video_core: add workarounds to enable GLES support

video_core: shorten GetGLSLVersionString

video_core: make GLES version and extensions consistent

video_core: move some logic to LoadShader

video_core: deduplicate fragment shader precision specifier

src/video_core/CMakeLists.txt

@@ -41,6 +41,8 @@ add_library(video_core STATIC
     renderer_opengl/gl_state.h
     renderer_opengl/gl_stream_buffer.cpp
     renderer_opengl/gl_stream_buffer.h
+    renderer_opengl/gl_vars.cpp
+    renderer_opengl/gl_vars.h
     renderer_opengl/pica_to_gl.h
     renderer_opengl/renderer_opengl.cpp
     renderer_opengl/renderer_opengl.h

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -28,6 +28,7 @@
 #include "video_core/renderer_base.h"
 #include "video_core/renderer_opengl/gl_rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_state.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 #include "video_core/utils.h"
 #include "video_core/video_core.h"
 
@@ -50,6 +51,17 @@ static constexpr std::array<FormatTuple, 5> fb_format_tuples = {{
     {GL_RGBA4, GL_RGBA, GL_UNSIGNED_SHORT_4_4_4_4},   // RGBA4
 }};
 
+// Same as above, with minor changes for OpenGL ES. Replaced
+// GL_UNSIGNED_INT_8_8_8_8 with GL_UNSIGNED_BYTE and
+// GL_BGR with GL_RGB
+static constexpr std::array<FormatTuple, 5> fb_format_tuples_oes = {{
+    {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE},            // RGBA8
+    {GL_RGB8, GL_RGB, GL_UNSIGNED_BYTE},              // RGB8
+    {GL_RGB5_A1, GL_RGBA, GL_UNSIGNED_SHORT_5_5_5_1}, // RGB5A1
+    {GL_RGB565, GL_RGB, GL_UNSIGNED_SHORT_5_6_5},     // RGB565
+    {GL_RGBA4, GL_RGBA, GL_UNSIGNED_SHORT_4_4_4_4},   // RGBA4
+}};
+
 static constexpr std::array<FormatTuple, 4> depth_format_tuples = {{
     {GL_DEPTH_COMPONENT16, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT}, // D16
     {},
@@ -63,6 +75,9 @@ static const FormatTuple& GetFormatTuple(PixelFormat pixel_format) {
     const SurfaceType type = SurfaceParams::GetFormatType(pixel_format);
     if (type == SurfaceType::Color) {
         ASSERT(static_cast<std::size_t>(pixel_format) < fb_format_tuples.size());
+        if (GLES) {
+            return fb_format_tuples_oes[static_cast<unsigned int>(pixel_format)];
+        }
         return fb_format_tuples[static_cast<unsigned int>(pixel_format)];
     } else if (type == SurfaceType::Depth || type == SurfaceType::DepthStencil) {
         std::size_t tuple_idx = static_cast<std::size_t>(pixel_format) - 14;
@@ -72,6 +87,77 @@ static const FormatTuple& GetFormatTuple(PixelFormat pixel_format) {
     return tex_tuple;
 }
 
+/**
+ * OpenGL ES does not support glGetTexImage. Obtain the pixels by attaching the
+ * texture to a framebuffer.
+ * Originally from https://github.com/apitrace/apitrace/blob/master/retrace/glstate_images.cpp
+ */
+static inline void GetTexImageOES(GLenum target, GLint level, GLenum format, GLenum type,
+                                  GLint height, GLint width, GLint depth, GLubyte* pixels) {
+
+    memset(pixels, 0x80, height * width * 4);
+
+    GLenum texture_binding = GL_NONE;
+    switch (target) {
+    case GL_TEXTURE_2D:
+        texture_binding = GL_TEXTURE_BINDING_2D;
+        break;
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
+        texture_binding = GL_TEXTURE_BINDING_CUBE_MAP;
+        break;
+    case GL_TEXTURE_3D_OES:
+        texture_binding = GL_TEXTURE_BINDING_3D_OES;
+    default:
+        return;
+    }
+
+    GLint texture = 0;
+    glGetIntegerv(texture_binding, &texture);
+    if (!texture) {
+        return;
+    }
+
+    GLint prev_fbo = 0;
+    GLuint fbo = 0;
+    glGetIntegerv(GL_FRAMEBUFFER_BINDING, &prev_fbo);
+    glGenFramebuffers(1, &fbo);
+    glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+
+    switch (target) {
+    case GL_TEXTURE_2D:
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
+    case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
+    case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z: {
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, level);
+        GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+        if (status != GL_FRAMEBUFFER_COMPLETE) {
+            LOG_DEBUG(Render_OpenGL, "Framebuffer is incomplete, status: {:X}", status);
+        }
+        glReadPixels(0, 0, width, height, format, type, pixels);
+        break;
+    }
+    case GL_TEXTURE_3D_OES:
+        for (int i = 0; i < depth; i++) {
+            glFramebufferTexture3D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_3D, texture,
+                                   level, i);
+            glReadPixels(0, 0, width, height, format, type, pixels + 4 * i * width * height);
+        }
+        break;
+    }
+
+    glBindFramebuffer(GL_FRAMEBUFFER, prev_fbo);
+
+    glDeleteFramebuffers(1, &fbo);
+}
+
 template <typename Map, typename Interval>
 constexpr auto RangeFromInterval(Map& map, const Interval& interval) {
     return boost::make_iterator_range(map.equal_range(interval));
@@ -841,7 +927,12 @@ void CachedSurface::DownloadGLTexture(const MathUtil::Rectangle<u32>& rect, GLui
         state.Apply();
 
         glActiveTexture(GL_TEXTURE0);
+        if (GLES) {
+            GetTexImageOES(GL_TEXTURE_2D, 0, tuple.format, tuple.type, height, width, 0,
+                           &gl_buffer[buffer_offset]);
+        } else {
             glGetTexImage(GL_TEXTURE_2D, 0, tuple.format, tuple.type, &gl_buffer[buffer_offset]);
+        }
     } else {
         state.ResetTexture(texture.handle);
         state.draw.read_framebuffer = read_fb_handle;
@@ -982,16 +1073,15 @@ RasterizerCacheOpenGL::RasterizerCacheOpenGL() {
     d24s8_abgr_buffer.Create();
     d24s8_abgr_buffer_size = 0;
 
-    const char* vs_source = R"(
+    std::string vs_source = R"(
-#version 330 core
 const vec2 vertices[4] = vec2[4](vec2(-1.0, -1.0), vec2(1.0, -1.0), vec2(-1.0, 1.0), vec2(1.0, 1.0));
 void main() {
     gl_Position = vec4(vertices[gl_VertexID], 0.0, 1.0);
 }
 )";
-    const char* fs_source = R"(
-#version 330 core
 
+    std::string fs_source = GLES ? fragment_shader_precision_OES : "";
+    fs_source += R"(
 uniform samplerBuffer tbo;
 uniform vec2 tbo_size;
 uniform vec4 viewport;
@@ -1004,7 +1094,7 @@ void main() {
     color = texelFetch(tbo, tbo_offset).rabg;
 }
 )";
-    d24s8_abgr_shader.Create(vs_source, fs_source);
+    d24s8_abgr_shader.Create(vs_source.c_str(), fs_source.c_str());
 
     OpenGLState state = OpenGLState::GetCurState();
     GLuint old_program = state.draw.shader_program;

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -18,6 +18,7 @@
 #include "video_core/renderer_opengl/gl_shader_decompiler.h"
 #include "video_core/renderer_opengl/gl_shader_gen.h"
 #include "video_core/renderer_opengl/gl_shader_util.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 #include "video_core/video_core.h"
 
 using Pica::FramebufferRegs;
@@ -1250,7 +1251,6 @@ std::string GenerateFragmentShader(const PicaFSConfig& config, bool separable_sh
     const auto& state = config.state;
 
     std::string out = R"(
-#version 330 core
 #extension GL_ARB_shader_image_load_store : enable
 #extension GL_ARB_shader_image_size : enable
 #define ALLOW_SHADOW (defined(GL_ARB_shader_image_load_store) && defined(GL_ARB_shader_image_size))
@@ -1260,10 +1260,16 @@ std::string GenerateFragmentShader(const PicaFSConfig& config, bool separable_sh
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
     }
 
+    if (GLES) {
+        out += fragment_shader_precision_OES;
+    }
+
     out += GetVertexInterfaceDeclaration(false, separable_shader);
 
     out += R"(
+#ifndef CITRA_GLES
 in vec4 gl_FragCoord;
+#endif // CITRA_GLES
 
 out vec4 color;
 
@@ -1300,13 +1306,13 @@ float LookupLightingLUT(int lut_index, int index, float delta) {
 
 float LookupLightingLUTUnsigned(int lut_index, float pos) {
     int index = clamp(int(pos * 256.0), 0, 255);
-    float delta = pos * 256.0 - index;
+    float delta = pos * 256.0 - float(index);
     return LookupLightingLUT(lut_index, index, delta);
 }
 
 float LookupLightingLUTSigned(int lut_index, float pos) {
     int index = clamp(int(pos * 128.0), -128, 127);
-    float delta = pos * 128.0 - index;
+    float delta = pos * 128.0 - float(index);
     if (index < 0) index += 256;
     return LookupLightingLUT(lut_index, index, delta);
 }
@@ -1494,10 +1500,10 @@ vec4 secondary_fragment_color = vec4(0.0);
         // Negate the condition if we have to keep only the pixels outside the scissor box
         if (state.scissor_test_mode == RasterizerRegs::ScissorMode::Include)
             out += "!";
-        out += "(gl_FragCoord.x >= scissor_x1 && "
+        out += "(gl_FragCoord.x >= float(scissor_x1) && "
-               "gl_FragCoord.y >= scissor_y1 && "
+               "gl_FragCoord.y >= float(scissor_y1) && "
-               "gl_FragCoord.x < scissor_x2 && "
+               "gl_FragCoord.x < float(scissor_x2) && "
-               "gl_FragCoord.y < scissor_y2)) discard;\n";
+               "gl_FragCoord.y < float(scissor_y2))) discard;\n";
     }
 
     // After perspective divide, OpenGL transform z_over_w from [-1, 1] to [near, far]. Here we use
@@ -1529,7 +1535,7 @@ vec4 secondary_fragment_color = vec4(0.0);
     if (state.fog_mode == TexturingRegs::FogMode::Fog) {
         // Get index into fog LUT
         if (state.fog_flip) {
-            out += "float fog_index = (1.0 - depth) * 128.0;\n";
+            out += "float fog_index = (1.0 - float(depth)) * 128.0;\n";
         } else {
             out += "float fog_index = depth * 128.0;\n";
         }
@@ -1591,7 +1597,7 @@ do {
 }
 
 std::string GenerateTrivialVertexShader(bool separable_shader) {
-    std::string out = "#version 330 core\n";
+    std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
     }
@@ -1626,8 +1632,10 @@ void main() {
     normquat = vert_normquat;
     view = vert_view;
     gl_Position = vert_position;
+#if !defined(CITRA_GLES) || defined(GL_EXT_clip_cull_distance)
     gl_ClipDistance[0] = -vert_position.z; // fixed PICA clipping plane z <= 0
     gl_ClipDistance[1] = dot(clip_coef, vert_position);
+#endif // !defined(CITRA_GLES) || defined(GL_EXT_clip_cull_distance)
 }
 )";
 
@@ -1636,7 +1644,7 @@ void main() {
 
 std::optional<std::string> GenerateVertexShader(const Pica::Shader::ShaderSetup& setup,
                                                 const PicaVSConfig& config, bool separable_shader) {
-    std::string out = "#version 330 core\n";
+    std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
     }
@@ -1744,9 +1752,12 @@ struct Vertex {
            semantic(VSOutputAttributes::POSITION_Y) + ", " +
            semantic(VSOutputAttributes::POSITION_Z) + ", " +
            semantic(VSOutputAttributes::POSITION_W) + ");\n";
+    semantic(VSOutputAttributes::POSITION_W) + ");\n";
     out += "    gl_Position = vtx_pos;\n";
+    out += "#if !defined(CITRA_GLES) || defined(GL_EXT_clip_cull_distance)\n";
     out += "    gl_ClipDistance[0] = -vtx_pos.z;\n"; // fixed PICA clipping plane z <= 0
-    out += "    gl_ClipDistance[1] = dot(clip_coef, vtx_pos);\n\n";
+    out += "    gl_ClipDistance[1] = dot(clip_coef, vtx_pos);\n";
+    out += "#endif // !defined(CITRA_GLES) || defined(GL_EXT_clip_cull_distance)\n\n";
 
     out += "    vec4 vtx_quat = GetVertexQuaternion(vtx);\n";
     out += "    normquat = mix(vtx_quat, -vtx_quat, bvec4(quats_opposite));\n\n";
@@ -1789,7 +1800,7 @@ void EmitPrim(Vertex vtx0, Vertex vtx1, Vertex vtx2) {
 };
 
 std::string GenerateFixedGeometryShader(const PicaFixedGSConfig& config, bool separable_shader) {
-    std::string out = "#version 330 core\n";
+    std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
     }
@@ -1824,7 +1835,7 @@ void main() {
 std::optional<std::string> GenerateGeometryShader(const Pica::Shader::ShaderSetup& setup,
                                                   const PicaGSConfig& config,
                                                   bool separable_shader) {
-    std::string out = "#version 330 core\n";
+    std::string out = "";
     if (separable_shader) {
         out += "#extension GL_ARB_separate_shader_objects : enable\n";
     }

src/video_core/renderer_opengl/gl_shader_util.cpp

@@ -2,15 +2,33 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <array>
+#include <string>
 #include <vector>
 #include <glad/glad.h>
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "video_core/renderer_opengl/gl_shader_util.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 
 namespace OpenGL {
 
 GLuint LoadShader(const char* source, GLenum type) {
+    const std::string version = GLES ? R"(
+#version 310 es
+
+#define CITRA_GLES
+
+#if defined(GL_ANDROID_extension_pack_es31a)
+#extension GL_ANDROID_extension_pack_es31a : enable
+#endif // defined(GL_ANDROID_extension_pack_es31a)
+
+#if defined(GL_EXT_clip_cull_distance)
+#extension GL_EXT_clip_cull_distance : enable
+#endif // defined(GL_EXT_clip_cull_distance)
+)"
+                                     : "#version 330\n";
+
     const char* debug_type;
     switch (type) {
     case GL_VERTEX_SHADER:
@@ -26,8 +44,9 @@ GLuint LoadShader(const char* source, GLenum type) {
         UNREACHABLE();
     }
 
+    std::array<const char*, 2> src_arr{version.data(), source};
     GLuint shader_id = glCreateShader(type);
-    glShaderSource(shader_id, 1, &source, nullptr);
+    glShaderSource(shader_id, static_cast<GLsizei>(src_arr.size()), src_arr.data(), nullptr);
     LOG_DEBUG(Render_OpenGL, "Compiling {} shader...", debug_type);
     glCompileShader(shader_id);
 
@@ -44,7 +63,7 @@ GLuint LoadShader(const char* source, GLenum type) {
         } else {
             LOG_ERROR(Render_OpenGL, "Error compiling {} shader:\n{}", debug_type,
                       &shader_error[0]);
-            LOG_ERROR(Render_OpenGL, "Shader source code:\n{}", source);
+            LOG_ERROR(Render_OpenGL, "Shader source code:\n{}{}", src_arr[0], src_arr[1]);
         }
     }
     return shader_id;

src/video_core/renderer_opengl/gl_shader_util.h

@@ -9,6 +9,17 @@
 
 namespace OpenGL {
 
+// High precision may or may not supported in GLES3. If it isn't, use medium precision instead.
+static constexpr char fragment_shader_precision_OES[] = R"(
+#ifdef GL_FRAGMENT_PRECISION_HIGH
+    precision highp float;
+precision highp samplerBuffer;
+#else
+    precision mediump float;
+precision mediump samplerBuffer;
+#endif // GL_FRAGMENT_PRECISION_HIGH
+)";
+
 /**
  * Utility function to create and compile an OpenGL GLSL shader
  * @param source String of the GLSL shader program

src/video_core/renderer_opengl/gl_state.cpp

@@ -6,6 +6,7 @@
 #include "common/common_funcs.h"
 #include "common/logging/log.h"
 #include "video_core/renderer_opengl/gl_state.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 
 namespace OpenGL {
 
@@ -193,9 +194,14 @@ void OpenGLState::Apply() const {
         glBlendEquationSeparate(blend.rgb_equation, blend.a_equation);
     }
 
+    // GLES3 does not support glLogicOp
+    if (!GLES) {
         if (logic_op != cur_state.logic_op) {
             glLogicOp(logic_op);
         }
+    } else {
+        LOG_TRACE(Render_OpenGL, "glLogicOps are unimplemented...");
+    }
 
     // Textures
     for (unsigned i = 0; i < ARRAY_SIZE(texture_units); ++i) {
@@ -319,7 +325,8 @@ void OpenGLState::Apply() const {
     }
 
     // Clip distance
-    for (std::size_t i = 0; i < clip_distance.size(); ++i) {
+    if (!GLES || GLAD_GL_EXT_clip_cull_distance) {
+        for (size_t i = 0; i < clip_distance.size(); ++i) {
             if (clip_distance[i] != cur_state.clip_distance[i]) {
                 if (clip_distance[i]) {
                     glEnable(GL_CLIP_DISTANCE0 + static_cast<GLenum>(i));
@@ -328,6 +335,7 @@ void OpenGLState::Apply() const {
                 }
             }
         }
+    }
 
     cur_state = *this;
 }

src/video_core/renderer_opengl/gl_vars.cpp

@@ -0,0 +1,9 @@
+// Copyright 2019 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/renderer_opengl/gl_vars.h"
+
+namespace OpenGL {
+bool GLES;
+}

src/video_core/renderer_opengl/gl_vars.h

@@ -0,0 +1,9 @@
+// Copyright 2019 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+namespace OpenGL {
+extern bool GLES;
+}

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -21,14 +21,13 @@
 #include "core/tracer/recorder.h"
 #include "video_core/debug_utils/debug_utils.h"
 #include "video_core/rasterizer_interface.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 #include "video_core/renderer_opengl/renderer_opengl.h"
 #include "video_core/video_core.h"
 
 namespace OpenGL {
 
 static const char vertex_shader[] = R"(
-#version 150 core
-
 in vec2 vert_position;
 in vec2 vert_tex_coord;
 out vec2 frag_tex_coord;
@@ -50,8 +49,6 @@ void main() {
 )";
 
 static const char fragment_shader[] = R"(
-#version 150 core
-
 in vec2 frag_tex_coord;
 out vec4 color;
 
@@ -279,7 +276,13 @@ void RendererOpenGL::InitOpenGLObjects() {
                  0.0f);
 
     // Link shaders and get variable locations
+    if (GLES) {
+        std::string frag_source(fragment_shader_precision_OES);
+        frag_source += fragment_shader;
+        shader.Create(vertex_shader, frag_source.data());
+    } else {
         shader.Create(vertex_shader, fragment_shader);
+    }
     state.draw.shader_program = shader.handle;
     state.Apply();
     uniform_modelview_matrix = glGetUniformLocation(shader.handle, "modelview_matrix");
@@ -344,7 +347,7 @@ void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
     case GPU::Regs::PixelFormat::RGBA8:
         internal_format = GL_RGBA;
         texture.gl_format = GL_RGBA;
-        texture.gl_type = GL_UNSIGNED_INT_8_8_8_8;
+        texture.gl_type = GLES ? GL_UNSIGNED_BYTE : GL_UNSIGNED_INT_8_8_8_8;
         break;
 
     case GPU::Regs::PixelFormat::RGB8:
@@ -353,7 +356,9 @@ void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
         // mostly everywhere) for words or half-words.
         // TODO: check how those behave on big-endian processors.
         internal_format = GL_RGB;
-        texture.gl_format = GL_BGR;
+
+        // GLES Dosen't support BGR , Use RGB instead
+        texture.gl_format = GLES ? GL_RGB : GL_BGR;
         texture.gl_type = GL_UNSIGNED_BYTE;
         break;
 
@@ -555,7 +560,7 @@ Core::System::ResultStatus RendererOpenGL::Init() {
         return Core::System::ResultStatus::ErrorVideoCore_ErrorGenericDrivers;
     }
 
-    if (!GLAD_GL_VERSION_3_3) {
+    if (!(GLAD_GL_VERSION_3_3 || GLAD_GL_ES_VERSION_3_1)) {
         return Core::System::ResultStatus::ErrorVideoCore_ErrorBelowGL33;
     }
 

src/video_core/video_core.cpp

@@ -7,6 +7,7 @@
 #include "core/settings.h"
 #include "video_core/pica.h"
 #include "video_core/renderer_base.h"
+#include "video_core/renderer_opengl/gl_vars.h"
 #include "video_core/renderer_opengl/renderer_opengl.h"
 #include "video_core/video_core.h"
 
@@ -36,6 +37,8 @@ Core::System::ResultStatus Init(EmuWindow& emu_window, Memory::MemorySystem& mem
     g_memory = &memory;
     Pica::Init();
 
+    OpenGL::GLES = Settings::values.use_gles;
+
     g_renderer = std::make_unique<OpenGL::RendererOpenGL>(emu_window);
     Core::System::ResultStatus result = g_renderer->Init();