video_core: Take factors into account with min/max blending functions (#6925)

* sw_framebuffer: Take factors into account for min/max blending

* renderer_gl: Take factors into account for min/max blending

* Address review comments

* gl_shader_gen: Fix frambuffer fetch on qcom and mali

* renderer_opengl: Add fallback path for mesa

* gl_shader_gen: Avoid emitting blend emulation if minmax_factor is present

externals/glad/include/glad/glad.h

@@ -1,28 +1,32 @@
 /*
 
-    OpenGL, OpenGL ES loader generated by glad 0.1.36 on Sat Apr  1 20:34:42 2023.
+    OpenGL, OpenGL ES loader generated by glad 0.1.34 on Sat Aug 26 18:38:43 2023.
 
     Language/Generator: C/C++
     Specification: gl
     APIs: gl=4.3, gles2=3.2
     Profile: core
     Extensions:
+        GL_AMD_blend_minmax_factor,
         GL_ARB_buffer_storage,
         GL_ARB_clear_texture,
         GL_ARB_get_texture_sub_image,
         GL_ARB_texture_compression_bptc,
+        GL_ARM_shader_framebuffer_fetch,
         GL_EXT_buffer_storage,
         GL_EXT_clip_cull_distance,
-        GL_EXT_texture_compression_s3tc
+        GL_EXT_shader_framebuffer_fetch,
+        GL_EXT_texture_compression_s3tc,
+        GL_NV_blend_minmax_factor
     Loader: True
     Local files: False
     Omit khrplatform: False
     Reproducible: False
 
     Commandline:
-        --profile="core" --api="gl=4.3,gles2=3.2" --generator="c" --spec="gl" --extensions="GL_ARB_buffer_storage,GL_ARB_clear_texture,GL_ARB_get_texture_sub_image,GL_ARB_texture_compression_bptc,GL_EXT_buffer_storage,GL_EXT_clip_cull_distance,GL_EXT_texture_compression_s3tc"
+        --profile="core" --api="gl=4.3,gles2=3.2" --generator="c" --spec="gl" --extensions="GL_AMD_blend_minmax_factor,GL_ARB_buffer_storage,GL_ARB_clear_texture,GL_ARB_get_texture_sub_image,GL_ARB_texture_compression_bptc,GL_ARM_shader_framebuffer_fetch,GL_EXT_buffer_storage,GL_EXT_clip_cull_distance,GL_EXT_shader_framebuffer_fetch,GL_EXT_texture_compression_s3tc,GL_NV_blend_minmax_factor"
     Online:
-        https://glad.dav1d.de/#profile=core&language=c&specification=gl&loader=on&api=gl%3D4.3&api=gles2%3D3.2&extensions=GL_ARB_buffer_storage&extensions=GL_ARB_clear_texture&extensions=GL_ARB_get_texture_sub_image&extensions=GL_ARB_texture_compression_bptc&extensions=GL_EXT_buffer_storage&extensions=GL_EXT_clip_cull_distance&extensions=GL_EXT_texture_compression_s3tc
+        https://glad.dav1d.de/#profile=core&language=c&specification=gl&loader=on&api=gl%3D4.3&api=gles2%3D3.2&extensions=GL_AMD_blend_minmax_factor&extensions=GL_ARB_buffer_storage&extensions=GL_ARB_clear_texture&extensions=GL_ARB_get_texture_sub_image&extensions=GL_ARB_texture_compression_bptc&extensions=GL_ARM_shader_framebuffer_fetch&extensions=GL_EXT_buffer_storage&extensions=GL_EXT_clip_cull_distance&extensions=GL_EXT_shader_framebuffer_fetch&extensions=GL_EXT_texture_compression_s3tc&extensions=GL_NV_blend_minmax_factor
 */
 
 
@@ -3320,6 +3324,8 @@ typedef void (APIENTRYP PFNGLGETNUNIFORMUIVPROC)(GLuint program, GLint location,
 GLAPI PFNGLGETNUNIFORMUIVPROC glad_glGetnUniformuiv;
 #define glGetnUniformuiv glad_glGetnUniformuiv
 #endif
+#define GL_FACTOR_MIN_AMD 0x901C
+#define GL_FACTOR_MAX_AMD 0x901D
 #define GL_MAP_PERSISTENT_BIT 0x0040
 #define GL_MAP_COHERENT_BIT 0x0080
 #define GL_DYNAMIC_STORAGE_BIT 0x0100
@@ -3332,10 +3338,13 @@ GLAPI PFNGLGETNUNIFORMUIVPROC glad_glGetnUniformuiv;
 #define GL_COMPRESSED_SRGB_ALPHA_BPTC_UNORM_ARB 0x8E8D
 #define GL_COMPRESSED_RGB_BPTC_SIGNED_FLOAT_ARB 0x8E8E
 #define GL_COMPRESSED_RGB_BPTC_UNSIGNED_FLOAT_ARB 0x8E8F
+#define GL_FRAGMENT_SHADER_DISCARDS_SAMPLES_EXT 0x8A52
 #define GL_COMPRESSED_RGB_S3TC_DXT1_EXT 0x83F0
 #define GL_COMPRESSED_RGBA_S3TC_DXT1_EXT 0x83F1
 #define GL_COMPRESSED_RGBA_S3TC_DXT3_EXT 0x83F2
 #define GL_COMPRESSED_RGBA_S3TC_DXT5_EXT 0x83F3
+#define GL_FETCH_PER_SAMPLE_ARM 0x8F65
+#define GL_FRAGMENT_SHADER_FRAMEBUFFER_FETCH_MRT_ARM 0x8F66
 #define GL_MAP_PERSISTENT_BIT_EXT 0x0040
 #define GL_MAP_COHERENT_BIT_EXT 0x0080
 #define GL_DYNAMIC_STORAGE_BIT_EXT 0x0100
@@ -3354,6 +3363,10 @@ GLAPI PFNGLGETNUNIFORMUIVPROC glad_glGetnUniformuiv;
 #define GL_CLIP_DISTANCE5_EXT 0x3005
 #define GL_CLIP_DISTANCE6_EXT 0x3006
 #define GL_CLIP_DISTANCE7_EXT 0x3007
+#ifndef GL_AMD_blend_minmax_factor
+#define GL_AMD_blend_minmax_factor 1
+GLAPI int GLAD_GL_AMD_blend_minmax_factor;
+#endif
 #ifndef GL_ARB_buffer_storage
 #define GL_ARB_buffer_storage 1
 GLAPI int GLAD_GL_ARB_buffer_storage;
@@ -3385,10 +3398,22 @@ GLAPI PFNGLGETCOMPRESSEDTEXTURESUBIMAGEPROC glad_glGetCompressedTextureSubImage;
 #define GL_ARB_texture_compression_bptc 1
 GLAPI int GLAD_GL_ARB_texture_compression_bptc;
 #endif
+#ifndef GL_EXT_shader_framebuffer_fetch
+#define GL_EXT_shader_framebuffer_fetch 1
+GLAPI int GLAD_GL_EXT_shader_framebuffer_fetch;
+#endif
 #ifndef GL_EXT_texture_compression_s3tc
 #define GL_EXT_texture_compression_s3tc 1
 GLAPI int GLAD_GL_EXT_texture_compression_s3tc;
 #endif
+#ifndef GL_NV_blend_minmax_factor
+#define GL_NV_blend_minmax_factor 1
+GLAPI int GLAD_GL_NV_blend_minmax_factor;
+#endif
+#ifndef GL_ARM_shader_framebuffer_fetch
+#define GL_ARM_shader_framebuffer_fetch 1
+GLAPI int GLAD_GL_ARM_shader_framebuffer_fetch;
+#endif
 #ifndef GL_EXT_buffer_storage
 #define GL_EXT_buffer_storage 1
 GLAPI int GLAD_GL_EXT_buffer_storage;
@@ -3400,10 +3425,18 @@ GLAPI PFNGLBUFFERSTORAGEEXTPROC glad_glBufferStorageEXT;
 #define GL_EXT_clip_cull_distance 1
 GLAPI int GLAD_GL_EXT_clip_cull_distance;
 #endif
+#ifndef GL_EXT_shader_framebuffer_fetch
+#define GL_EXT_shader_framebuffer_fetch 1
+GLAPI int GLAD_GL_EXT_shader_framebuffer_fetch;
+#endif
 #ifndef GL_EXT_texture_compression_s3tc
 #define GL_EXT_texture_compression_s3tc 1
 GLAPI int GLAD_GL_EXT_texture_compression_s3tc;
 #endif
+#ifndef GL_NV_blend_minmax_factor
+#define GL_NV_blend_minmax_factor 1
+GLAPI int GLAD_GL_NV_blend_minmax_factor;
+#endif
 
 #ifdef __cplusplus
 }

externals/glad/src/glad.c

@@ -1,28 +1,32 @@
 /*
 
-    OpenGL, OpenGL ES loader generated by glad 0.1.36 on Sat Apr  1 20:34:42 2023.
+    OpenGL, OpenGL ES loader generated by glad 0.1.34 on Sat Aug 26 18:38:43 2023.
 
     Language/Generator: C/C++
     Specification: gl
     APIs: gl=4.3, gles2=3.2
     Profile: core
     Extensions:
+        GL_AMD_blend_minmax_factor,
         GL_ARB_buffer_storage,
         GL_ARB_clear_texture,
         GL_ARB_get_texture_sub_image,
         GL_ARB_texture_compression_bptc,
+        GL_ARM_shader_framebuffer_fetch,
         GL_EXT_buffer_storage,
         GL_EXT_clip_cull_distance,
-        GL_EXT_texture_compression_s3tc
+        GL_EXT_shader_framebuffer_fetch,
+        GL_EXT_texture_compression_s3tc,
+        GL_NV_blend_minmax_factor
     Loader: True
     Local files: False
     Omit khrplatform: False
     Reproducible: False
 
     Commandline:
-        --profile="core" --api="gl=4.3,gles2=3.2" --generator="c" --spec="gl" --extensions="GL_ARB_buffer_storage,GL_ARB_clear_texture,GL_ARB_get_texture_sub_image,GL_ARB_texture_compression_bptc,GL_EXT_buffer_storage,GL_EXT_clip_cull_distance,GL_EXT_texture_compression_s3tc"
+        --profile="core" --api="gl=4.3,gles2=3.2" --generator="c" --spec="gl" --extensions="GL_AMD_blend_minmax_factor,GL_ARB_buffer_storage,GL_ARB_clear_texture,GL_ARB_get_texture_sub_image,GL_ARB_texture_compression_bptc,GL_ARM_shader_framebuffer_fetch,GL_EXT_buffer_storage,GL_EXT_clip_cull_distance,GL_EXT_shader_framebuffer_fetch,GL_EXT_texture_compression_s3tc,GL_NV_blend_minmax_factor"
     Online:
-        https://glad.dav1d.de/#profile=core&language=c&specification=gl&loader=on&api=gl%3D4.3&api=gles2%3D3.2&extensions=GL_ARB_buffer_storage&extensions=GL_ARB_clear_texture&extensions=GL_ARB_get_texture_sub_image&extensions=GL_ARB_texture_compression_bptc&extensions=GL_EXT_buffer_storage&extensions=GL_EXT_clip_cull_distance&extensions=GL_EXT_texture_compression_s3tc
+        https://glad.dav1d.de/#profile=core&language=c&specification=gl&loader=on&api=gl%3D4.3&api=gles2%3D3.2&extensions=GL_AMD_blend_minmax_factor&extensions=GL_ARB_buffer_storage&extensions=GL_ARB_clear_texture&extensions=GL_ARB_get_texture_sub_image&extensions=GL_ARB_texture_compression_bptc&extensions=GL_ARM_shader_framebuffer_fetch&extensions=GL_EXT_buffer_storage&extensions=GL_EXT_clip_cull_distance&extensions=GL_EXT_shader_framebuffer_fetch&extensions=GL_EXT_texture_compression_s3tc&extensions=GL_NV_blend_minmax_factor
 */
 
 #include <stdio.h>
@@ -853,13 +857,17 @@ PFNGLVIEWPORTARRAYVPROC glad_glViewportArrayv = NULL;
 PFNGLVIEWPORTINDEXEDFPROC glad_glViewportIndexedf = NULL;
 PFNGLVIEWPORTINDEXEDFVPROC glad_glViewportIndexedfv = NULL;
 PFNGLWAITSYNCPROC glad_glWaitSync = NULL;
+int GLAD_GL_AMD_blend_minmax_factor = 0;
 int GLAD_GL_ARB_buffer_storage = 0;
 int GLAD_GL_ARB_clear_texture = 0;
 int GLAD_GL_ARB_get_texture_sub_image = 0;
 int GLAD_GL_ARB_texture_compression_bptc = 0;
+int GLAD_GL_ARM_shader_framebuffer_fetch = 0;
 int GLAD_GL_EXT_buffer_storage = 0;
 int GLAD_GL_EXT_clip_cull_distance = 0;
+int GLAD_GL_EXT_shader_framebuffer_fetch = 0;
 int GLAD_GL_EXT_texture_compression_s3tc = 0;
+int GLAD_GL_NV_blend_minmax_factor = 0;
 PFNGLBUFFERSTORAGEPROC glad_glBufferStorage = NULL;
 PFNGLCLEARTEXIMAGEPROC glad_glClearTexImage = NULL;
 PFNGLCLEARTEXSUBIMAGEPROC glad_glClearTexSubImage = NULL;
@@ -1498,11 +1506,14 @@ static void load_GL_ARB_get_texture_sub_image(GLADloadproc load) {
 }
 static int find_extensionsGL(void) {
 	if (!get_exts()) return 0;
+	GLAD_GL_AMD_blend_minmax_factor = has_ext("GL_AMD_blend_minmax_factor");
 	GLAD_GL_ARB_buffer_storage = has_ext("GL_ARB_buffer_storage");
 	GLAD_GL_ARB_clear_texture = has_ext("GL_ARB_clear_texture");
 	GLAD_GL_ARB_get_texture_sub_image = has_ext("GL_ARB_get_texture_sub_image");
 	GLAD_GL_ARB_texture_compression_bptc = has_ext("GL_ARB_texture_compression_bptc");
+	GLAD_GL_EXT_shader_framebuffer_fetch = has_ext("GL_EXT_shader_framebuffer_fetch");
 	GLAD_GL_EXT_texture_compression_s3tc = has_ext("GL_EXT_texture_compression_s3tc");
+	GLAD_GL_NV_blend_minmax_factor = has_ext("GL_NV_blend_minmax_factor");
 	free_exts();
 	return 1;
 }
@@ -1971,9 +1982,12 @@ static void load_GL_EXT_buffer_storage(GLADloadproc load) {
 }
 static int find_extensionsGLES2(void) {
 	if (!get_exts()) return 0;
+	GLAD_GL_ARM_shader_framebuffer_fetch = has_ext("GL_ARM_shader_framebuffer_fetch");
 	GLAD_GL_EXT_buffer_storage = has_ext("GL_EXT_buffer_storage");
 	GLAD_GL_EXT_clip_cull_distance = has_ext("GL_EXT_clip_cull_distance");
+	GLAD_GL_EXT_shader_framebuffer_fetch = has_ext("GL_EXT_shader_framebuffer_fetch");
 	GLAD_GL_EXT_texture_compression_s3tc = has_ext("GL_EXT_texture_compression_s3tc");
+	GLAD_GL_NV_blend_minmax_factor = has_ext("GL_NV_blend_minmax_factor");
 	free_exts();
 	return 1;
 }

src/video_core/renderer_opengl/gl_driver.cpp

@@ -170,6 +170,9 @@ void Driver::CheckExtensionSupport() {
     arb_texture_compression_bptc = GLAD_GL_ARB_texture_compression_bptc;
     ext_clip_cull_distance = GLAD_GL_EXT_clip_cull_distance;
     ext_texture_compression_s3tc = GLAD_GL_EXT_texture_compression_s3tc;
+    shader_framebuffer_fetch =
+        GLAD_GL_EXT_shader_framebuffer_fetch || GLAD_GL_ARM_shader_framebuffer_fetch;
+    blend_minmax_factor = GLAD_GL_AMD_blend_minmax_factor || GLAD_GL_NV_blend_minmax_factor;
     is_suitable = GLAD_GL_VERSION_4_3 || GLAD_GL_ES_VERSION_3_1;
 }
 

src/video_core/renderer_opengl/gl_driver.h

@@ -105,6 +105,16 @@ public:
         return ext_clip_cull_distance;
     }
 
+    /// Returns true if the implementation supports (EXT/ARM)_shader_framebuffer_fetch
+    bool HasShaderFramebufferFetch() const {
+        return shader_framebuffer_fetch;
+    }
+
+    /// Returns true if the implementation supports (NV/AMD)_blend_minmax_factor
+    bool HasBlendMinMaxFactor() const {
+        return blend_minmax_factor;
+    }
+
 private:
     void ReportDriverInfo();
     void DeduceVendor();
@@ -125,6 +135,8 @@ private:
     bool ext_clip_cull_distance{};
     bool ext_texture_compression_s3tc{};
     bool arb_texture_compression_bptc{};
+    bool shader_framebuffer_fetch{};
+    bool blend_minmax_factor{};
 
     std::string_view gl_version{};
     std::string_view gpu_vendor{};

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -468,14 +468,16 @@ bool RasterizerOpenGL::Draw(bool accelerate, bool is_indexed) {
                         GL_TEXTURE_UPDATE_BARRIER_BIT | GL_FRAMEBUFFER_BARRIER_BIT);
     }
 
-    use_custom_normal = false;
-
     return succeeded;
 }
 
 void RasterizerOpenGL::SyncTextureUnits(const Framebuffer* framebuffer) {
     using TextureType = Pica::TexturingRegs::TextureConfig::TextureType;
 
+    // Reset transient draw state
+    state.color_buffer.texture_2d = 0;
+    use_custom_normal = false;
+
     const auto pica_textures = regs.texturing.GetTextures();
     for (u32 texture_index = 0; texture_index < pica_textures.size(); ++texture_index) {
         const auto& texture = pica_textures[texture_index];
@@ -519,6 +521,10 @@ void RasterizerOpenGL::SyncTextureUnits(const Framebuffer* framebuffer) {
             state.texture_units[texture_index].texture_2d = surface.Handle();
         }
     }
+
+    if (emulate_minmax_blend && !driver.HasShaderFramebufferFetch()) {
+        state.color_buffer.texture_2d = framebuffer->Attachment(SurfaceType::Color);
+    }
 }
 
 void RasterizerOpenGL::BindShadowCube(const Pica::TexturingRegs::FullTextureConfig& texture) {
@@ -760,17 +766,14 @@ void RasterizerOpenGL::SyncCullMode() {
     case Pica::RasterizerRegs::CullMode::KeepAll:
         state.cull.enabled = false;
         break;
-
     case Pica::RasterizerRegs::CullMode::KeepClockWise:
         state.cull.enabled = true;
         state.cull.front_face = GL_CW;
         break;
-
     case Pica::RasterizerRegs::CullMode::KeepCounterClockWise:
         state.cull.enabled = true;
         state.cull.front_face = GL_CCW;
         break;
-
     default:
         LOG_CRITICAL(Render_OpenGL, "Unknown cull mode {}",
                      static_cast<u32>(regs.rasterizer.cull_mode.Value()));
@@ -784,10 +787,12 @@ void RasterizerOpenGL::SyncBlendEnabled() {
 }
 
 void RasterizerOpenGL::SyncBlendFuncs() {
-    state.blend.rgb_equation =
+    const bool has_minmax_factor = driver.HasBlendMinMaxFactor();
-        PicaToGL::BlendEquation(regs.framebuffer.output_merger.alpha_blending.blend_equation_rgb);
+
-    state.blend.a_equation =
+    state.blend.rgb_equation = PicaToGL::BlendEquation(
-        PicaToGL::BlendEquation(regs.framebuffer.output_merger.alpha_blending.blend_equation_a);
+        regs.framebuffer.output_merger.alpha_blending.blend_equation_rgb, has_minmax_factor);
+    state.blend.a_equation = PicaToGL::BlendEquation(
+        regs.framebuffer.output_merger.alpha_blending.blend_equation_a, has_minmax_factor);
     state.blend.src_rgb_func =
         PicaToGL::BlendFunc(regs.framebuffer.output_merger.alpha_blending.factor_source_rgb);
     state.blend.dst_rgb_func =
@@ -796,14 +801,39 @@ void RasterizerOpenGL::SyncBlendFuncs() {
         PicaToGL::BlendFunc(regs.framebuffer.output_merger.alpha_blending.factor_source_a);
     state.blend.dst_a_func =
         PicaToGL::BlendFunc(regs.framebuffer.output_merger.alpha_blending.factor_dest_a);
+
+    if (has_minmax_factor) {
+        return;
+    }
+
+    // Blending with min/max equations is emulated in the fragment shader so
+    // configure blending to not modify the incoming fragment color.
+    emulate_minmax_blend = false;
+    if (state.EmulateColorBlend()) {
+        emulate_minmax_blend = true;
+        state.blend.rgb_equation = GL_FUNC_ADD;
+        state.blend.src_rgb_func = GL_ONE;
+        state.blend.dst_rgb_func = GL_ZERO;
+    }
+    if (state.EmulateAlphaBlend()) {
+        emulate_minmax_blend = true;
+        state.blend.a_equation = GL_FUNC_ADD;
+        state.blend.src_a_func = GL_ONE;
+        state.blend.dst_a_func = GL_ZERO;
+    }
 }
 
 void RasterizerOpenGL::SyncBlendColor() {
-    auto blend_color = PicaToGL::ColorRGBA8(regs.framebuffer.output_merger.blend_const.raw);
+    const auto blend_color = PicaToGL::ColorRGBA8(regs.framebuffer.output_merger.blend_const.raw);
     state.blend.color.red = blend_color[0];
     state.blend.color.green = blend_color[1];
     state.blend.color.blue = blend_color[2];
     state.blend.color.alpha = blend_color[3];
+
+    if (blend_color != uniform_block_data.data.blend_color) {
+        uniform_block_data.data.blend_color = blend_color;
+        uniform_block_data.dirty = true;
+    }
 }
 
 void RasterizerOpenGL::SyncLogicOp() {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -155,6 +155,7 @@ private:
     OGLTexture texture_buffer_lut_rg;
     OGLTexture texture_buffer_lut_rgba;
     bool use_custom_normal{};
+    bool emulate_minmax_blend{};
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -9,6 +9,7 @@
 #include "core/core.h"
 #include "core/telemetry_session.h"
 #include "video_core/pica_state.h"
+#include "video_core/renderer_opengl/gl_driver.h"
 #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"
@@ -60,7 +61,8 @@ out gl_PerVertex {
     return out;
 }
 
-PicaFSConfig PicaFSConfig::BuildFromRegs(const Pica::Regs& regs, bool use_normal) {
+PicaFSConfig PicaFSConfig::BuildFromRegs(const Pica::Regs& regs, bool has_blend_minmax_factor,
+                                         bool use_normal) {
     PicaFSConfig res{};
 
     auto& state = res.state;
@@ -229,6 +231,29 @@ PicaFSConfig PicaFSConfig::BuildFromRegs(const Pica::Regs& regs, bool use_normal
         state.proctex.lut_filter = regs.texturing.proctex_lut.filter;
     }
 
+    const auto alpha_eq = regs.framebuffer.output_merger.alpha_blending.blend_equation_a.Value();
+    const auto rgb_eq = regs.framebuffer.output_merger.alpha_blending.blend_equation_rgb.Value();
+    if (regs.framebuffer.output_merger.alphablend_enable && !has_blend_minmax_factor) {
+        if (rgb_eq == Pica::FramebufferRegs::BlendEquation::Max ||
+            rgb_eq == Pica::FramebufferRegs::BlendEquation::Min) {
+            state.rgb_blend.emulate_blending = true;
+            state.rgb_blend.eq = rgb_eq;
+            state.rgb_blend.src_factor =
+                regs.framebuffer.output_merger.alpha_blending.factor_source_rgb;
+            state.rgb_blend.dst_factor =
+                regs.framebuffer.output_merger.alpha_blending.factor_dest_rgb;
+        }
+        if (alpha_eq == Pica::FramebufferRegs::BlendEquation::Max ||
+            alpha_eq == Pica::FramebufferRegs::BlendEquation::Min) {
+            state.alpha_blend.emulate_blending = true;
+            state.alpha_blend.eq = alpha_eq;
+            state.alpha_blend.src_factor =
+                regs.framebuffer.output_merger.alpha_blending.factor_source_a;
+            state.alpha_blend.dst_factor =
+                regs.framebuffer.output_merger.alpha_blending.factor_dest_a;
+        }
+    }
+
     state.shadow_rendering = regs.framebuffer.output_merger.fragment_operation_mode ==
                              FramebufferRegs::FragmentOperationMode::Shadow;
     if (state.shadow_rendering) {
@@ -1222,6 +1247,103 @@ float ProcTexNoiseCoef(vec2 x) {
     }
 }
 
+static void WriteLogicOp(std::string& out, const PicaFSConfig& config) {
+    if (!GLES || config.state.alphablend_enable) {
+        return;
+    }
+    switch (config.state.logic_op) {
+    case FramebufferRegs::LogicOp::Clear:
+        out += "color = vec4(0);\n";
+        break;
+    case FramebufferRegs::LogicOp::Set:
+        out += "color = vec4(1);\n";
+        break;
+    case FramebufferRegs::LogicOp::Copy:
+        // Take the color output as-is
+        break;
+    case FramebufferRegs::LogicOp::CopyInverted:
+        out += "color = ~color;\n";
+        break;
+    case FramebufferRegs::LogicOp::NoOp:
+        // We need to discard the color, but not necessarily the depth. This is not possible
+        // with fragment shader alone, so we emulate this behavior on GLES with glColorMask.
+        break;
+    default:
+        LOG_CRITICAL(HW_GPU, "Unhandled logic_op {:x}", static_cast<int>(config.state.logic_op));
+        UNIMPLEMENTED();
+    }
+}
+
+static void WriteBlending(std::string& out, const PicaFSConfig& config) {
+    if (!config.state.rgb_blend.emulate_blending && !config.state.alpha_blend.emulate_blending)
+        [[likely]] {
+        return;
+    }
+
+    using BlendFactor = Pica::FramebufferRegs::BlendFactor;
+    out += R"(
+vec4 source_color = last_tex_env_out;
+#if defined(GL_EXT_shader_framebuffer_fetch)
+vec4 dest_color = color;
+#elif defined(GL_ARM_shader_framebuffer_fetch)
+vec4 dest_color = gl_LastFragColorARM;
+#else
+vec4 dest_color = texelFetch(colorBuffer, ivec2(gl_FragCoord.xy), 0);
+#endif
+)";
+    const auto get_factor = [&](BlendFactor factor) -> std::string {
+        switch (factor) {
+        case BlendFactor::Zero:
+            return "vec4(0.f)";
+        case BlendFactor::One:
+            return "vec4(1.f)";
+        case BlendFactor::SourceColor:
+            return "source_color";
+        case BlendFactor::OneMinusSourceColor:
+            return "vec4(1.f) - source_color";
+        case BlendFactor::DestColor:
+            return "dest_color";
+        case BlendFactor::OneMinusDestColor:
+            return "vec4(1.f) - dest_color";
+        case BlendFactor::SourceAlpha:
+            return "source_color.aaaa";
+        case BlendFactor::OneMinusSourceAlpha:
+            return "vec4(1.f) - source_color.aaaa";
+        case BlendFactor::DestAlpha:
+            return "dest_color.aaaa";
+        case BlendFactor::OneMinusDestAlpha:
+            return "vec4(1.f) - dest_color.aaaa";
+        case BlendFactor::ConstantColor:
+            return "blend_color";
+        case BlendFactor::OneMinusConstantColor:
+            return "vec4(1.f) - blend_color";
+        case BlendFactor::ConstantAlpha:
+            return "blend_color.aaaa";
+        case BlendFactor::OneMinusConstantAlpha:
+            return "vec4(1.f) - blend_color.aaaa";
+        default:
+            LOG_CRITICAL(Render_OpenGL, "Unknown blend factor {}", factor);
+            return "vec4(1.f)";
+        }
+    };
+    const auto get_func = [](Pica::FramebufferRegs::BlendEquation eq) {
+        return eq == Pica::FramebufferRegs::BlendEquation::Min ? "min" : "max";
+    };
+
+    if (config.state.rgb_blend.emulate_blending) {
+        out += fmt::format(
+            "last_tex_env_out.rgb = {}(source_color.rgb * ({}).rgb, dest_color.rgb * ({}).rgb);\n",
+            get_func(config.state.rgb_blend.eq), get_factor(config.state.rgb_blend.src_factor),
+            get_factor(config.state.rgb_blend.dst_factor));
+    }
+    if (config.state.alpha_blend.emulate_blending) {
+        out += fmt::format(
+            "last_tex_env_out.a = {}(source_color.a * ({}).a, dest_color.a * ({}).a);\n",
+            get_func(config.state.alpha_blend.eq), get_factor(config.state.alpha_blend.src_factor),
+            get_factor(config.state.alpha_blend.dst_factor));
+    }
+}
+
 ShaderDecompiler::ProgramResult GenerateFragmentShader(const PicaFSConfig& config,
                                                        bool separable_shader) {
     const auto& state = config.state;
@@ -1235,6 +1357,17 @@ ShaderDecompiler::ProgramResult GenerateFragmentShader(const PicaFSConfig& confi
         out += fragment_shader_precision_OES;
     }
 
+    out += R"(
+#if defined(GL_EXT_shader_framebuffer_fetch)
+#extension GL_EXT_shader_framebuffer_fetch : enable
+#elif defined(GL_ARM_shader_framebuffer_fetch)
+#extension GL_ARM_shader_framebuffer_fetch : enable
+#else
+layout(location = 10) uniform sampler2D colorBuffer;
+#endif
+
+)";
+
     out += GetVertexInterfaceDeclaration(false, separable_shader);
 
     out += R"(
@@ -1242,7 +1375,7 @@ ShaderDecompiler::ProgramResult GenerateFragmentShader(const PicaFSConfig& confi
 in vec4 gl_FragCoord;
 #endif // CITRA_GLES
 
-out vec4 color;
+layout(location = 0) out vec4 color;
 
 uniform sampler2D tex0;
 uniform sampler2D tex1;
@@ -1552,34 +1685,12 @@ do {
     } else {
         out += "gl_FragDepth = depth;\n";
         // Round the final fragment color to maintain the PICA's 8 bits of precision
-        out += "color = byteround(last_tex_env_out);\n";
+        out += "last_tex_env_out = byteround(last_tex_env_out);\n";
+        WriteBlending(out, config);
+        out += "color = last_tex_env_out;\n";
     }
 
-    if (GLES) {
+    WriteLogicOp(out, config);
-        if (!state.alphablend_enable) {
-            switch (state.logic_op) {
-            case FramebufferRegs::LogicOp::Clear:
-                out += "color = vec4(0);\n";
-                break;
-            case FramebufferRegs::LogicOp::Set:
-                out += "color = vec4(1);\n";
-                break;
-            case FramebufferRegs::LogicOp::Copy:
-                // Take the color output as-is
-                break;
-            case FramebufferRegs::LogicOp::CopyInverted:
-                out += "color = ~color;\n";
-                break;
-            case FramebufferRegs::LogicOp::NoOp:
-                // We need to discard the color, but not necessarily the depth. This is not possible
-                // with fragment shader alone, so we emulate this behavior on GLES with glColorMask.
-                break;
-            default:
-                LOG_CRITICAL(HW_GPU, "Unhandled logic_op {:x}", static_cast<int>(state.logic_op));
-                UNIMPLEMENTED();
-            }
-        }
-    }
 
     out += '}';
 

src/video_core/renderer_opengl/gl_shader_gen.h

@@ -11,6 +11,8 @@
 
 namespace OpenGL {
 
+class Driver;
+
 namespace ShaderDecompiler {
 struct ProgramResult;
 }
@@ -115,6 +117,13 @@ struct PicaFSConfigState {
         Pica::TexturingRegs::ProcTexFilter lut_filter;
     } proctex;
 
+    struct {
+        bool emulate_blending;
+        Pica::FramebufferRegs::BlendEquation eq;
+        Pica::FramebufferRegs::BlendFactor src_factor;
+        Pica::FramebufferRegs::BlendFactor dst_factor;
+    } rgb_blend, alpha_blend;
+
     bool shadow_rendering;
     bool shadow_texture_orthographic;
     bool use_custom_normal_map;
@@ -131,7 +140,8 @@ struct PicaFSConfigState {
 struct PicaFSConfig : Common::HashableStruct<PicaFSConfigState> {
 
     /// Construct a PicaFSConfig with the given Pica register configuration.
-    static PicaFSConfig BuildFromRegs(const Pica::Regs& regs, bool use_normal = false);
+    static PicaFSConfig BuildFromRegs(const Pica::Regs& regs, bool has_blend_minmax_factor,
+                                      bool use_normal = false);
 
     bool TevStageUpdatesCombinerBufferColor(unsigned stage_index) const {
         return (stage_index < 4) && (state.combiner_buffer_input & (1 << stage_index));

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -418,7 +418,8 @@ void ShaderProgramManager::UseTrivialGeometryShader() {
 }
 
 void ShaderProgramManager::UseFragmentShader(const Pica::Regs& regs, bool use_normal) {
-    PicaFSConfig config = PicaFSConfig::BuildFromRegs(regs, use_normal);
+    PicaFSConfig config =
+        PicaFSConfig::BuildFromRegs(regs, driver.HasBlendMinMaxFactor(), use_normal);
     auto [handle, result] = impl->fragment_shaders.Get(config);
     impl->current.fs = handle;
     impl->current.fs_hash = config.Hash();
@@ -543,7 +544,8 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                     impl->programmable_vertex_shaders.Inject(conf, decomp->second.result.code,
                                                              std::move(shader));
                 } else if (raw.GetProgramType() == ProgramType::FS) {
-                    PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig());
+                    PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig(),
+                                                                    driver.HasBlendMinMaxFactor());
                     std::scoped_lock lock(mutex);
                     impl->fragment_shaders.Inject(conf, std::move(shader));
                 } else {
@@ -655,7 +657,8 @@ void ShaderProgramManager::LoadDiskCache(const std::atomic_bool& stop_loading,
                 std::scoped_lock lock(mutex);
                 impl->programmable_vertex_shaders.Inject(conf, result->code, std::move(stage));
             } else if (raw.GetProgramType() == ProgramType::FS) {
-                PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig());
+                PicaFSConfig conf = PicaFSConfig::BuildFromRegs(raw.GetRawShaderConfig(),
+                                                                driver.HasBlendMinMaxFactor());
                 result = GenerateFragmentShader(conf, impl->separable);
                 OGLShaderStage stage{impl->separable};
                 stage.Create(result->code.c_str(), GL_FRAGMENT_SHADER);

src/video_core/renderer_opengl/gl_state.cpp

@@ -247,6 +247,12 @@ void OpenGLState::Apply() const {
         glBindTexture(GL_TEXTURE_BUFFER, texture_buffer_lut_rgba.texture_buffer);
     }
 
+    // Color buffer
+    if (color_buffer.texture_2d != cur_state.color_buffer.texture_2d) {
+        glActiveTexture(TextureUnits::TextureColorBuffer.Enum());
+        glBindTexture(GL_TEXTURE_2D, color_buffer.texture_2d);
+    }
+
     // Shadow Images
     if (image_shadow_buffer != cur_state.image_shadow_buffer) {
         glBindImageTexture(ImageUnits::ShadowBuffer, image_shadow_buffer, 0, GL_FALSE, 0,

src/video_core/renderer_opengl/gl_state.h

@@ -27,6 +27,7 @@ constexpr TextureUnit TextureBufferLUT_LF{3};
 constexpr TextureUnit TextureBufferLUT_RG{4};
 constexpr TextureUnit TextureBufferLUT_RGBA{5};
 constexpr TextureUnit TextureNormalMap{7};
+constexpr TextureUnit TextureColorBuffer{10};
 
 } // namespace TextureUnits
 
@@ -115,6 +116,10 @@ public:
         GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } texture_buffer_lut_rgba;
 
+    struct {
+        GLuint texture_2d; // GL_TEXTURE_BINDING_2D
+    } color_buffer;
+
     // GL_IMAGE_BINDING_NAME
     GLuint image_shadow_buffer;
     union {
@@ -165,6 +170,14 @@ public:
         return cur_state;
     }
 
+    bool EmulateColorBlend() const {
+        return blend.rgb_equation == GL_MIN || blend.rgb_equation == GL_MAX;
+    }
+
+    bool EmulateAlphaBlend() const {
+        return blend.a_equation == GL_MIN || blend.a_equation == GL_MAX;
+    }
+
     /// Apply this state as the current OpenGL state
     void Apply() const;
 

src/video_core/renderer_opengl/pica_to_gl.h

@@ -95,7 +95,7 @@ inline GLenum WrapMode(Pica::TexturingRegs::TextureConfig::WrapMode mode) {
     return gl_mode;
 }
 
-inline GLenum BlendEquation(Pica::FramebufferRegs::BlendEquation equation) {
+inline GLenum BlendEquation(Pica::FramebufferRegs::BlendEquation equation, bool factor_minmax) {
     static constexpr std::array<GLenum, 5> blend_equation_table{{
         GL_FUNC_ADD,              // BlendEquation::Add
         GL_FUNC_SUBTRACT,         // BlendEquation::Subtract
@@ -103,6 +103,13 @@ inline GLenum BlendEquation(Pica::FramebufferRegs::BlendEquation equation) {
         GL_MIN,                   // BlendEquation::Min
         GL_MAX,                   // BlendEquation::Max
     }};
+    static constexpr std::array<GLenum, 5> blend_equation_table_minmax{{
+        GL_FUNC_ADD,              // BlendEquation::Add
+        GL_FUNC_SUBTRACT,         // BlendEquation::Subtract
+        GL_FUNC_REVERSE_SUBTRACT, // BlendEquation::ReverseSubtract
+        GL_FACTOR_MIN_AMD,        // BlendEquation::Min
+        GL_FACTOR_MAX_AMD,        // BlendEquation::Max
+    }};
 
     const auto index = static_cast<std::size_t>(equation);
 
@@ -114,7 +121,7 @@ inline GLenum BlendEquation(Pica::FramebufferRegs::BlendEquation equation) {
         return GL_FUNC_ADD;
     }
 
-    return blend_equation_table[index];
+    return (factor_minmax ? blend_equation_table_minmax : blend_equation_table)[index];
 }
 
 inline GLenum BlendFunc(Pica::FramebufferRegs::BlendFactor factor) {

src/video_core/renderer_software/sw_framebuffer.cpp

@@ -311,19 +311,17 @@ Common::Vec4<u8> EvaluateBlendEquation(const Common::Vec4<u8>& src,
     case FramebufferRegs::BlendEquation::ReverseSubtract:
         result = (dst_result - src_result) / 255;
         break;
-    // TODO: How do these two actually work?  OpenGL doesn't include the blend factors in the
-    //       min/max computations, but is this what the 3DS actually does?
     case FramebufferRegs::BlendEquation::Min:
-        result.r() = std::min(src.r(), dest.r());
+        result.r() = std::min(src_result.r(), dst_result.r()) / 255;
-        result.g() = std::min(src.g(), dest.g());
+        result.g() = std::min(src_result.g(), dst_result.g()) / 255;
-        result.b() = std::min(src.b(), dest.b());
+        result.b() = std::min(src_result.b(), dst_result.b()) / 255;
-        result.a() = std::min(src.a(), dest.a());
+        result.a() = std::min(src_result.a(), dst_result.a()) / 255;
         break;
     case FramebufferRegs::BlendEquation::Max:
-        result.r() = std::max(src.r(), dest.r());
+        result.r() = std::max(src_result.r(), dst_result.r()) / 255;
-        result.g() = std::max(src.g(), dest.g());
+        result.g() = std::max(src_result.g(), dst_result.g()) / 255;
-        result.b() = std::max(src.b(), dest.b());
+        result.b() = std::max(src_result.b(), dst_result.b()) / 255;
-        result.a() = std::max(src.a(), dest.a());
+        result.a() = std::max(src_result.a(), dst_result.a()) / 255;
         break;
     default:
         LOG_CRITICAL(HW_GPU, "Unknown RGB blend equation 0x{:x}", equation);

src/video_core/shader/shader_uniforms.cpp

@@ -69,6 +69,7 @@ layout ({}std140) uniform shader_data {{
     vec3 tex_lod_bias;
     vec4 tex_border_color[3];
     vec4 clip_coef;
+    vec4 blend_color;
 }};
 )";
 

src/video_core/shader/shader_uniforms.h

@@ -66,9 +66,10 @@ struct UniformData {
     alignas(16) Common::Vec3f tex_lod_bias;
     alignas(16) Common::Vec4f tex_border_color[3];
     alignas(16) Common::Vec4f clip_coef;
+    alignas(16) Common::Vec4f blend_color;
 };
 
-static_assert(sizeof(UniformData) == 0x530,
+static_assert(sizeof(UniformData) == 0x540,
               "The size of the UniformData does not match the structure in the shader");
 static_assert(sizeof(UniformData) < 16384,
               "UniformData structure must be less than 16kb as per the OpenGL spec");