, destY, 0, r.width(), r.height(), 1);
	}
	else if (GLAD_GL_NV_copy_image)
	{
		glCopyImageSubDataNV(sid, GL_TEXTURE_2D, 0, r.x, r.y, 0, did, GL_TEXTURE_2D,
			0, destX, destY, 0, r.width(), r.height(), 1);
	}
	else
	{
		const bool draw_in_depth = sTex->IsDepthStencil();
		const GLenum attachment = draw_in_depth ? GL_DEPTH_STENCIL_ATTACHMENT : GL_COLOR_ATTACHMENT0;

		// Bind attachments.
		glBindFramebuffer(GL_READ_FRAMEBUFFER, m_fbo_read);
		glFramebufferTexture2D(GL_READ_FRAMEBUFFER, attachment, GL_TEXTURE_2D, sid, 0);
		glReadBuffer(draw_in_depth ? GL_NONE : GL_COLOR_ATTACHMENT0);

		// Do copy.
		glCopyTextureSubImage2D(did, 0, destX, destY, r.x, r.y, r.width(), r.height());

		// Unbind attachments.
		if (draw_in_depth)
			glReadBuffer(GL_COLOR_ATTACHMENT0);
		glFramebufferTexture2D(GL_READ_FRAMEBUFFER, attachment, GL_TEXTURE_2D, 0, 0);
		glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
	}

	dTex->SetState(GSTexture::State::Dirty);
}

void GSDeviceOGL::DoStretchRect(GSTexture* sTex, const GSVector4& sRect, GSTexture* dTex, const GSVector4& dRect,
	ShaderConvertSelector shader, Filter filter)
{
	const u8 mask = shader.Mask();
	shader = shader.SetMask(); // Mask is handled separately from program.
	filter = shader.SupportsBilinear() ? Nearest : filter; // Don't allow HW bilinear if SW bilinear is needed.
	DoStretchRect(sTex, sRect, dTex, dRect, GetConvertProgram(shader), false, OMColorMaskSelector(mask), filter);
}

void GSDeviceOGL::DoStretchRect(GSTexture* sTex, const GSVector4& sRect, GSTexture* dTex, const GSVector4& dRect,
	const GLProgram& ps, Filter filter)
{
	DoStretchRect(sTex, sRect, dTex, dRect, ps, false, OMColorMaskSelector(), filter);
}

void GSDeviceOGL::DoStretchRect(GSTexture* sTex, const GSVector4& sRect, GSTexture* dTex, const GSVector4& dRect,
	const GLProgram& ps, bool alpha_blend, OMColorMaskSelector cms, Filter filter)
{
	CommitClear(sTex, true);

	const bool draw_in_depth = dTex->IsDepthStencil();

	// ************************************
	// Init
	// ************************************

	GL_PUSH("StretchRect from %d to %d", static_cast<GSTextureOGL*>(sTex)->GetID(), static_cast<GSTextureOGL*>(dTex)->GetID());
	if (draw_in_depth)
		OMSetRenderTargets(nullptr, nullptr, dTex);
	else
		OMSetRenderTargets(dTex, nullptr, nullptr);

	ps.Bind();

	// ************************************
	// om
	// ************************************

	if (draw_in_depth)
		OMSetDepthStencilState(m_convert.dss_write);
	else
		OMSetDepthStencilState(m_convert.dss);

	OMSetBlendState(alpha_blend, GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_FUNC_ADD);
	OMSetColorMaskState(cms);

	// ************************************
	// Texture
	// ************************************

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(filter == Biln ? m_convert.ln : m_convert.pt);

	// ************************************
	// Draw
	// ************************************
	DrawStretchRect(sRect, dRect, dTex->GetSize());
}

void GSDeviceOGL::PresentRect(GSTexture* sTex, const GSVector4& sRect, GSTexture* dTex, const GSVector4& dRect, PresentShader shader, float shaderTime, Filter filter)
{
	CommitClear(sTex, true);

	const GSVector2i ds(dTex ? dTex->GetSize() : GSVector2i(GetWindowWidth(), GetWindowHeight()));
	DisplayConstantBuffer cb;
	cb.SetSource(sRect, sTex->GetSize());
	cb.SetTarget(dRect, ds);
	cb.SetTime(shaderTime);

	GLProgram& prog = m_present[static_cast<int>(shader)];
	prog.Bind();
	prog.Uniform4fv(0, cb.SourceRect.F32);
	prog.Uniform4fv(1, cb.TargetRect.F32);
	prog.Uniform2fv(2, &cb.SourceSize.x);
	prog.Uniform2fv(3, &cb.TargetSize.x);
	prog.Uniform2fv(4, &cb.TargetResolution.x);
	prog.Uniform2fv(5, &cb.RcpTargetResolution.x);
	prog.Uniform2fv(6, &cb.SourceResolution.x);
	prog.Uniform2fv(7, &cb.RcpSourceResolution.x);
	prog.Uniform1f(8, cb.TimeAndPad.x);

	OMSetDepthStencilState(m_convert.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(filter == Biln ? m_convert.ln : m_convert.pt);

	// Flip y axis only when we render in the backbuffer
	// By default everything is render in the wrong order (ie dx).
	// 1/ consistency between several pass rendering (interlace)
	// 2/ in case some GS code expect thing in dx order.
	// Only flipping the backbuffer is transparent (I hope)...
	const GSVector4 flip_sr(sRect.xwzy());
	DrawStretchRect(flip_sr, dRect, ds);
}

void GSDeviceOGL::DoUpdateCLUTTexture(GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, GSTexture* dTex, u32 dOffset, u32 dSize)
{
	CommitClear(sTex, false);

	const ShaderConvert shader = (dSize == 16) ? ShaderConvert::CLUT_4 : ShaderConvert::CLUT_8;
	GLProgram& prog = GetConvertProgram(shader);
	prog.Bind();
	prog.Uniform3ui(0, offsetX, offsetY, dOffset);
	prog.Uniform1f(1, sScale);

	OMSetDepthStencilState(m_convert.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();
	OMSetRenderTargets(dTex, nullptr, nullptr);

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(m_convert.pt);

	const GSVector4 dRect(0, 0, dSize, 1);
	DrawStretchRect(GSVector4::zero(), dRect, dTex->GetSize());
}

void GSDeviceOGL::DoConvertToIndexedTexture(GSTexture* sTex, float sScale, u32 offsetX, u32 offsetY, u32 SBW, u32 SPSM, GSTexture* dTex, u32 DBW, u32 DPSM)
{
	CommitClear(sTex, false);

	const ShaderConvert shader = ((SPSM & 0xE) == 0) ? ShaderConvert::RGBA_TO_8I : ShaderConvert::RGB5A1_TO_8I;
	GLProgram& prog = GetConvertProgram(shader);
	prog.Bind();
	prog.Uniform1ui(0, SBW);
	prog.Uniform1ui(1, DBW);
	prog.Uniform1ui(2, SPSM);
	prog.Uniform1f(3, sScale);

	OMSetDepthStencilState(m_convert.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();
	OMSetRenderTargets(dTex, nullptr, nullptr);

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(m_convert.pt);

	const GSVector4 dRect(0, 0, dTex->GetWidth(), dTex->GetHeight());
	DrawStretchRect(GSVector4::zero(), dRect, dTex->GetSize());
}

void GSDeviceOGL::DoFilteredDownsampleTexture(GSTexture* sTex, GSTexture* dTex, u32 downsample_factor, const GSVector2i& clamp_min, const GSVector4& dRect)
{
	CommitClear(sTex, false);

	constexpr ShaderConvert shader = ShaderConvert::DOWNSAMPLE_COPY;
	GLProgram& prog = GetConvertProgram(shader);
	prog.Bind();
	prog.Uniform2iv(0, clamp_min.v);
	prog.Uniform1i(1, downsample_factor);
	prog.Uniform1f(2, static_cast<float>(downsample_factor * downsample_factor));
	prog.Uniform1f(3, (GSConfig.UserHacks_NativeScaling > GSNativeScaling::Aggressive) ? 2.0f : 1.0f);

	OMSetDepthStencilState(m_convert.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();
	OMSetRenderTargets(dTex, nullptr, nullptr);

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	PSSetSamplerState(m_convert.pt);

	//const GSVector4 dRect = GSVector4(dTex->GetRect());
	DrawStretchRect(GSVector4::zero(), dRect, dTex->GetSize());
}

void GSDeviceOGL::DrawStretchRect(const GSVector4& sRect, const GSVector4& dRect, const GSVector2i& ds)
{
	g_perfmon.Put(GSPerfMon::TextureCopies, 1);

	const float inv_x = 2.0f / ds.x;
	const float inv_y = 2.0f / ds.y;

	const float left = dRect.x * inv_x - 1.0f;
	const float right = dRect.z * inv_x - 1.0f;
	const float top = -1.0f + dRect.y * inv_y;
	const float bottom = -1.0f + dRect.w * inv_y;

	const GSVertexPT1 vertices[] =
	{
		{GSVector4(left  , top   , 0.0f, 0.0f) , GSVector2(sRect.x , sRect.y)} ,
		{GSVector4(right , top   , 0.0f, 0.0f) , GSVector2(sRect.z , sRect.y)} ,
		{GSVector4(left  , bottom, 0.0f, 0.0f) , GSVector2(sRect.x , sRect.w)} ,
		{GSVector4(right , bottom, 0.0f, 0.0f) , GSVector2(sRect.z , sRect.w)} ,
	};

	IASetVAO(m_vao);
	IASetVertexBuffer(vertices, 4);
	IASetPrimitiveTopology(GL_TRIANGLE_STRIP);
	DrawPrimitive();
}

void GSDeviceOGL::DoDrawMultiStretchRects(
	const MultiStretchRect* rects, u32 num_rects, GSTexture* dTex, ShaderConvertSelector shader)
{
	shader = shader.SetMask(); // Mask is handled separately from program.

	IASetVAO(m_vao);
	IASetPrimitiveTopology(GL_TRIANGLE_STRIP);
	OMSetDepthStencilState(shader.DepthOutput() ? m_convert.dss_write : m_convert.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();
	if (!dTex->IsDepthStencil())
		OMSetRenderTargets(dTex, nullptr, nullptr);
	else
		OMSetRenderTargets(nullptr, nullptr, dTex);
	GetConvertProgram(shader).Bind();

	const GSVector2 ds(static_cast<float>(dTex->GetWidth()), static_cast<float>(dTex->GetHeight()));
	GSTexture* last_tex = rects[0].src;
	Filter last_filter = rects[0].filter;
	u8 last_wmask = rects[0].wmask.wrgba;

	u32 first = 0;
	u32 count = 1;

	for (u32 i = 1; i < num_rects; i++)
	{
		if (rects[i].src == last_tex && rects[i].filter == last_filter && rects[i].wmask.wrgba == last_wmask)
		{
			count++;
			continue;
		}

		DoMultiStretchRects(rects + first, count, ds);
		last_tex = rects[i].src;
		last_filter = rects[i].filter;
		last_wmask = rects[i].wmask.wrgba;
		first += count;
		count = 1;
	}

	DoMultiStretchRects(rects + first, count, ds);
}

void GSDeviceOGL::DoMultiStretchRects(const MultiStretchRect* rects, u32 num_rects, const GSVector2& ds)
{
	g_perfmon.Put(GSPerfMon::TextureCopies, 1);

	const u32 vertex_reserve_size = num_rects * 4 * sizeof(GSVertexPT1);
	const u32 index_reserve_size = num_rects * 6 * sizeof(u16);
	auto vertex_map = m_vertex_stream_buffer->Map(sizeof(GSVertexPT1), vertex_reserve_size);
	auto index_map = m_index_stream_buffer->Map(sizeof(u16), index_reserve_size);
	m_vertex.start = vertex_map.index_aligned;
	m_index.start = index_map.index_aligned;

	// Don't use primitive restart here, it ends up slower on some drivers.
	GSVertexPT1* verts = reinterpret_cast<GSVertexPT1*>(vertex_map.pointer);
	u16* idx = reinterpret_cast<u16*>(index_map.pointer);
	u32 icount = 0;
	u32 vcount = 0;
	for (u32 i = 0; i < num_rects; i++)
	{
		const GSVector4& sRect = rects[i].src_rect;
		const GSVector4& dRect = rects[i].dst_rect;

		const float inv_x = 2.0f / ds.x;
		const float inv_y = 2.0f / ds.y;

		const float left = dRect.x * inv_x - 1.0f;
		const float right = dRect.z * inv_x - 1.0f;
		const float top = -1.0f + dRect.y * inv_y;
		const float bottom = -1.0f + dRect.w * inv_y;

		const u32 vstart = vcount;
		verts[vcount++] = { GSVector4(left  , top   , 0.0f, 0.0f) , GSVector2(sRect.x , sRect.y) };
		verts[vcount++] = { GSVector4(right , top   , 0.0f, 0.0f) , GSVector2(sRect.z , sRect.y) };
		verts[vcount++] = { GSVector4(left  , bottom, 0.0f, 0.0f) , GSVector2(sRect.x , sRect.w) };
		verts[vcount++] = { GSVector4(right , bottom, 0.0f, 0.0f) , GSVector2(sRect.z , sRect.w) };

		if (i > 0)
			idx[icount++] = vstart;

		idx[icount++] = vstart;
		idx[icount++] = vstart + 1;
		idx[icount++] = vstart + 2;
		idx[icount++] = vstart + 3;
		idx[icount++] = vstart + 3;
	};

	m_vertex.count = vcount;
	m_index.count = icount;
	m_vertex_stream_buffer->Unmap(vcount * sizeof(GSVertexPT1));
	m_index_stream_buffer->Unmap(icount * sizeof(u16));

	PSSetShaderResource(TEXTURE_TEXTURE, rects[0].src);
	PSSetSamplerState(rects[0].filter == Biln ? m_convert.ln : m_convert.pt);
	OMSetColorMaskState(rects[0].wmask);
	DrawIndexedPrimitive();
}

void GSDeviceOGL::DoMerge(GSTexture* sTex[3], GSVector4* sRect, GSTexture* dTex, GSVector4* dRect, const GSRegPMODE& PMODE, const GSRegEXTBUF& EXTBUF, u32 c, const Filter filter)
{
	GL_PUSH("DoMerge");

	const GSVector4 full_r(0.0f, 0.0f, 1.0f, 1.0f);
	const bool feedback_write_2 = PMODE.EN2 && sTex[2] != nullptr && EXTBUF.FBIN == 1;
	const bool feedback_write_1 = PMODE.EN1 && sTex[2] != nullptr && EXTBUF.FBIN == 0;
	const bool feedback_write_2_but_blend_bg = feedback_write_2 && PMODE.SLBG == 1;

	// Merge the 2 source textures (sTex[0],sTex[1]). Final results go to dTex. Feedback write will go to sTex[2].
	// If either 2nd output is disabled or SLBG is 1, a background color will be used.
	// Note: background color is also used when outside of the unit rectangle area
	OMSetColorMaskState();
	ClearRenderTarget(dTex, c);

	if (sTex[1] && (PMODE.SLBG == 0 || feedback_write_2_but_blend_bg))
	{
		// 2nd output is enabled and selected. Copy it to destination so we can blend it with 1st output
		// Note: value outside of dRect must contains the background color (c)
		StretchRect(sTex[1], sRect[1], dTex, PMODE.SLBG ? dRect[2] : dRect[1], ShaderConvert::COPY, filter);
	}

	// Upload constant to select YUV algo
	if (feedback_write_2 || feedback_write_1)
	{
		// Write result to feedback loop
		GetConvertProgram(ShaderConvert::YUV).Bind();
		GetConvertProgram(ShaderConvert::YUV).Uniform2i(0, EXTBUF.EMODA, EXTBUF.EMODC);
	}

	// Save 2nd output
	if (feedback_write_2)
		StretchRect(dTex, full_r, sTex[2], dRect[2], ShaderConvert::YUV, filter);

	// Restore background color to process the normal merge
	if (feedback_write_2_but_blend_bg)
		ClearRenderTarget(dTex, c);

	if (sTex[0])
	{
		if (PMODE.AMOD == 1) // Keep the alpha from the 2nd output
			OMSetColorMaskState(OMColorMaskSelector(0x7));

		// 1st output is enabled. It must be blended
		if (PMODE.MMOD == 1)
		{
			// Blend with a constant alpha
			m_merge_obj.ps[1].Bind();
			m_merge_obj.ps[1].Uniform4fv(0, GSVector4::unorm8(c).v);
			DoStretchRect(sTex[0], sRect[0], dTex, dRect[0], m_merge_obj.ps[1], true, OMColorMaskSelector(), filter);
		}
		else
		{
			// Blend with 2 * input alpha
			DoStretchRect(sTex[0], sRect[0], dTex, dRect[0], m_merge_obj.ps[0], true, OMColorMaskSelector(), filter);
		}
	}

	if (feedback_write_1)
		StretchRect(dTex, full_r, sTex[2], dRect[2], ShaderConvert::YUV, filter);
}

void GSDeviceOGL::DoInterlace(GSTexture* sTex, const GSVector4& sRect, GSTexture* dTex, const GSVector4& dRect, ShaderInterlace shader, Filter filter, const InterlaceConstantBuffer& cb)
{
	OMSetColorMaskState();

	m_interlace.ps[static_cast<int>(shader)].Bind();
	m_interlace.ps[static_cast<int>(shader)].Uniform4fv(0, cb.ZrH.F32);

	DoStretchRect(sTex, sRect, dTex, dRect, m_interlace.ps[static_cast<int>(shader)], filter);
}

bool GSDeviceOGL::CompileFXAAProgram()
{
	const std::string_view fxaa_macro = "#define FXAA_GLSL_130 1\n";
	const std::optional<std::string> shader = ReadShaderSource("shaders/common/fxaa.fx");
	if (!shader.has_value())
	{
		Console.Error("GL: Failed to read fxaa.fs");
		return false;
	}

	const std::string ps(GetShaderSource("main", GL_FRAGMENT_SHADER, shader->c_str(), fxaa_macro));
	std::optional<GLProgram> prog = m_shader_cache.GetProgram(m_convert.vs, ps);
	if (!prog.has_value())
	{
		Console.Error("GL: Failed to compile FXAA fragment shader");
		return false;
	}

	m_fxaa.ps = std::move(prog.value());
	return true;
}

void GSDeviceOGL::DoFXAA(GSTexture* sTex, GSTexture* dTex)
{
	if (!m_fxaa.ps.IsValid())
		return;

	GL_PUSH("DoFxaa");

	OMSetColorMaskState();

	const GSVector2i s = dTex->GetSize();

	const GSVector4 sRect(0, 0, 1, 1);
	const GSVector4 dRect(0, 0, s.x, s.y);

	DoStretchRect(sTex, sRect, dTex, dRect, m_fxaa.ps, Biln);
}

#ifdef ARMSX2_HAS_LIBRASHADER

static void ReportShaderChainError(const char* what, libra_error_t err)
{
	char* msg = nullptr;
	if (libra_error_write(err, &msg) == 0 && msg)
	{
		Console.Error("(GS) librashader GL: %s failed: %s", what, msg);
		libra_error_free_string(&msg);
	}
	else
	{
		Console.Error("(GS) librashader GL: %s failed (errno %d)", what, static_cast<int>(libra_error_errno(err)));
	}
	libra_error_free(&err);
}

// libra_gl_loader_t is a bare C function pointer with no userdata argument, so the context
// has to be reachable without one. Only ever one GL device exists at a time (g_gs_device),
// and the chain is created and run on the GS thread, so a file-static set at create time is
// enough — no need to thread the device through.
static GLContext* s_shader_chain_gl_context = nullptr;

static const void* ShaderChainGLLoader(const char* name)
{
	return s_shader_chain_gl_context ? s_shader_chain_gl_context->GetProcAddress(name) : nullptr;
}

#endif

void GSDeviceOGL::DestroyShaderChain()
{
#ifdef ARMSX2_HAS_LIBRASHADER
	if (m_shader_chain)
	{
		libra_gl_filter_chain_t chain = static_cast<libra_gl_filter_chain_t>(m_shader_chain);
		libra_gl_filter_chain_free(&chain);
		m_shader_chain = nullptr;
	}
#endif
	m_shader_chain_preset.clear();
	m_shader_chain_failed = false;
	m_shader_frame_count = 0;
	m_shader_param_generation = 0;
}

void GSDeviceOGL::ApplyShaderChainParams()
{
#ifdef ARMSX2_HAS_LIBRASHADER
	// Per-frame fast path: one atomic load. The lock and the copy only happen on the
	// frames where the user actually moved something.
	const u64 generation = GetShaderChainParamGeneration();
	if (generation == m_shader_param_generation)
		return;

	std::vector<std::pair<std::string, float>> params;
	if (GetShaderChainParams(m_shader_chain_preset, &params))
	{
		libra_gl_filter_chain_t chain = static_cast<libra_gl_filter_chain_t>(m_shader_chain);
		for (const auto& [name, value] : params)
		{
			// A preset can be swapped under a stale override set, so an unknown parameter
			// name is a routine miss, not a fault: report nothing and keep going, since
			// the remaining names are still valid.
			if (libra_error_t err = libra_gl_filter_chain_set_param(&chain, name.c_str(), value))
				libra_error_free(&err);
		}
	}

	// Set even when the store held another preset's values or none at all — otherwise this
	// re-runs the lookup on every frame for as long as the generation stays ahead.
	m_shader_param_generation = generation;
#endif
}

void GSDeviceOGL::RestoreGLStateAfterShaderChain()
{
	// librashader backs up and restores the fixed-function *enables* itself (BLEND,
	// CULL_FACE, DEPTH_TEST, STENCIL_TEST, SCISSOR_TEST, and PRIMITIVE_RESTART /
	// FRAMEBUFFER_SRGB where supported) via an RAII guard, and it never touches the blend
	// func/equation, depth func/mask, stencil func/op, or the scissor box — so all of those
	// stay in sync with GLState on their own. What it does clobber and leave behind is the
	// bound FBO (documented as "GL_FRAMEBUFFER is bound to 0" on return), VAO, program,
	// texture units, sampler bindings, the viewport and the colour mask.
	//
	// GLState's setters are all early-return-if-unchanged, so a stale cache means the state
	// would never be re-pushed and the next draw would silently render wrong. Rather than
	// invalidate the cache (it has no "unknown" sentinel — GLState::Clear() asserts defaults,
	// which is only true on a fresh context), push the cached values back into the driver so
	// reality matches the cache again. Same shape as the end of RenderImGui().
	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, GLState::fbo);
	glBindVertexArray(GLState::vao);
	glViewport(0, 0, GLState::viewport.x, GLState::viewport.y);

	const OMColorMaskSelector cms(GLState::wrgba);
	glColorMaski(0, cms.wr, cms.wg, cms.wb, cms.wa);

	// The chain binds a texture and a sampler per pass, starting at unit 0. Only the units
	// GLState tracks matter; anything above them is never sampled by the TFX/utility draws.
	for (u32 i = 0; i < std::size(GLState::tex_unit); i++)
		glBindTextureUnit(i, GLState::tex_unit[i]);

	// Sampler unit 0 is the tracked one. Unit 1 is special: m_palette_ss is bound there once
	// at device creation and is expected to stay put for the device's lifetime (a sampler
	// object can't be shared across image units, hence the dedicated one), so it has to be
	// put back explicitly or every paletted texture silently samples wrong from here on.
	// GSDeviceOGL never binds a sampler to the remaining units and expects them to fall back
	// to the texture's own parameters, so any the chain left behind have to come back off.
	glBindSampler(0, GLState::ps_ss);
	glBindSampler(1, m_palette_ss);
	for (u32 i = 2; i < std::size(GLState::tex_unit); i++)
		glBindSampler(i, 0);

	// The program cache is a file-static in GLProgram rather than part of GLState, and it
	// has a purpose-built invalidator; the next Bind() re-pushes.
	GLProgram::ResetLastProgram();
}

bool GSDeviceOGL::DoApplyShaderChain(GSTexture* sTex, GSTexture* dTex)
{
#ifndef ARMSX2_HAS_LIBRASHADER
	return false;
#else
	// A preset that fails to compile must not be retried every frame — that would run a
	// full slang compile 60x/sec. Latch the failure until the user picks another preset.
	if (m_shader_chain_failed && m_shader_chain_preset == GSConfig.ShaderChainPreset)
		return false;

	if (!m_shader_chain || m_shader_chain_preset != GSConfig.ShaderChainPreset)
	{
		DestroyShaderChain();
		m_shader_chain_preset = GSConfig.ShaderChainPreset;

		libra_shader_preset_t preset = nullptr;
		if (libra_error_t err = libra_preset_create(m_shader_chain_preset.c_str(), &preset))
		{
			ReportShaderChainError("preset load", err);
			m_shader_chain_failed = true;
			return false;
		}

		filter_chain_gl_opt_t opt = {};
		opt.version = LIBRASHADER_CURRENT_VERSION;
		// 0 means "detect from the context": librashader asks glow for the version and picks
		// GLSL 300/310/320 ES on an ES context. Hardcoding 330 would emit desktop GLSL and
		// break every GLES device.
		opt.glsl_version = 0;
		// DSA is GL 4.5+ and simply does not exist on GLES. This implicitly disables
		// librashader's own shader cache, which is the documented trade-off.
		opt.use_dsa = false;
		opt.force_no_mipmaps = false;
		opt.disable_cache = false;

		// librashader resolves GL through this loader rather than linking it, the same way
		// the Vulkan chain rides the user's ICD.
		s_shader_chain_gl_context = m_gl_context.get();

		// create() invalidates `preset` unconditionally ("the shader preset is
		// immediately invalidated"), so it must NOT be freed afterwards on either path.
		libra_gl_filter_chain_t chain = nullptr;
		if (libra_error_t err = libra_gl_filter_chain_create(&preset, ShaderChainGLLoader, &opt, &chain))
		{
			ReportShaderChainError("chain create", err);
			m_shader_chain_failed = true;
			return false;
		}

		m_shader_chain = chain;
		m_shader_frame_count = 0;
		// The new chain sits at the preset's initial values, so whatever we last pushed is
		// gone with the old one — force ApplyShaderChainParams to feed it again.
		m_shader_param_generation = 0;
		Console.WriteLn("(GS) librashader GL: loaded preset '%s'", m_shader_chain_preset.c_str());
	}

	// GS thread, chain alive, before the frame call — the only place a set_param is safe.
	// Touches no GL state of its own (it writes librashader's own parameter map), so it is
	// fine ahead of the GL_PUSH and needs no state restore.
	ApplyShaderChainParams();

	GL_PUSH("ApplyShaderChain");

	GSTextureOGL* const src = static_cast<GSTextureOGL*>(sTex);
	GSTextureOGL* const dst = static_cast<GSTextureOGL*>(dTex);

	// librashader wants the sized internal format the texture was allocated with — that's
	// m_gl_format (GL_RGBA8), the one handed to glTextureStorage2D, NOT m_int_format, which
	// despite the name is the pixel-transfer format (GL_RGBA).
	const libra_image_gl_t in = {src->GetID(), src->GetGLFormat(),
		static_cast<uint32_t>(src->GetWidth()), static_cast<uint32_t>(src->GetHeight())};
	const libra_image_gl_t out = {dst->GetID(), dst->GetGLFormat(),
		static_cast<uint32_t>(dst->GetWidth()), static_cast<uint32_t>(dst->GetHeight())};
	const libra_viewport_t vp = {0.0f, 0.0f,
		static_cast<uint32_t>(dst->GetWidth()), static_cast<uint32_t>(dst->GetHeight())};

	// Every librashader entry point takes the chain handle by address, not by value.
	// Unlike Vulkan there's no command buffer — the chain issues its draws immediately.
	libra_gl_filter_chain_t chain = static_cast<libra_gl_filter_chain_t>(m_shader_chain);
	const libra_error_t err = libra_gl_filter_chain_frame(&chain, m_shader_frame_count, in, out, &vp, nullptr, nullptr);

	// Unconditional: the chain can fail part-way through, having already clobbered state.
	RestoreGLStateAfterShaderChain();

	if (err)
	{
		ReportShaderChainError("frame", err);
		m_shader_chain_failed = true;
		return false;
	}
	m_shader_frame_count++;

	dst->SetState(GSTexture::State::Dirty);
	return true;
#endif
}

bool GSDeviceOGL::CompileShadeBoostProgram()
{
	const std::optional<std::string> shader = ReadShaderSource("shaders/opengl/shadeboost.glsl");
	if (!shader.has_value())
	{
		Host::ReportErrorAsync("GS", "Failed to read shaders/opengl/shadeboost.glsl.");
		return false;
	}

	const std::string ps(GetShaderSource("ps_main", GL_FRAGMENT_SHADER, *shader));
	if (!m_shader_cache.GetProgram(&m_shadeboost.ps, m_convert.vs, ps))
		return false;
	m_shadeboost.ps.RegisterUniform("params");
	m_shadeboost.ps.SetName("Shadeboost pipe");
	return true;
}

void GSDeviceOGL::DoShadeBoost(GSTexture* sTex, GSTexture* dTex, const float params[4])
{
	GL_PUSH("DoShadeBoost");

	m_shadeboost.ps.Bind();
	m_shadeboost.ps.Uniform4fv(0, params);

	OMSetColorMaskState();

	const GSVector2i s = dTex->GetSize();

	const GSVector4 sRect(0, 0, 1, 1);
	const GSVector4 dRect(0, 0, s.x, s.y);

	DoStretchRect(sTex, sRect, dTex, dRect, m_shadeboost.ps, Nearest);
}

void GSDeviceOGL::SetupDATE(GSTexture* rt, GSTexture* ds, SetDATM datm, const GSVector4i& bbox)
{
	g_perfmon.Put(GSPerfMon::TextureCopies, 1);

	GL_PUSH("DATE First Pass");

	// sfex3 (after the capcom logo), vf4 (first menu fading in), ffxii shadows, rumble roses shadows, persona4 shadows

	OMSetRenderTargets(nullptr, nullptr, ds, &GLState::scissor);
	{
		constexpr GLint clear_color = 0;
		glClearBufferiv(GL_STENCIL, 0, &clear_color);
	}
	GetConvertProgram(SetDATMShader(datm)).Bind();

	// om

	OMSetDepthStencilState(m_date.dss);
	OMSetBlendState(false);
	OMSetColorMaskState();

	// ia

	const GSVector4 src = GSVector4(bbox) / GSVector4(ds->GetSize()).xyxy();
	const GSVector4 dst = src * 2.f - 1.f;

	const GSVertexPT1 vertices[] =
	{
		{GSVector4(dst.x, dst.y, 0.0f, 0.0f), GSVector2(src.x, src.y)},
		{GSVector4(dst.z, dst.y, 0.0f, 0.0f), GSVector2(src.z, src.y)},
		{GSVector4(dst.x, dst.w, 0.0f, 0.0f), GSVector2(src.x, src.w)},
		{GSVector4(dst.z, dst.w, 0.0f, 0.0f), GSVector2(src.z, src.w)},
	};

	IASetVAO(m_vao);
	IASetVertexBuffer(vertices, 4);
	IASetPrimitiveTopology(GL_TRIANGLE_STRIP);

	// Texture

	PSSetShaderResource(TEXTURE_TEXTURE, rt);
	PSSetSamplerState(m_convert.pt);

	DrawPrimitive();
}

__fi static void WriteToStreamBuffer(GLStreamBuffer* sb, u32 index, u32 align, const void* data, u32 size)
{
	const auto res = sb->Map(align, size);
	std::memcpy(res.pointer, data, size);
	sb->Unmap(size);

	glBindBufferRange(GL_UNIFORM_BUFFER, index, sb->GetGLBufferId(), res.buffer_offset, size);
}

void GSDeviceOGL::VSSetUniformBuffer(GSHWDrawConfig::VSConstantBuffer& cb)
{
	WriteToStreamBuffer(m_vertex_uniform_stream_buffer.get(), g_vs_cb_index,
		m_uniform_buffer_alignment, &cb, sizeof(cb));
}

void GSDeviceOGL::PSSetUniformBuffer(GSHWDrawConfig::PSConstantBuffer& cb)
{
	WriteToStreamBuffer(m_fragment_uniform_stream_buffer.get(), g_ps_cb_index,
		m_uniform_buffer_alignment, &cb, sizeof(cb));
}

void GSDeviceOGL::VSSetPushConstants(u32 base_vertex, u32 base_index, bool force_update)
{
	GSHWDrawConfig::VSPushConstants vs_pc;
	vs_pc.base_vertex = base_vertex;
	vs_pc.base_index = base_index;

	if (m_vs_pc_cache.Update(vs_pc) || force_update)
	{
		WriteToStreamBuffer(m_vertex_push_constants_stream_buffer.get(), g_vs_pc_index,
			m_uniform_buffer_alignment, &vs_pc, sizeof(vs_pc));
	}
}

void GSDeviceOGL::IASetVAO(GLuint vao)
{
	if (GLState::vao == vao)
		return;

	GLState::vao = vao;
	glBindVertexArray(vao);
}

void GSDeviceOGL::IASetVertexBuffer(const void* vertices, size_t count, size_t align_multiplier)
{
	const u32 size = static_cast<u32>(count) * sizeof(GSVertexPT1);
	auto res = m_vertex_stream_buffer->Map(sizeof(GSVertexPT1) * align_multiplier, size);
	std::memcpy(res.pointer, vertices, size);
	m_vertex.start = res.index_aligned * align_multiplier;
	m_vertex.count = count;
	m_vertex_stream_buffer->Unmap(size);
}

void GSDeviceOGL::SetIndexBuffer(std::unique_ptr<GLStreamBuffer>& buffer, const void* index, size_t count)
{
	const u32 size = static_cast<u32>(count) * sizeof(u16);
	auto res = buffer->Map(sizeof(u16), size);
	m_index.start = res.index_aligned;
	m_index.count = count;
	std::memcpy(res.pointer, index, size);
	buffer->Unmap(size);
}

void GSDeviceOGL::IASetIndexBuffer(const void* index, size_t count)
{
	SetIndexBuffer(m_index_stream_buffer, index, count);
}

void GSDeviceOGL::VSSetIndexBuffer(const void* index, size_t count)
{
	SetIndexBuffer(m_expand_index_stream_buffer, index, count);
}

void GSDeviceOGL::IASetPrimitiveTopology(GLenum topology)
{
	m_draw_topology = topology;
}

void GSDeviceOGL::PSSetShaderResource(int i, GSTexture* sr)
{
	pxAssert(i < static_cast<int>(std::size(GLState::tex_unit)));

	const GLuint id = sr ?  static_cast<GSTextureOGL*>(sr)->GetID() : 0;
	if (GLState::tex_unit[i] != id)
	{
		GLState::tex_unit[i] = id;
		glBindTextureUnit(i, id);
	}
}

void GSDeviceOGL::PSSetSamplerState(GLuint ss)
{
	if (GLState::ps_ss != ss)
	{
		GLState::ps_ss = ss;
		glBindSampler(0, ss);
	}
}

void GSDeviceOGL::ClearSamplerCache()
{
	glDeleteSamplers(std::size(m_ps_ss), m_ps_ss);

	for (u32 key = 0; key < std::size(m_ps_ss); key++)
	{
		m_ps_ss[key] = CreateSampler(PSSamplerSelector(key));
	}
}

bool GSDeviceOGL::CreateCASPrograms()
{
	std::optional<std::string> cas_source = ReadShaderSource("shaders/opengl/cas.glsl");
	if (!cas_source.has_value() || !GetCASShaderSource(&cas_source.value()))
	{
		m_features.cas_sharpening = false;
		return false;
	}

	const char* header =
		"#version 420\n"
		"#extension GL_ARB_compute_shader : require\n";
	const char* sharpen_params[2] = {
		"#define CAS_SHARPEN_ONLY false\n",
		"#define CAS_SHARPEN_ONLY true\n"};

	if (!m_shader_cache.GetComputeProgram(&m_cas.upscale_ps, fmt::format("{}{}{}", header, sharpen_params[0], cas_source.value())) ||
		!m_shader_cache.GetComputeProgram(&m_cas.sharpen_ps, fmt::format("{}{}{}", header, sharpen_params[1], cas_source.value())))
	{
		m_features.cas_sharpening = false;
		return false;
	}

	const auto link_uniforms = [](GLProgram& prog) {
		prog.RegisterUniform("const0");
		prog.RegisterUniform("const1");
		prog.RegisterUniform("srcOffset");
	};
	link_uniforms(m_cas.upscale_ps);
	link_uniforms(m_cas.sharpen_ps);

	return true;
}

bool GSDeviceOGL::DoCAS(GSTexture* sTex, GSTexture* dTex, bool sharpen_only, const std::array<u32, NUM_CAS_CONSTANTS>& constants)
{
	g_perfmon.Put(GSPerfMon::TextureCopies, 1);

	const GLProgram& prog = sharpen_only ? m_cas.sharpen_ps : m_cas.upscale_ps;
	prog.Bind();
	prog.Uniform4uiv(0, &constants[0]);
	prog.Uniform4uiv(1, &constants[4]);
	prog.Uniform2iv(2, reinterpret_cast<const s32*>(&constants[8]));

	PSSetShaderResource(TEXTURE_TEXTURE, sTex);
	glBindImageTexture(0, static_cast<GSTextureOGL*>(dTex)->GetID(), 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RGBA8);

	static const int threadGroupWorkRegionDim = 16;
	const int dispatchX = (dTex->GetWidth() + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim;
	const int dispatchY = (dTex->GetHeight() + (threadGroupWorkRegionDim - 1)) / threadGroupWorkRegionDim;
	glDispatchCompute(dispatchX, dispatchY, 1);

	// dTex is written through an image binding, but the caller turns straight around and
	// samples it for the present blit. Image stores are incoherent without an explicit
	// barrier, so the fetch is otherwise free to observe the pre-dispatch contents.
	glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT | GL_TEXTURE_FETCH_BARRIER_BIT);

	return true;
}

bool GSDeviceOGL::CreateImGuiProgram()
{
	const std::optional<std::string> glsl = ReadShaderSource("shaders/opengl/imgui.glsl");
	if (!glsl.has_value())
	{
		Console.Error("GL: Failed to read imgui.glsl");
		return false;
	}

	std::optional<GLProgram> prog = m_shader_cache.GetProgram(
		GetShaderSource("vs_main", GL_VERTEX_SHADER, glsl.value()),
		GetShaderSource("ps_main", GL_FRAGMENT_SHADER, glsl.value()));
	if (!prog.has_value())
	{
		Console.Error("GL: Failed to compile imgui shaders");
		return false;
	}

	prog->SetName("ImGui Render");
	prog->RegisterUniform("ProjMtx");
	m_imgui.ps = std::move(prog.value());

	// Need a different VAO because the layout doesn't match GS
	glGenVertexArrays(1, &m_imgui.vao);
	glBindVertexArray(m_imgui.vao);
	m_vertex_stream_buffer->Bind();
	m_index_stream_buffer->Bind();

	glEnableVertexAttribArray(0);
	glEnableVertexAttribArray(1);
	glEnableVertexAttribArray(2);
	glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(ImDrawVert), (GLvoid*)offsetof(ImDrawVert, pos));
	glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, sizeof(ImDrawVert), (GLvoid*)offsetof(ImDrawVert, uv));
	glVertexAttribPointer(2, 4, GL_UNSIGNED_BYTE, GL_TRUE, sizeof(ImDrawVert), (GLvoid*)offsetof(ImDrawVert, col));

	glBindVertexArray(GLState::vao);
	return true;
}

void GSDeviceOGL::RenderImGui()
{
	ImGui::Render();
	const ImDrawData* draw_data = ImGui::GetDrawData();
	if (draw_data->CmdListsCount == 0)
		return;

	UpdateImGuiTextures();

	constexpr float L = 0.0f;
	const float R = static_cast<float>(m_window_info.surface_width);
	constexpr float T = 0.0f;
	const float B = static_cast<float>(m_window_info.surface_height);

	// clang-format off
	const float ortho_projection[4][4] =
	{
		{ 2.0f/(R-L),   0.0f,         0.0f,   0.0f },
		{ 0.0f,         2.0f/(T-B),   0.0f,   0.0f },
		{ 0.0f,         0.0f,        -1.0f,   0.0f },
		{ (R+L)/(L-R),  (T+B)/(B-T),  0.0f,   1.0f },
	};
	// clang-format on

	m_imgui.ps.Bind();
	m_imgui.ps.UniformMatrix4fv(0, &ortho_projection[0][0]);
	IASetVAO(m_imgui.vao);
	OMSetBlendState(true, GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_FUNC_ADD);
	OMSetDepthStencilState(m_convert.dss);
	PSSetSamplerState(m_convert.ln);

	// Need to flip the scissor due to lower-left on the window framebuffer
	GSVector4i last_scissor = GSVector4i::xffffffff();

	// Render command lists
	for (int n = 0; n < draw_data->CmdListsCount; n++)
	{
		const ImDrawList* cmd_list = draw_data->CmdLists[n];

		// Different vertex format.
		u32 vertex_start;
		{
			const u32 size = static_cast<u32>(cmd_list->VtxBuffer.Size) * sizeof(ImDrawVert);
			auto res = m_vertex_stream_buffer->Map(sizeof(ImDrawVert), size);
			std::memcpy(res.pointer, cmd_list->VtxBuffer.Data, size);
			vertex_start = res.index_aligned;
			m_vertex_stream_buffer->Unmap(size);
		}

		IASetIndexBuffer(cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.Size);

		for (int cmd_i = 0; cmd_i < cmd_list->CmdBuffer.Size; cmd_i++)
		{
			const ImDrawCmd* pcmd = &cmd_list->CmdBuffer[cmd_i];
			pxAssert(!pcmd->UserCallback);

			const GSVector4 clip = GSVector4::load<false>(&pcmd->ClipRect);
			if ((clip.zwzw() <= clip.xyxy()).mask() != 0)
				continue;

			// Apply scissor/clipping rectangle (Y is inverted in OpenGL)
			const GSVector4i iclip = GSVector4i(clip);
			if (!last_scissor.eq(iclip))
			{
				glScissor(iclip.x, m_window_info.surface_height - iclip.w, iclip.width(), iclip.height());
				last_scissor = iclip;
			}

			// Since we don't have the GSTexture...
			const GLuint texture_id = static_cast<GLuint>(pcmd->GetTexID());
			if (GLState::tex_unit[0] != texture_id)
			{
				GLState::tex_unit[0] = texture_id;
				glBindTextureUnit(0, texture_id);
			}

			glDrawElementsBaseVertex(GL_TRIANGLES, (GLsizei)pcmd->ElemCount, GL_UNSIGNED_SHORT,
				(void*)(intptr_t)((pcmd->IdxOffset + m_index.start) * sizeof(ImDrawIdx)), pcmd->VtxOffset + vertex_start);
		}

		g_perfmon.Put(GSPerfMon::DrawCalls, cmd_list->CmdBuffer.Size);
	}

	IASetVAO(m_vao);
	glScissor(GLState::scissor.x, GLState::scissor.y, GLState::scissor.width(), GLState::scissor.height());
}

void GSDeviceOGL::RenderBlankFrame()
{
	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
	glDisable(GL_SCISSOR_TEST);
	if (m_is_gles) // GLES/TBDR-only tile-bandwidth hint; inert on desktop, gated to keep it canonical
	{
		const GLenum pre[] = {GL_COLOR, GL_DEPTH, GL_STENCIL};
		glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(pre), pre);
	}
	glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
	glClear(GL_COLOR_BUFFER_BIT);
	if (GLAD_GL_VERSION_4_3 || m_is_gles)
	{
		const GLenum post[] = {GL_DEPTH, GL_STENCIL};
		glInvalidateFramebuffer(GL_DRAW_FRAMEBUFFER, std::size(post), post);
	}
	m_gl_context->SwapBuffers();
	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, GLState::fbo);
	glEnable(GL_SCISSOR_TEST);
}

void GSDeviceOGL::OMAttachRt(GSTexture* rt)
{
	if (GLState::rt == rt)
		return;

	GLState::rt = static_cast<GSTextureOGL*>(rt);
	glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D,
		rt ? static_cast<GSTextureOGL*>(rt)->GetID() : 0, 0);
}

void GSDeviceOGL::OMAttachDsAsRt(GSTexture* ds_as_rt)
{
	if (GLState::ds_as_rt == ds_as_rt)
		return;

	GLState::ds_as_rt = static_cast<GSTextureOGL*>(ds_as_rt);
	glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, GL_TEXTURE_2D,
		ds_as_rt ? static_cast<GSTextureOGL*>(ds_as_rt)->GetID() : 0, 0);
}

void GSDeviceOGL::OMAttachDs(GSTexture* ds)
{
	if (GLState::ds == ds)
		return;

	GLState::ds = static_cast<GSTextureOGL*>(ds);

	const GLenum target = m_features.framebuffer_fetch ? GL_DEPTH_ATTACHMENT : GL_DEPTH_STENCIL_ATTACHMENT;
	glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, target, GL_TEXTURE_2D, ds ? static_cast<GSTextureOGL*>(ds)->GetID() : 0, 0);
}

void GSDeviceOGL::OMSetFBO(GLuint fbo)
{
	if (GLState::fbo != fbo)
	{
		GLState::fbo = fbo;
		glBindFramebuffer(GL_DRAW_FRAMEBUFFER, fbo);
	}
}

void GSDeviceOGL::OMSetDepthStencilState(GSDepthStencilOGL* dss)
{
	dss->SetupDepth();
	dss->SetupStencil();
}

void GSDeviceOGL::OMSetColorMaskState(OMColorMaskSelector sel)
{
	if (sel.wrgba != GLState::wrgba)
	{
		GLState::wrgba = sel.wrgba;

		glColorMaski(0, sel.wr, sel.wg, sel.wb, sel.wa);
	}
}

void GSDeviceOGL::OMUnbindTexture(GSTextureOGL* tex)
{
	if (GLState::rt != tex && GLState::ds_as_rt != tex && GLState::ds != tex)
		return;

	OMSetFBO(m_fbo);
	if (GLState::rt == tex)
		OMAttachRt();
	if (GLState::ds_as_rt == tex)
		OMAttachDsAsRt();
	if (GLState::ds == tex)
		OMAttachDs();
}

void GSDeviceOGL::OMSetBlendState(bool enable, GLenum src_factor, GLenum dst_factor, GLenum op,
	GLenum src_factor_alpha, GLenum dst_factor_alpha, bool is_constant, u8 constant)
{
	if (enable)
	{
		if (!GLState::blend)
		{
			GLState::blend = true;
			glEnable(GL_BLEND);
		}

		if (is_constant && GLState::bf != constant)
		{
			GLState::bf = constant;
			const float bf = static_cast<float>(constant) / 128.0f;
			glBlendColor(bf, bf, bf, bf);
		}

		if (GLState::eq_RGB != op)
		{
			GLState::eq_RGB = op;
			glBlendEquationSeparate(op, GL_FUNC_ADD);
		}

		if (GLState::f_sRGB != src_factor || GLState::f_dRGB != dst_factor ||
			GLState::f_sA != src_factor_alpha || GLState::f_dA != dst_factor_alpha)
		{
			GLState::f_sRGB = src_factor;
			GLState::f_dRGB = dst_factor;
			GLState::f_sA = src_factor_alpha;
			GLState::f_dA = dst_factor_alpha;
			glBlendFuncSeparate(src_factor, dst_factor, src_factor_alpha, dst_factor_alpha);
		}
	}
	else
	{
		if (GLState::blend)
		{
			GLState::blend = false;
			glDisable(GL_BLEND);
		}
	}
}

void GSDeviceOGL::OMSetRenderTargets(GSTexture* rt, GSTexture* ds_as_rt, GSTexture* ds, const GSVector4i* scissor)
{
	const bool rt_changed = (rt != GLState::rt);
	const bool ds_as_rt_changed = (ds_as_rt != GLState::ds_as_rt);
	const bool ds_changed = (ds != GLState::ds);
	const u32 draw_buffers = GLState::draw_buffers;

	g_perfmon.Put(GSPerfMon::RenderPasses, static_cast<double>(rt_changed || ds_as_rt_changed || ds_changed));
	// Split up to avoid unbind/bind calls when clearing.

	OMSetFBO(m_fbo);

	GLState::rt_written = false;
	GLState::ds_as_rt_written = false;
	GLState::ds_written = false;

	if (rt)
	{
		OMAttachRt(rt);
		CommitClear(rt, false);
		GLState::rt_written = rt_changed;
	}
	else
		OMAttachRt();

	if (ds_as_rt)
	{
		OMAttachDsAsRt(ds_as_rt);
		CommitClear(ds_as_rt, false);
		GLState::ds_as_rt_written = ds_as_rt_changed;
	}
	else
		OMAttachDsAsRt();

	if (ds)
	{
		OMAttachDs(ds);
		CommitClear(ds, false);
		GLState::ds_written = ds_changed;
	}
	else
		OMAttachDs();

	if (rt || ds_as_rt || ds)
	{
		const GSVector2i size = rt ? rt->GetSize() : (ds_as_rt ? ds_as_rt->GetSize() : ds->GetSize());
		SetViewport(size);
		SetScissor(scissor ? *scissor : GSVector4i::loadh(size));
	}

	if (draw_buffers != GLState::UpdateDrawBuffers())
	{
		// Always write to the first and second buffer
		static constexpr GLenum target[2] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
		glDrawBuffers(GLState::draw_buffers, target);
	}
}

void GSDeviceOGL::SetViewport(const GSVector2i& viewport)
{
	if (GLState::viewport != viewport)
	{
		GLState::viewport = viewport;
		glViewport(0, 0, viewport.x, viewport.y);
	}
}

void GSDeviceOGL::SetScissor(const GSVector4i& scissor)
{
	if (!GLState::scissor.eq(scissor))
	{
		GLState::scissor = scissor;
		glScissor(scissor.x, scissor.y, scissor.width(), scissor.height());
	}
}

void GSDeviceOGL::SetupPipeline(const ProgramSelector& psel)
{
	auto it = m_programs.find(psel);
	if (it != m_programs.end())
	{
		it->second.Bind();
		return;
	}

	const std::string vs(GetVSSource(psel.vs));
	const std::string ps(GetPSSource(psel.ps));

	GLProgram prog;
	m_shader_cache.GetProgram(&prog, vs, ps);
	it = m_programs.emplace(psel, std::move(prog)).first;
	it->second.Bind();
}

void GSDeviceOGL::SetupSampler(PSSamplerSelector ssel)
{
	PSSetSamplerState(m_ps_ss[ssel.key]);
}

GLuint GSDeviceOGL::GetPaletteSamplerID()
{
	return m_palette_ss;
}

void GSDeviceOGL::SetupOM(OMDepthStencilSelector dssel)
{
	OMSetDepthStencilState(m_om_dss[dssel.key]);
}

// clang-format off
static constexpr std::array<GLenum, 16> s_gl_blend_factors = { {
	GL_SRC_COLOR, GL_ONE_MINUS_SRC_COLOR, GL_DST_COLOR, GL_ONE_MINUS_DST_COLOR,
	GL_SRC1_COLOR, GL_ONE_MINUS_SRC1_COLOR, GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA,
	GL_DST_ALPHA, GL_ONE_MINUS_DST_ALPHA, GL_SRC1_ALPHA, GL_ONE_MINUS_SRC1_ALPHA,
	GL_CONSTANT_COLOR, GL_ONE_MINUS_CONSTANT_COLOR, GL_ONE, GL_ZERO
} };
static constexpr std::array<GLenum, 3> s_gl_blend_ops = { {
		GL_FUNC_ADD, GL_FUNC_SUBTRACT, GL_FUNC_REVERSE_SUBTRACT
} };
// clang-format on

void GSDeviceOGL::DoRenderHW(GSHWDrawConfig& config)
{
	if (!GLState::scissor.eq(config.scissor))
	{
		glScissor(config.scissor.x, config.scissor.y, config.scissor.width(), config.scissor.height());
		GLState::scissor = config.scissor;
	}

	if (config.tex && (m_features.texture_barrier || (config.tex != config.rt)))
		CommitClear(config.tex, true);
	if (config.pal)
		CommitClear(config.pal, true);

	const GSVector2i rtsize = (config.rt ? config.rt : config.ds)->GetSize();
	GSTexture* colclip_rt = g_gs_device->GetColorClipTexture();
	GSTexture* draw_rt = config.rt;
	GSTexture* draw_ds = config.ds;
	GSTexture* draw_ds_as_rt = m_ds_as_rt;
	GSTexture* draw_rt_clone = nullptr;
	GSTexture* draw_ds_clone = nullptr;
	GSTexture* primid_texture = nullptr;

	ScopedGuard recycle_temp_textures([&]() {
		if (draw_rt_clone)
			Recycle(draw_rt_clone);
		if (draw_ds_clone)
			Recycle(draw_ds_clone);
		if (primid_texture)
			Recycle(primid_texture);
	});

	if (colclip_rt)
	{
		if (config.colclip_mode == GSHWDrawConfig::ColClipMode::EarlyResolve)
		{
			const GSVector2i size = config.rt->GetSize();
			const GSVector4 dRect(config.colclip_update_area);
			const GSVector4 sRect = dRect / GSVector4(size.x, size.y).xyxy();
			StretchRect(colclip_rt, sRect, config.rt, dRect, ShaderConvert::COLCLIP_RESOLVE, Nearest);
			Recycle(colclip_rt);

			g_gs_device->SetColorClipTexture(nullptr);

			colclip_rt = nullptr;
		}
		else
		{
			config.ps.colclip_hw = 1;
		}
	}

	if (config.ps.colclip_hw)
	{
		if (!colclip_rt)
		{
			config.colclip_update_area = config.drawarea;

			colclip_rt = CreateFeedbackTarget(rtsize.x, rtsize.y, GSTexture::Format::ColorClip, false);

			if (!colclip_rt)
			{
				Console.Warning("GL: Failed to allocate ColorClip render target, aborting draw.");

				return;
			}

			OMSetRenderTargets(colclip_rt, nullptr, config.ds, nullptr);

			g_gs_device->SetColorClipTexture(colclip_rt);

			const GSVector4 dRect = GSVector4((config.colclip_mode == GSHWDrawConfig::ColClipMode::ConvertOnly) ? GSVector4i::loadh(rtsize) : config.drawarea);
			const GSVector4 sRect = dRect / GSVector4(rtsize.x, rtsize.y).xyxy();
			StretchRect(config.rt, sRect, colclip_rt, dRect, ShaderConvert::COLCLIP_INIT, Nearest);
		}

		draw_rt = colclip_rt ? colclip_rt : config.rt;
	}

	// Destination Alpha Setup
	const bool need_barrier = config.require_one_barrier || (config.require_full_barrier && m_features.feedback_loops());
	switch (config.destination_alpha)
	{
		case GSHWDrawConfig::DestinationAlphaMode::Off:
		case GSHWDrawConfig::DestinationAlphaMode::Full:
			break; // No setup
		case GSHWDrawConfig::DestinationAlphaMode::PrimIDTracking:
			primid_texture = InitPrimDateTexture(colclip_rt ? colclip_rt : config.rt, config.drawarea, config.datm);
			if (!primid_texture)
			{
				Console.Warning("GL: Failed to allocate DATE image, aborting draw.");
				return;
			}
			break;
		case GSHWDrawConfig::DestinationAlphaMode::StencilOne:
			if (need_barrier)
			{
				// Cleared after RT bind.
				break;
			}
			[[fallthrough]];
		case GSHWDrawConfig::DestinationAlphaMode::Stencil:
			SetupDATE(colclip_rt ? colclip_rt : config.rt, config.ds, config.datm, config.drawarea);
			break;
	}

	IASetVertexBuffer(config.verts, config.nverts, GetVertexAlignment(config.vs.expand));

	if (config.vs.UseFixedExpandIndexBuffer())
	{
		IASetVAO(m_expand_vao);
		m_index.start = 0;
		m_index.count = config.nindices;
	}
	else if (config.vs.UseVSExpandIndexBuffer())
	{
		IASetVAO(m_dummy_vao); // Unbind vertex buffer from IA to prevent unwanted fetches.
		VSSetIndexBuffer(config.indices, config.nindices);
	}
	else
	{
		IASetVAO(m_vao);
		IASetIndexBuffer(config.indices, config.nindices);
	}

	GLenum topology = 0;
	switch (config.topology)
	{
		case GSHWDrawConfig::Topology::Point:    topology = GL_POINTS;    break;
		case GSHWDrawConfig::Topology::Line:     topology = GL_LINES;     break;
		case GSHWDrawConfig::Topology::Triangle: topology = GL_TRIANGLES; break;
	}
	IASetPrimitiveTopology(topology);

	if (config.tex && (m_features.texture_barrier || config.tex_hazard == GSHWDrawConfig::TEX_HAZARD_NONE))
		PSSetShaderResource(TEXTURE_TEXTURE, config.tex);
	if (config.pal)
		PSSetShaderResource(TEXTURE_PALETTE, config.pal);
	if (m_features.texture_barrier && (config.require_one_barrier || config.require_full_barrier))
		PSSetShaderResource(TEXTURE_RT, colclip_rt ? colclip_rt : config.rt);
	if (m_features.texture_barrier && (config.require_one_barrier || config.require_full_barrier) && config.ps.IsFeedbackLoopDepth())
		// With ARM depth-stencil fetch the shader reads gl_LastFragDepthARM, not a
		// sampler, so don't bind the live depth attachment as a texture (avoids a
		// feedback-loop bind the driver may flag).
		PSSetShaderResource(TEXTURE_DEPTH, (m_features.depth_feedback && !m_arm_depth_fetch) ? config.ds : m_ds_as_rt);

	SetupSampler(config.sampler);

	if (m_vs_cb_cache.Update(config.cb_vs))
		VSSetUniformBuffer(m_vs_cb_cache);
	if (m_ps_cb_cache.Update(config.cb_ps))
		PSSetUniformBuffer(m_ps_cb_cache);

	ProgramSelector psel;
	psel.vs = config.vs;
	psel.ps.key_hi = config.ps.key_hi;
	psel.ps.key_lo = config.ps.key_lo;
	std::memset(psel.pad, 0, sizeof(psel.pad));

	SetupPipeline(psel);

	// In Time Crisis:
	// 1. Fullscreen sprite reads depth and writes alpha (rt_hazard_barrier true from config.ds == config.tex)
	// 2. Fullscreen sprite writes gray, rta hw blend blends based on dst alpha.
	// On Nvidia, 2 seems to not pick up the data written by 1 unless we add a second barrier.
	// Pretty sure GL is supposed to guarantee that the blend unit is coherent with previous pixel write out, so calling this a bug.
	bool broken_blend_coherency_barrier = false;
	if (m_bugs.broken_blend_coherency)
		broken_blend_coherency_barrier = (config.IsFeedbackLoopRT(psel.ps) || psel.ps.blend_c == 1) && GLState::rt == config.rt;
	if (config.require_one_barrier || !m_features.texture_barrier)
	{
		broken_blend_coherency_barrier = false;
	}

	// additional non-pipeline config stuff
	const bool point_size_enabled = config.vs.point_size;
	if (GLState::point_size != point_size_enabled)
	{
		if (point_size_enabled)
			glEnable(GL_PROGRAM_POINT_SIZE);
		else
			glDisable(GL_PROGRAM_POINT_SIZE);
		GLState::point_size = point_size_enabled;
	}
	if (config.topology == GSHWDrawConfig::Topology::Line)
	{
		const float line_width = config.line_expand ? config.cb_ps.ScaleFactor.z : 1.0f;
		if (GLState::line_width != line_width)
		{
			GLState::line_width = line_width;
			glLineWidth(line_width);
		}
	}

	if (primid_texture)
	{
		GL_PUSH("Destination Alpha PrimID Init");

		OMSetRenderTargets(primid_texture, nullptr, config.ds, &config.scissor);
		OMColorMaskSelector mask;
		mask.wrgba = 0;
		mask.wr = true;
		OMSetColorMaskState(mask);
		OMSetBlendState(true, GL_ONE, GL_ONE, GL_MIN);
		OMDepthStencilSelector dss = config.depth;
		dss.zwe = 0; // Don't write depth
		SetupOM(dss);

		// Compute primitiveID max that pass the date test (Draw without barrier)
		Draw(config);

		psel.ps.date = 3;
		config.alpha_second_pass.ps.date = 3;
		SetupPipeline(psel);
		PSSetShaderResource(TEXTURE_PRIMID, primid_texture);
	}

	if (draw_ds_as_rt)
	{
		// We must clear the blend equation of any dual source blending factors or
		// it may interact badly with MRTs, even if blending is disabled.
		OMSetBlendState(true);
	}

	if (config.blend.IsEffective(config.colormask))
	{
		OMSetBlendState(config.blend.enable, s_gl_blend_factors[config.blend.src_factor],
			s_gl_blend_factors[config.blend.dst_factor], s_gl_blend_ops[config.blend.op],
			s_gl_blend_factors[config.blend.src_factor_alpha], s_gl_blend_factors[config.blend.dst_factor_alpha],
			config.blend.constant_enable, config.blend.constant);
	}
	else
	{
		OMSetBlendState();
	}

	// Clear texture binding when it's bound to RT or DS.
	if (!config.tex && ((draw_rt && static_cast<GSTextureOGL*>(draw_rt)->GetID() == GLState::tex_unit[0]) ||
		(draw_ds && static_cast<GSTextureOGL*>(draw_ds)->GetID() == GLState::tex_unit[0])))
		PSSetShaderResource(TEXTURE_TEXTURE, nullptr);

	// Avoid changing framebuffer just to switch from rt+depth to rt and vice versa.
	bool fb_optimization_needs_barrier = false;
	if (!draw_rt && GLState::rt && GLState::ds == draw_ds && config.tex != GLState::rt &&
		draw_ds && GLState::rt->GetSize() == draw_ds->GetSize() && !draw_ds_as_rt)
	{
		draw_rt = GLState::rt;
		fb_optimization_needs_barrier = !GLState::rt_written;
	}
	else if (!draw_ds && GLState::ds && GLState::rt == draw_rt && config.tex != GLState::ds &&
		draw_rt && GLState::ds->GetSize() == draw_rt->GetSize() && !draw_ds_as_rt)
	{
		draw_ds = GLState::ds;
		fb_optimization_needs_barrier = !GLState::ds_written;
	}

	// Be careful of the rt already being bound and the blend using the RT without a barrier.
	if (fb_optimization_needs_barrier && broken_blend_coherency_barrier)
	{
		// Ensure all depth writes are finished before sampling
		GL_INS("GL: Texture barrier to flush depth or rt before reading");
		g_perfmon.Put(GSPerfMon::Barriers, 1);
		glTextureBarrier();
	}

	const bool rt_feedbackloop_pass1 = config.IsFeedbackLoopRT(config.ps);
	const bool rt_feedbackloop_pass2 = config.IsFeedbackLoopRT(config.alpha_second_pass.ps);
	if (draw_rt && !m_features.texture_barrier && (((config.require_one_barrier || (config.require_full_barrier && m_features.multidraw_fb_copy)) &&
		(rt_feedbackloop_pass1 || rt_feedbackloop_pass2))))
	{
		// Requires a copy of the RT.
		draw_rt_clone = CreateTexture(rtsize.x, rtsize.y, 1, draw_rt->GetFormat(), true);
		if (!draw_rt_clone)
			Console.Warning("GL: Failed to allocate temp texture for RT copy.");
	}

	const bool ds_feedbackloop_pass1 = config.ps.IsFeedbackLoopDepth();
	const bool ds_feedbackloop_pass2 = config.alpha_second_pass.ps.IsFeedbackLoopDepth();
	if (draw_ds && !m_features.texture_barrier && m_features.depth_feedback &&
		(config.require_one_barrier || (config.require_full_barrier && m_features.multidraw_fb_copy)) && (ds_feedbackloop_pass1 || ds_feedbackloop_pass2))
	{
		// Requires a copy of the DS.
		draw_ds_clone = CreateTexture(rtsize.x, rtsize.y, 1, draw_ds->GetFormat(), true);
		if (!draw_ds_clone)
			Console.Warning("GL: Failed to allocate temp texture for DS copy.");
	}

	OMSetRenderTargets(draw_rt, draw_ds_as_rt, draw_ds, &config.scissor);
	OMSetColorMaskState(config.colormask);
	SetupOM(config.depth);

	// Clear stencil as close as possible to the RT bind, to avoid framebuffer swaps.
	if (config.destination_alpha == GSHWDrawConfig::DestinationAlphaMode::StencilOne && need_barrier)
	{
		constexpr GLint clear_color = 1;
		glClearBufferiv(GL_STENCIL, 0, &clear_color);
	}

	SendHWDraw(config, rt_feedbackloop_pass1 ? draw_rt_clone : nullptr, draw_rt, ds_feedbackloop_pass1 ? draw_ds_clone : nullptr, draw_ds,
		config.require_one_barrier, config.require_full_barrier);

	if (config.blend_multi_pass.enable)
	{
		if (config.blend.IsEffective(config.colormask))
		{
			OMSetBlendState(config.blend_multi_pass.blend.enable, s_gl_blend_factors[config.blend_multi_pass.blend.src_factor],
				s_gl_blend_factors[config.blend_multi_pass.blend.dst_factor], s_gl_blend_ops[config.blend_multi_pass.blend.op],
				s_gl_blend_factors[config.blend_multi_pass.blend.src_factor_alpha], s_gl_blend_factors[config.blend_multi_pass.blend.dst_factor_alpha],
				config.blend_multi_pass.blend.constant_enable, config.blend_multi_pass.blend.constant);
		}
		else
		{
			OMSetBlendState();
		}
		psel.ps.no_color1 = config.blend_multi_pass.no_color1;
		psel.ps.blend_hw = config.blend_multi_pass.blend_hw;
		psel.ps.dither = config.blend_multi_pass.dither;
		SetupPipeline(psel);
		Draw(config);
	}

	if (config.alpha_second_pass.enable)
	{
		// cbuffer will definitely be dirty if aref changes, no need to check it
		if (config.cb_ps.FogColor_AREF.a != config.alpha_second_pass.ps_aref)
		{
			config.cb_ps.FogColor_AREF.a = config.alpha_second_pass.ps_aref;
			PSSetUniformBuffer(config.cb_ps);
		}

		psel.ps = config.alpha_second_pass.ps;
		SetupPipeline(psel);
		OMSetColorMaskState(config.alpha_second_pass.colormask);
		if (config.blend.IsEffective(config.alpha_second_pass.colormask))
		{
			OMSetBlendState(config.blend.enable, s_gl_blend_factors[config.blend.src_factor],
				s_gl_blend_factors[config.blend.dst_factor], s_gl_blend_ops[config.blend.op],
				s_gl_blend_factors[config.blend.src_factor_alpha], s_gl_blend_factors[config.blend.dst_factor_alpha],
				config.blend.constant_enable, config.blend.constant);
		}
		else
		{
			OMSetBlendState();
		}
		const bool one_barrier = config.alpha_second_pass.require_one_barrier && m_features.feedback_loops();
		SetupOM(config.alpha_second_pass.depth);
		SendHWDraw(config, rt_feedbackloop_pass2 ? draw_rt_clone : nullptr, draw_rt, ds_feedbackloop_pass2 ? draw_ds_clone : nullptr, draw_ds,
			one_barrier, config.alpha_second_pass.require_full_barrier);
	}

	if (colclip_rt)
	{
		config.colclip_update_area = config.colclip_update_area.runion(config.drawarea);

		if ((config.colclip_mode == GSHWDrawConfig::ColClipMode::ResolveOnly || config.colclip_mode == GSHWDrawConfig::ColClipMode::ConvertAndResolve))
		{
			const GSVector2i size = config.rt->GetSize();
			const GSVector4 dRect(config.colclip_update_area);
			const GSVector4 sRect = dRect / GSVector4(size.x, size.y).xyxy();
			StretchRect(colclip_rt, sRect, config.rt, dRect, ShaderConvert::COLCLIP_RESOLVE, Nearest);
			Recycle(colclip_rt);

			g_gs_device->SetColorClipTexture(nullptr);
		}
	}
}

void GSDeviceOGL::FeedbackCopyAndBind(const GSHWDrawConfig& config,
	GSTexture* rt, GSTexture* rt_clone, GSTexture* ds, GSTexture* ds_clone, const GSVector4i& copyarea)
{
	if (rt_clone)
	{
		DoCopyRect(rt, rt_clone, copyarea, copyarea.left, copyarea.top);
		PSSetShaderResource(2, rt_clone);
		if (config.tex_hazard == GSHWDrawConfig::TEX_HAZARD_RT)
			PSSetShaderResource(0, rt_clone);
	}
	if (ds_clone)
	{
		DoCopyRect(ds, ds_clone, copyarea, copyarea.left, copyarea.top);
		PSSetShaderResource(4, ds_clone);
		if (config.tex_hazard == GSHWDrawConfig::TEX_HAZARD_DEPTH)
			PSSetShaderResource(0, ds_clone);
	}
}

// Choose the best copy area based on the hazards and whether we need RT and/or DS copies.
void GSDeviceOGL::FeedbackCopyAndBind(const GSHWDrawConfig& config,
	GSTexture* rt, GSTexture* rt_clone, GSTexture* ds, GSTexture* ds_clone,
	const GSVector4i& copyarea, const GSVector4i& samplearea)
{
	const GSVector4i rtsize = (rt ? rt : ds)->GetRect();

	if (config.tex_hazard != GSHWDrawConfig::TEX_HAZARD_NONE)
	{
		const GSVector4i union_rect = config.drawarea.runion(config.samplearea);
		const u32 size_union = union_rect.width() * union_rect.height();
		const u32 size_indiv = config.drawarea.width() * config.drawarea.height() +
			config.samplearea.width() * config.samplearea.height();

		// Do an individual copy if the union is larger than the sum of individual areas.
		if (size_union > size_indiv)
		{
			FeedbackCopyAndBind(config, rt, rt_clone, ds, ds_clone, ProcessCopyArea(rtsize, config.drawarea));
			FeedbackCopyAndBind(config, rt, rt_clone, ds, ds_clone, ProcessCopyArea(rtsize, config.samplearea));
		}
		else
		{
			FeedbackCopyAndBind(config, rt, rt_clone, ds, ds_clone, ProcessCopyArea(rtsize, union_rect));
		}
	}
	else
	{
		// No RT/DS hazards so just need the draw area.
		FeedbackCopyAndBind(config, rt, rt_clone, ds, ds_clone, ProcessCopyArea(rtsize, config.drawarea));
	}
}

void GSDeviceOGL::SendHWDraw(const GSHWDrawConfig& config,
	GSTexture* draw_rt_clone, GSTexture* draw_rt, GSTexture* draw_ds_clone, GSTexture* draw_ds,
	const bool one_barrier, const bool full_barrier)
{
#ifdef PCSX2_DEVBUILD
	if ((one_barrier || full_barrier) && !(config.IsFeedbackLoopRT(config.ps) || config.IsFeedbackLoopDepth(config.ps))) [[unlikely]]
		Console.Warning("OpenGL: Possible unnecessary barrier detected.");
#endif

	if (full_barrier)
	{
		pxAssert(config.drawlist && !config.drawlist->empty());
		
		const u32 indices_per_prim = config.indices_per_prim;
		const u32 draw_list_size = static_cast<u32>(config.drawlist->size());

		if (m_features.texture_barrier)
			g_perfmon.Put(GSPerfMon::Barriers, static_cast<u32>(draw_list_size));
		else
			pxAssert(config.drawlist_bbox && static_cast<u32>(config.drawlist_bbox->size()) == draw_list_size);

		if (!m_features.texture_barrier && config.tex_hazard != config.TEX_HAZARD_NONE)
			FeedbackCopyAndBind(config, draw_rt, draw_rt_clone, draw_ds, draw_ds_clone, config.samplearea);

		for (u32 n = 0, p = 0; n < draw_list_size; n++)
		{
			const u32 count = config.drawlist->at(n) * indices_per_prim;

			if (m_features.texture_barrier)
			{
				glTextureBarrier();
			}
			else
			{
				const GSVector4i bbox = config.drawlist_bbox->at(n).rintersect(config.drawarea);
				FeedbackCopyAndBind(config, draw_rt, draw_rt_clone, draw_ds, draw_ds_clone, bbox);
			}

			Draw(config, p, count);
			p += count;
		}

		return;
	}

	if (one_barrier)
	{
		if (m_features.texture_barrier)
		{
			g_perfmon.Put(GSPerfMon::Barriers, 1);
			glTextureBarrier();
		}
		else
		{
			// Optimization: For alpha second pass we can reuse the copy snapshot from the first pass.
			FeedbackCopyAndBind(config, draw_rt, draw_rt_clone, draw_ds, draw_ds_clone, config.drawarea, config.samplearea);
		}
	}

	Draw(config);
}

// Note: used as a callback of DebugMessageCallback. Don't change the signature
void GSDeviceOGL::DebugMessageCallback(GLenum gl_source, GLenum gl_type, GLuint id, GLenum gl_severity, GLsizei gl_length, const GLchar* gl_message, const void* userParam)
{
	std::string message(gl_message, gl_length >= 0 ? gl_length : strlen(gl_message));
	std::string type, severity, source;
	switch (gl_type)
	{
		case GL_DEBUG_TYPE_ERROR_ARB               : type = "Error"; break;
		case GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR_ARB : type = "Deprecated bhv"; break;
		case GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR_ARB  : type = "Undefined bhv"; break;
		case GL_DEBUG_TYPE_PORTABILITY_ARB         : type = "Portability"; break;
		case GL_DEBUG_TYPE_PERFORMANCE_ARB         : type = "Perf"; break;
		case GL_DEBUG_TYPE_OTHER_ARB               : type = "Oth"; break;
		case GL_DEBUG_TYPE_PUSH_GROUP              : return; // Don't print message injected by myself
		case GL_DEBUG_TYPE_POP_GROUP               : return; // Don't print message injected by myself
		default                                    : type = "TTT"; break;
	}
	switch (gl_severity)
	{
		case GL_DEBUG_SEVERITY_HIGH_ARB   : severity = "High"; break;
		case GL_DEBUG_SEVERITY_MEDIUM_ARB : severity = "Mid"; break;
		case GL_DEBUG_SEVERITY_LOW_ARB    : severity = "Low"; break;
		default:
			if (id == 0xFEAD)
				severity = "Cache";
			else if (id == 0xB0B0)
				severity = "REG";
			else if (id == 0xD0D0)
				severity = "EXTRA";
			break;
	}
	switch (gl_source)
	{
		case GL_DEBUG_SOURCE_API_ARB             : source = "API"; break;
		case GL_DEBUG_SOURCE_WINDOW_SYSTEM_ARB   : source = "WINDOW"; break;
		case GL_DEBUG_SOURCE_SHADER_COMPILER_ARB : source = "COMPILER"; break;
		case GL_DEBUG_SOURCE_THIRD_PARTY_ARB     : source = "3rdparty"; break;
		case GL_DEBUG_SOURCE_APPLICATION_ARB     : source = "Application"; break;
		case GL_DEBUG_SOURCE_OTHER_ARB           : source = "Others"; break;
		default                                  : source = "???"; break;
	}

	// Don't spam noisy information on the terminal
	if (gl_severity != GL_DEBUG_SEVERITY_NOTIFICATION && gl_source != GL_DEBUG_SOURCE_APPLICATION)
	{
		Console.Error("T:%s\tID:%d\tS:%s\t=> %s", type.c_str(), g_gs_renderer ? g_gs_renderer->s_n : 0, severity.c_str(), message.c_str());
	}
}

#ifdef ENABLE_OGL_DEBUG
static int s_debugGroupDepth = 0;
#endif
void GSDeviceOGL::PushDebugGroup(const char* fmt, ...)
{
#ifdef ENABLE_OGL_DEBUG
	if (!glPushDebugGroup || !GSConfig.UseDebugDevice)
		return;

	std::va_list ap;
	va_start(ap, fmt);
	const std::string buf(StringUtil::StdStringFromFormatV(fmt, ap));
	va_end(ap);

	if (!buf.empty())
	{
		glPushDebugGroup(GL_DEBUG_SOURCE_APPLICATION, 0xBAD, -1, buf.c_str());

		// Make sure the calls succeed first.
		if (glGetError() == GL_NO_ERROR)
			s_debugGroupDepth++;
	}
#endif
}

void GSDeviceOGL::PopDebugGroup()
{
#ifdef ENABLE_OGL_DEBUG
	if (!glPopDebugGroup || !GSConfig.UseDebugDevice || (s_debugGroupDepth <= 0))
		return;

	glPopDebugGroup();
	s_debugGroupDepth--;
#endif
}

void GSDeviceOGL::InsertDebugMessage(DebugMessageCategory category, const char* fmt, ...)
{
#ifdef ENABLE_OGL_DEBUG
	if (!glDebugMessageInsert || !GSConfig.UseDebugDevice)
		return;

	GLenum type, id, severity;
	switch (category)
	{
	case GSDevice::DebugMessageCategory::Cache:
			type = GL_DEBUG_TYPE_OTHER;
			id = 0xFEAD;
			severity = GL_DEBUG_SEVERITY_NOTIFICATION;
		break;
	case GSDevice::DebugMessageCategory::Reg:
		type = GL_DEBUG_TYPE_OTHER;
		id = 0xB0B0;
		severity = GL_DEBUG_SEVERITY_NOTIFICATION;
		break;
	case GSDevice::DebugMessageCategory::Debug:
		type = GL_DEBUG_TYPE_OTHER;
		id = 0xD0D0;
		severity = GL_DEBUG_SEVERITY_NOTIFICATION;
		break;
	case GSDevice::DebugMessageCategory::Message:
		type = GL_DEBUG_TYPE_ERROR;
		id = 0xDEAD;
		severity = GL_DEBUG_SEVERITY_MEDIUM;
		break;
	case GSDevice::DebugMessageCategory::Performance:
	default:
		type = GL_DEBUG_TYPE_PERFORMANCE;
		id = 0xFEE1;
		severity = GL_DEBUG_SEVERITY_NOTIFICATION;
		break;
	}

	std::va_list ap;
	va_start(ap, fmt);
	const std::string buf(StringUtil::StdStringFromFormatV(fmt, ap));
	va_end(ap);
	if (!buf.empty())
		glDebugMessageInsert(GL_DEBUG_SOURCE_APPLICATION, type, id, severity, buf.size(), buf.c_str());
#endif
}