gl_rasterizer/lighting: fix LUT interpolation

7 files changed, 102 insertions, 116 deletions

diff --git a/src/video_core/pica_state.h b/src/video_core/pica_state.h
index f46db09fb..3b00df0b3 100644
--- a/src/video_core/pica_state.h
+++ b/src/video_core/pica_state.h
@@ -87,12 +87,18 @@ struct State {
             // LUT value, encoded as 12-bit fixed point, with 12 fraction bits
             BitField<0, 12, u32> value; // 0.0.12 fixed point
 
-            // Used by HW for efficient interpolation, Citra does not use these
-            BitField<12, 12, s32> difference; // 1.0.11 fixed point
+            // Used for efficient interpolation.
+            BitField<12, 11, u32> difference; // 0.0.11 fixed point
+            BitField<23, 1, u32> neg_difference;
 
-            float ToFloat() {
+            float ToFloat() const {
                 return static_cast<float>(value) / 4095.f;
             }
+
+            float DiffToFloat() const {
+                float diff = static_cast<float>(difference) / 2047.f;
+                return neg_difference ? -diff : diff;
+            }
         };
 
         std::array<std::array<LutEntry, 256>, 24> luts;
diff --git a/src/video_core/regs_lighting.h b/src/video_core/regs_lighting.h
index 7221d1688..b89709cfe 100644
--- a/src/video_core/regs_lighting.h
+++ b/src/video_core/regs_lighting.h
@@ -26,6 +26,8 @@ struct LightingRegs {
         DistanceAttenuation = 16,
     };
 
+    static constexpr unsigned NumLightingSampler = 24;
+
     static LightingSampler SpotlightAttenuationSampler(unsigned index) {
         return static_cast<LightingSampler>(
             static_cast<unsigned>(LightingSampler::SpotlightAttenuation) + index);
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index e6cccebf6..c73e1d6e2 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -49,9 +49,7 @@ RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
 
     uniform_block_data.dirty = true;
 
-    for (unsigned index = 0; index < lighting_luts.size(); index++) {
-        uniform_block_data.lut_dirty[index] = true;
-    }
+    uniform_block_data.lut_dirty.fill(true);
 
     uniform_block_data.fog_lut_dirty = true;
 
@@ -96,18 +94,16 @@ RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
     framebuffer.Create();
 
     // Allocate and bind lighting lut textures
-    for (size_t i = 0; i < lighting_luts.size(); ++i) {
-        lighting_luts[i].Create();
-        state.lighting_luts[i].texture_1d = lighting_luts[i].handle;
-    }
+    lighting_lut_buffer.Create();
+    state.lighting_lut.texture_buffer = lighting_lut.handle;
     state.Apply();
-
-    for (size_t i = 0; i < lighting_luts.size(); ++i) {
-        glActiveTexture(static_cast<GLenum>(GL_TEXTURE3 + i));
-        glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA32F, 256, 0, GL_RGBA, GL_FLOAT, nullptr);
-        glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-        glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
-    }
+    lighting_lut.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, lighting_lut_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER,
+                 sizeof(GLfloat) * 2 * 256 * Pica::LightingRegs::NumLightingSampler, nullptr,
+                 GL_DYNAMIC_DRAW);
+    glActiveTexture(GL_TEXTURE15);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RG32F, lighting_lut_buffer.handle);
 
     // Setup the LUT for the fog
     {
@@ -313,7 +309,7 @@ void RasterizerOpenGL::DrawTriangles() {
     }
 
     // Sync the lighting luts
-    for (unsigned index = 0; index < lighting_luts.size(); index++) {
+    for (unsigned index = 0; index < uniform_block_data.lut_dirty.size(); index++) {
         if (uniform_block_data.lut_dirty[index]) {
             SyncLightingLUT(index);
             uniform_block_data.lut_dirty[index] = false;
@@ -851,7 +847,7 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
     case PICA_REG_INDEX_WORKAROUND(lighting.lut_data[6], 0x1ce):
     case PICA_REG_INDEX_WORKAROUND(lighting.lut_data[7], 0x1cf): {
         auto& lut_config = regs.lighting.lut_config;
-        uniform_block_data.lut_dirty[lut_config.type / 4] = true;
+        uniform_block_data.lut_dirty[lut_config.type] = true;
         break;
     }
     }
@@ -1201,29 +1197,9 @@ void RasterizerOpenGL::SetShader() {
         }
 
         // Set the texture samplers to correspond to different lookup table texture units
-        GLuint uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[0]");
-        if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 3);
-        }
-        uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[1]");
-        if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 4);
-        }
-        uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[2]");
-        if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 5);
-        }
-        uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[3]");
+        GLuint uniform_lut = glGetUniformLocation(shader->shader.handle, "lighting_lut");
         if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 6);
-        }
-        uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[4]");
-        if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 7);
-        }
-        uniform_lut = glGetUniformLocation(shader->shader.handle, "lut[5]");
-        if (uniform_lut != -1) {
-            glUniform1i(uniform_lut, 8);
+            glUniform1i(uniform_lut, 15);
         }
 
         GLuint uniform_fog_lut = glGetUniformLocation(shader->shader.handle, "fog_lut");
@@ -1571,20 +1547,17 @@ void RasterizerOpenGL::SyncGlobalAmbient() {
 }
 
 void RasterizerOpenGL::SyncLightingLUT(unsigned lut_index) {
-    std::array<GLvec4, 256> new_data;
-
-    for (unsigned offset = 0; offset < new_data.size(); ++offset) {
-        new_data[offset][0] = Pica::g_state.lighting.luts[(lut_index * 4) + 0][offset].ToFloat();
-        new_data[offset][1] = Pica::g_state.lighting.luts[(lut_index * 4) + 1][offset].ToFloat();
-        new_data[offset][2] = Pica::g_state.lighting.luts[(lut_index * 4) + 2][offset].ToFloat();
-        new_data[offset][3] = Pica::g_state.lighting.luts[(lut_index * 4) + 3][offset].ToFloat();
-    }
+    std::array<GLvec2, 256> new_data;
+    const auto& source_lut = Pica::g_state.lighting.luts[lut_index];
+    std::transform(source_lut.begin(), source_lut.end(), new_data.begin(), [](const auto& entry) {
+        return GLvec2{entry.ToFloat(), entry.DiffToFloat()};
+    });
 
     if (new_data != lighting_lut_data[lut_index]) {
         lighting_lut_data[lut_index] = new_data;
-        glActiveTexture(GL_TEXTURE3 + lut_index);
-        glTexSubImage1D(GL_TEXTURE_1D, 0, 0, 256, GL_RGBA, GL_FLOAT,
-                        lighting_lut_data[lut_index].data());
+        glBindBuffer(GL_TEXTURE_BUFFER, lighting_lut_buffer.handle);
+        glBufferSubData(GL_TEXTURE_BUFFER, lut_index * new_data.size() * sizeof(GLvec2),
+                        new_data.size() * sizeof(GLvec2), new_data.data());
     }
 }
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index d9a3e9d1c..79acd4230 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -263,7 +263,7 @@ private:
 
     struct {
         UniformData data;
-        bool lut_dirty[6];
+        std::array<bool, Pica::LightingRegs::NumLightingSampler> lut_dirty;
         bool fog_lut_dirty;
         bool proctex_noise_lut_dirty;
         bool proctex_color_map_dirty;
@@ -279,8 +279,9 @@ private:
     OGLBuffer uniform_buffer;
     OGLFramebuffer framebuffer;
 
-    std::array<OGLTexture, 6> lighting_luts;
-    std::array<std::array<GLvec4, 256>, 6> lighting_lut_data{};
+    OGLBuffer lighting_lut_buffer;
+    OGLTexture lighting_lut;
+    std::array<std::array<GLvec2, 256>, Pica::LightingRegs::NumLightingSampler> lighting_lut_data{};
 
     OGLTexture fog_lut;
     std::array<GLuint, 128> fog_lut_data{};
diff --git a/src/video_core/renderer_opengl/gl_shader_gen.cpp b/src/video_core/renderer_opengl/gl_shader_gen.cpp
index 540cbb9d0..0c7c4dd5c 100644
--- a/src/video_core/renderer_opengl/gl_shader_gen.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_gen.cpp
@@ -562,9 +562,9 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
     out += "vec3 normal = quaternion_rotate(normalized_normquat, surface_normal);\n";
     out += "vec3 tangent = quaternion_rotate(normalized_normquat, surface_tangent);\n";
 
-    // Gets the index into the specified lookup table for specular lighting
-    auto GetLutIndex = [&lighting](unsigned light_num, LightingRegs::LightingLutInput input,
-                                   bool abs) {
+    // Samples the specified lookup table for specular lighting
+    auto GetLutValue = [&lighting](LightingRegs::LightingSampler sampler, unsigned light_num,
+                                   LightingRegs::LightingLutInput input, bool abs) {
         std::string index;
         switch (input) {
         case LightingRegs::LightingLutInput::NH:
@@ -610,22 +610,18 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
             break;
         }
 
+        std::string sampler_string = std::to_string(static_cast<unsigned>(sampler));
+
         if (abs) {
             // LUT index is in the range of (0.0, 1.0)
             index = lighting.light[light_num].two_sided_diffuse ? "abs(" + index + ")"
                                                                 : "max(" + index + ", 0.0)";
+            return "LookupLightingLUTUnsigned(" + sampler_string + ", " + index + ")";
         } else {
             // LUT index is in the range of (-1.0, 1.0)
-            index = "((" + index + " < 0) ? " + index + " + 2.0 : " + index + ") / 2.0";
+            return "LookupLightingLUTSigned(" + sampler_string + ", " + index + ")";
         }
 
-        return "(OFFSET_256 + SCALE_256 * clamp(" + index + ", 0.0, 1.0))";
-    };
-
-    // Gets the lighting lookup table value given the specified sampler and index
-    auto GetLutValue = [](LightingRegs::LightingSampler sampler, std::string lut_index) {
-        return std::string("texture(lut[" + std::to_string((unsigned)sampler / 4) + "], " +
-                           lut_index + ")[" + std::to_string((unsigned)sampler & 3) + "]");
     };
 
     // Write the code to emulate each enabled light
@@ -653,21 +649,21 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
         if (light_config.spot_atten_enable &&
             LightingRegs::IsLightingSamplerSupported(
                 lighting.config, LightingRegs::LightingSampler::SpotlightAttenuation)) {
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_sp.type, lighting.lut_sp.abs_input);
-            auto sampler = LightingRegs::SpotlightAttenuationSampler(light_config.num);
-            spot_atten = "(" + std::to_string(lighting.lut_sp.scale) + " * " +
-                         GetLutValue(sampler, index) + ")";
+            std::string value =
+                GetLutValue(LightingRegs::SpotlightAttenuationSampler(light_config.num),
+                            light_config.num, lighting.lut_sp.type, lighting.lut_sp.abs_input);
+            spot_atten = "(" + std::to_string(lighting.lut_sp.scale) + " * " + value + ")";
         }
 
         // If enabled, compute distance attenuation value
         std::string dist_atten = "1.0";
         if (light_config.dist_atten_enable) {
-            std::string index = "(" + light_src + ".dist_atten_scale * length(-view - " +
-                                light_src + ".position) + " + light_src + ".dist_atten_bias)";
-            index = "(OFFSET_256 + SCALE_256 * clamp(" + index + ", 0.0, 1.0))";
+            std::string index = "clamp(" + light_src + ".dist_atten_scale * length(-view - " +
+                                light_src + ".position) + " + light_src +
+                                ".dist_atten_bias, 0.0, 1.0)";
             auto sampler = LightingRegs::DistanceAttenuationSampler(light_config.num);
-            dist_atten = GetLutValue(sampler, index);
+            dist_atten = "LookupLightingLUTUnsigned(" +
+                         std::to_string(static_cast<unsigned>(sampler)) + "," + index + ")";
         }
 
         // If enabled, clamp specular component if lighting result is negative
@@ -686,10 +682,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
             LightingRegs::IsLightingSamplerSupported(
                 lighting.config, LightingRegs::LightingSampler::Distribution0)) {
             // Lookup specular "distribution 0" LUT value
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_d0.type, lighting.lut_d0.abs_input);
-            d0_lut_value = "(" + std::to_string(lighting.lut_d0.scale) + " * " +
-                           GetLutValue(LightingRegs::LightingSampler::Distribution0, index) + ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::Distribution0, light_config.num,
+                            lighting.lut_d0.type, lighting.lut_d0.abs_input);
+            d0_lut_value = "(" + std::to_string(lighting.lut_d0.scale) + " * " + value + ")";
         }
         std::string specular_0 = "(" + d0_lut_value + " * " + light_src + ".specular_0)";
         if (light_config.geometric_factor_0) {
@@ -700,10 +696,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
         if (lighting.lut_rr.enable &&
             LightingRegs::IsLightingSamplerSupported(lighting.config,
                                                      LightingRegs::LightingSampler::ReflectRed)) {
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_rr.type, lighting.lut_rr.abs_input);
-            std::string value = "(" + std::to_string(lighting.lut_rr.scale) + " * " +
-                                GetLutValue(LightingRegs::LightingSampler::ReflectRed, index) + ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::ReflectRed, light_config.num,
+                            lighting.lut_rr.type, lighting.lut_rr.abs_input);
+            value = "(" + std::to_string(lighting.lut_rr.scale) + " * " + value + ")";
             out += "refl_value.r = " + value + ";\n";
         } else {
             out += "refl_value.r = 1.0;\n";
@@ -713,11 +709,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
         if (lighting.lut_rg.enable &&
             LightingRegs::IsLightingSamplerSupported(lighting.config,
                                                      LightingRegs::LightingSampler::ReflectGreen)) {
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_rg.type, lighting.lut_rg.abs_input);
-            std::string value = "(" + std::to_string(lighting.lut_rg.scale) + " * " +
-                                GetLutValue(LightingRegs::LightingSampler::ReflectGreen, index) +
-                                ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::ReflectGreen, light_config.num,
+                            lighting.lut_rg.type, lighting.lut_rg.abs_input);
+            value = "(" + std::to_string(lighting.lut_rg.scale) + " * " + value + ")";
             out += "refl_value.g = " + value + ";\n";
         } else {
             out += "refl_value.g = refl_value.r;\n";
@@ -727,11 +722,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
         if (lighting.lut_rb.enable &&
             LightingRegs::IsLightingSamplerSupported(lighting.config,
                                                      LightingRegs::LightingSampler::ReflectBlue)) {
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_rb.type, lighting.lut_rb.abs_input);
-            std::string value = "(" + std::to_string(lighting.lut_rb.scale) + " * " +
-                                GetLutValue(LightingRegs::LightingSampler::ReflectBlue, index) +
-                                ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::ReflectBlue, light_config.num,
+                            lighting.lut_rb.type, lighting.lut_rb.abs_input);
+            value = "(" + std::to_string(lighting.lut_rb.scale) + " * " + value + ")";
             out += "refl_value.b = " + value + ";\n";
         } else {
             out += "refl_value.b = refl_value.r;\n";
@@ -743,10 +737,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
             LightingRegs::IsLightingSamplerSupported(
                 lighting.config, LightingRegs::LightingSampler::Distribution1)) {
             // Lookup specular "distribution 1" LUT value
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_d1.type, lighting.lut_d1.abs_input);
-            d1_lut_value = "(" + std::to_string(lighting.lut_d1.scale) + " * " +
-                           GetLutValue(LightingRegs::LightingSampler::Distribution1, index) + ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::Distribution1, light_config.num,
+                            lighting.lut_d1.type, lighting.lut_d1.abs_input);
+            d1_lut_value = "(" + std::to_string(lighting.lut_d1.scale) + " * " + value + ")";
         }
         std::string specular_1 =
             "(" + d1_lut_value + " * refl_value * " + light_src + ".specular_1)";
@@ -759,10 +753,10 @@ static void WriteLighting(std::string& out, const PicaShaderConfig& config) {
             LightingRegs::IsLightingSamplerSupported(lighting.config,
                                                      LightingRegs::LightingSampler::Fresnel)) {
             // Lookup fresnel LUT value
-            std::string index =
-                GetLutIndex(light_config.num, lighting.lut_fr.type, lighting.lut_fr.abs_input);
-            std::string value = "(" + std::to_string(lighting.lut_fr.scale) + " * " +
-                                GetLutValue(LightingRegs::LightingSampler::Fresnel, index) + ")";
+            std::string value =
+                GetLutValue(LightingRegs::LightingSampler::Fresnel, light_config.num,
+                            lighting.lut_fr.type, lighting.lut_fr.abs_input);
+            value = "(" + std::to_string(lighting.lut_fr.scale) + " * " + value + ")";
 
             // Enabled for difffuse lighting alpha component
             if (lighting.fresnel_selector == LightingRegs::LightingFresnelSelector::PrimaryAlpha ||
@@ -1016,10 +1010,6 @@ std::string GenerateFragmentShader(const PicaShaderConfig& config) {
 #define NUM_TEV_STAGES 6
 #define NUM_LIGHTS 8
 
-// Texture coordinate offsets and scales
-#define OFFSET_256 (0.5 / 256.0)
-#define SCALE_256 (255.0 / 256.0)
-
 in vec4 primary_color;
 in vec2 texcoord[3];
 in float texcoord0_w;
@@ -1061,7 +1051,7 @@ layout (std140) uniform shader_data {
 };
 
 uniform sampler2D tex[3];
-uniform sampler1D lut[6];
+uniform samplerBuffer lighting_lut;
 uniform usampler1D fog_lut;
 uniform sampler1D proctex_noise_lut;
 uniform sampler1D proctex_color_map;
@@ -1074,6 +1064,24 @@ vec3 quaternion_rotate(vec4 q, vec3 v) {
     return v + 2.0 * cross(q.xyz, cross(q.xyz, v) + q.w * v);
 }
 
+float LookupLightingLUT(int lut_index, int index, float delta) {
+    vec2 entry = texelFetch(lighting_lut, lut_index * 256 + index).rg;
+    return entry.r + entry.g * delta;
+}
+
+float LookupLightingLUTUnsigned(int lut_index, float pos) {
+    int index = clamp(int(pos * 256.0), 0, 255);
+    float delta = pos * 256.0 - 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;
+    if (index < 0) index += 256;
+    return LookupLightingLUT(lut_index, index, delta);
+}
+
 )";
 
     if (config.state.proctex.enable)
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index bf837a7fb..40d7cee96 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -52,9 +52,7 @@ OpenGLState::OpenGLState() {
         texture_unit.sampler = 0;
     }
 
-    for (auto& lut : lighting_luts) {
-        lut.texture_1d = 0;
-    }
+    lighting_lut.texture_buffer = 0;
 
     fog_lut.texture_1d = 0;
 
@@ -194,11 +192,9 @@ void OpenGLState::Apply() const {
     }
 
     // Lighting LUTs
-    for (unsigned i = 0; i < ARRAY_SIZE(lighting_luts); ++i) {
-        if (lighting_luts[i].texture_1d != cur_state.lighting_luts[i].texture_1d) {
-            glActiveTexture(GL_TEXTURE3 + i);
-            glBindTexture(GL_TEXTURE_1D, lighting_luts[i].texture_1d);
-        }
+    if (lighting_lut.texture_buffer != cur_state.lighting_lut.texture_buffer) {
+        glActiveTexture(GL_TEXTURE15);
+        glBindTexture(GL_TEXTURE_BUFFER, cur_state.lighting_lut.texture_buffer);
     }
 
     // Fog LUT
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 7dcc03bd5..d524d0625 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -64,8 +64,8 @@ public:
     } texture_units[3];
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
-    } lighting_luts[6];
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
+    } lighting_lut;
 
     struct {
         GLuint texture_1d; // GL_TEXTURE_BINDING_1D