Merge pull request #2816 from wwylele/proctex-lutlutlut

gl_rasterizer: use texture buffer for proctex LUT

5 files changed, 80 insertions, 70 deletions

diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index ff3f69ba3..1c6c15a58 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -117,48 +117,53 @@ RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
 
     // Setup the noise LUT for proctex
     proctex_noise_lut.Create();
-    state.proctex_noise_lut.texture_1d = proctex_noise_lut.handle;
+    state.proctex_noise_lut.texture_buffer = proctex_noise_lut.handle;
     state.Apply();
+    proctex_noise_lut_buffer.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, proctex_noise_lut_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER, sizeof(GLfloat) * 2 * 128, nullptr, GL_DYNAMIC_DRAW);
     glActiveTexture(TextureUnits::ProcTexNoiseLUT.Enum());
-    glTexImage1D(GL_TEXTURE_1D, 0, GL_RG32F, 128, 0, GL_RG, GL_FLOAT, nullptr);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RG32F, proctex_noise_lut_buffer.handle);
 
     // Setup the color map for proctex
     proctex_color_map.Create();
-    state.proctex_color_map.texture_1d = proctex_color_map.handle;
+    state.proctex_color_map.texture_buffer = proctex_color_map.handle;
     state.Apply();
+    proctex_color_map_buffer.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, proctex_color_map_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER, sizeof(GLfloat) * 2 * 128, nullptr, GL_DYNAMIC_DRAW);
     glActiveTexture(TextureUnits::ProcTexColorMap.Enum());
-    glTexImage1D(GL_TEXTURE_1D, 0, GL_RG32F, 128, 0, GL_RG, GL_FLOAT, nullptr);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RG32F, proctex_color_map_buffer.handle);
 
     // Setup the alpha map for proctex
     proctex_alpha_map.Create();
-    state.proctex_alpha_map.texture_1d = proctex_alpha_map.handle;
+    state.proctex_alpha_map.texture_buffer = proctex_alpha_map.handle;
     state.Apply();
+    proctex_alpha_map_buffer.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, proctex_alpha_map_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER, sizeof(GLfloat) * 2 * 128, nullptr, GL_DYNAMIC_DRAW);
     glActiveTexture(TextureUnits::ProcTexAlphaMap.Enum());
-    glTexImage1D(GL_TEXTURE_1D, 0, GL_RG32F, 128, 0, GL_RG, GL_FLOAT, nullptr);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RG32F, proctex_alpha_map_buffer.handle);
 
     // Setup the LUT for proctex
     proctex_lut.Create();
-    state.proctex_lut.texture_1d = proctex_lut.handle;
+    state.proctex_lut.texture_buffer = proctex_lut.handle;
     state.Apply();
+    proctex_lut_buffer.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, proctex_lut_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER, sizeof(GLfloat) * 4 * 256, nullptr, GL_DYNAMIC_DRAW);
     glActiveTexture(TextureUnits::ProcTexLUT.Enum());
-    glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA32F, 256, 0, GL_RGBA, GL_FLOAT, nullptr);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RGBA32F, proctex_lut_buffer.handle);
 
     // Setup the difference LUT for proctex
     proctex_diff_lut.Create();
-    state.proctex_diff_lut.texture_1d = proctex_diff_lut.handle;
+    state.proctex_diff_lut.texture_buffer = proctex_diff_lut.handle;
     state.Apply();
+    proctex_diff_lut_buffer.Create();
+    glBindBuffer(GL_TEXTURE_BUFFER, proctex_diff_lut_buffer.handle);
+    glBufferData(GL_TEXTURE_BUFFER, sizeof(GLfloat) * 4 * 256, nullptr, GL_DYNAMIC_DRAW);
     glActiveTexture(TextureUnits::ProcTexDiffLUT.Enum());
-    glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA32F, 256, 0, GL_RGBA, GL_FLOAT, nullptr);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
-    glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glTexBuffer(GL_TEXTURE_BUFFER, GL_RGBA32F, proctex_diff_lut_buffer.handle);
 
     // Sync fixed function OpenGL state
     SyncCullMode();
@@ -1387,7 +1392,7 @@ void RasterizerOpenGL::SyncProcTexNoise() {
 
 // helper function for SyncProcTexNoiseLUT/ColorMap/AlphaMap
 static void SyncProcTexValueLUT(const std::array<Pica::State::ProcTex::ValueEntry, 128>& lut,
-                                std::array<GLvec2, 128>& lut_data, GLenum texture) {
+                                std::array<GLvec2, 128>& lut_data, GLuint buffer) {
     std::array<GLvec2, 128> new_data;
     std::transform(lut.begin(), lut.end(), new_data.begin(), [](const auto& entry) {
         return GLvec2{entry.ToFloat(), entry.DiffToFloat()};
@@ -1395,24 +1400,24 @@ static void SyncProcTexValueLUT(const std::array<Pica::State::ProcTex::ValueEntr
 
     if (new_data != lut_data) {
         lut_data = new_data;
-        glActiveTexture(texture);
-        glTexSubImage1D(GL_TEXTURE_1D, 0, 0, 128, GL_RG, GL_FLOAT, lut_data.data());
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glBufferSubData(GL_TEXTURE_BUFFER, 0, new_data.size() * sizeof(GLvec2), new_data.data());
     }
 }
 
 void RasterizerOpenGL::SyncProcTexNoiseLUT() {
     SyncProcTexValueLUT(Pica::g_state.proctex.noise_table, proctex_noise_lut_data,
-                        TextureUnits::ProcTexNoiseLUT.Enum());
+                        proctex_noise_lut_buffer.handle);
 }
 
 void RasterizerOpenGL::SyncProcTexColorMap() {
     SyncProcTexValueLUT(Pica::g_state.proctex.color_map_table, proctex_color_map_data,
-                        TextureUnits::ProcTexColorMap.Enum());
+                        proctex_color_map_buffer.handle);
 }
 
 void RasterizerOpenGL::SyncProcTexAlphaMap() {
     SyncProcTexValueLUT(Pica::g_state.proctex.alpha_map_table, proctex_alpha_map_data,
-                        TextureUnits::ProcTexAlphaMap.Enum());
+                        proctex_alpha_map_buffer.handle);
 }
 
 void RasterizerOpenGL::SyncProcTexLUT() {
@@ -1427,8 +1432,8 @@ void RasterizerOpenGL::SyncProcTexLUT() {
 
     if (new_data != proctex_lut_data) {
         proctex_lut_data = new_data;
-        glActiveTexture(TextureUnits::ProcTexLUT.Enum());
-        glTexSubImage1D(GL_TEXTURE_1D, 0, 0, 256, GL_RGBA, GL_FLOAT, proctex_lut_data.data());
+        glBindBuffer(GL_TEXTURE_BUFFER, proctex_lut_buffer.handle);
+        glBufferSubData(GL_TEXTURE_BUFFER, 0, new_data.size() * sizeof(GLvec4), new_data.data());
     }
 }
 
@@ -1444,8 +1449,8 @@ void RasterizerOpenGL::SyncProcTexDiffLUT() {
 
     if (new_data != proctex_diff_lut_data) {
         proctex_diff_lut_data = new_data;
-        glActiveTexture(TextureUnits::ProcTexDiffLUT.Enum());
-        glTexSubImage1D(GL_TEXTURE_1D, 0, 0, 256, GL_RGBA, GL_FLOAT, proctex_diff_lut_data.data());
+        glBindBuffer(GL_TEXTURE_BUFFER, proctex_diff_lut_buffer.handle);
+        glBufferSubData(GL_TEXTURE_BUFFER, 0, new_data.size() * sizeof(GLvec4), new_data.data());
     }
 }
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index a433c1d4a..78e218efe 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -287,18 +287,23 @@ private:
     OGLTexture fog_lut;
     std::array<GLvec2, 128> fog_lut_data{};
 
+    OGLBuffer proctex_noise_lut_buffer;
     OGLTexture proctex_noise_lut;
     std::array<GLvec2, 128> proctex_noise_lut_data{};
 
+    OGLBuffer proctex_color_map_buffer;
     OGLTexture proctex_color_map;
     std::array<GLvec2, 128> proctex_color_map_data{};
 
+    OGLBuffer proctex_alpha_map_buffer;
     OGLTexture proctex_alpha_map;
     std::array<GLvec2, 128> proctex_alpha_map_data{};
 
+    OGLBuffer proctex_lut_buffer;
     OGLTexture proctex_lut;
     std::array<GLvec4, 256> proctex_lut_data{};
 
+    OGLBuffer proctex_diff_lut_buffer;
     OGLTexture proctex_diff_lut;
     std::array<GLvec4, 256> proctex_diff_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 c93b108fb..bb192affd 100644
--- a/src/video_core/renderer_opengl/gl_shader_gen.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_gen.cpp
@@ -886,12 +886,12 @@ void AppendProcTexSampler(std::string& out, const PicaShaderConfig& config) {
     // coord=1.0 is lut[127]+lut_diff[127]. For other indices, the result is interpolated using
     // value entries and difference entries.
     out += R"(
-float ProcTexLookupLUT(sampler1D lut, float coord) {
+float ProcTexLookupLUT(samplerBuffer lut, float coord) {
     coord *= 128;
     float index_i = clamp(floor(coord), 0.0, 127.0);
     float index_f = coord - index_i; // fract() cannot be used here because 128.0 needs to be
                                      // extracted as index_i = 127.0 and index_f = 1.0
-    vec2 entry = texelFetch(lut, int(index_i), 0).rg;
+    vec2 entry = texelFetch(lut, int(index_i)).rg;
     return clamp(entry.r + entry.g * index_f, 0.0, 1.0);
 }
     )";
@@ -979,14 +979,14 @@ float ProcTexNoiseCoef(vec2 x) {
         out += "int lut_index_i = int(lut_coord) + " +
                std::to_string(config.state.proctex.lut_offset) + ";\n";
         out += "float lut_index_f = fract(lut_coord);\n";
-        out += "vec4 final_color = texelFetch(proctex_lut, lut_index_i, 0) + lut_index_f * "
-               "texelFetch(proctex_diff_lut, lut_index_i, 0);\n";
+        out += "vec4 final_color = texelFetch(proctex_lut, lut_index_i) + lut_index_f * "
+               "texelFetch(proctex_diff_lut, lut_index_i);\n";
         break;
     case ProcTexFilter::Nearest:
     case ProcTexFilter::NearestMipmapLinear:
     case ProcTexFilter::NearestMipmapNearest:
         out += "lut_coord += " + std::to_string(config.state.proctex.lut_offset) + ";\n";
-        out += "vec4 final_color = texelFetch(proctex_lut, int(round(lut_coord)), 0);\n";
+        out += "vec4 final_color = texelFetch(proctex_lut, int(round(lut_coord)));\n";
         break;
     }
 
@@ -1053,11 +1053,11 @@ layout (std140) uniform shader_data {
 uniform sampler2D tex[3];
 uniform samplerBuffer lighting_lut;
 uniform samplerBuffer fog_lut;
-uniform sampler1D proctex_noise_lut;
-uniform sampler1D proctex_color_map;
-uniform sampler1D proctex_alpha_map;
-uniform sampler1D proctex_lut;
-uniform sampler1D proctex_diff_lut;
+uniform samplerBuffer proctex_noise_lut;
+uniform samplerBuffer proctex_color_map;
+uniform samplerBuffer proctex_alpha_map;
+uniform samplerBuffer proctex_lut;
+uniform samplerBuffer proctex_diff_lut;
 
 // Rotate the vector v by the quaternion q
 vec3 quaternion_rotate(vec4 q, vec3 v) {
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index eface2dea..bc9d34b84 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -56,11 +56,11 @@ OpenGLState::OpenGLState() {
 
     fog_lut.texture_buffer = 0;
 
-    proctex_lut.texture_1d = 0;
-    proctex_diff_lut.texture_1d = 0;
-    proctex_color_map.texture_1d = 0;
-    proctex_alpha_map.texture_1d = 0;
-    proctex_noise_lut.texture_1d = 0;
+    proctex_lut.texture_buffer = 0;
+    proctex_diff_lut.texture_buffer = 0;
+    proctex_color_map.texture_buffer = 0;
+    proctex_alpha_map.texture_buffer = 0;
+    proctex_noise_lut.texture_buffer = 0;
 
     draw.read_framebuffer = 0;
     draw.draw_framebuffer = 0;
@@ -204,33 +204,33 @@ void OpenGLState::Apply() const {
     }
 
     // ProcTex Noise LUT
-    if (proctex_noise_lut.texture_1d != cur_state.proctex_noise_lut.texture_1d) {
+    if (proctex_noise_lut.texture_buffer != cur_state.proctex_noise_lut.texture_buffer) {
         glActiveTexture(TextureUnits::ProcTexNoiseLUT.Enum());
-        glBindTexture(GL_TEXTURE_1D, proctex_noise_lut.texture_1d);
+        glBindTexture(GL_TEXTURE_BUFFER, proctex_noise_lut.texture_buffer);
     }
 
     // ProcTex Color Map
-    if (proctex_color_map.texture_1d != cur_state.proctex_color_map.texture_1d) {
+    if (proctex_color_map.texture_buffer != cur_state.proctex_color_map.texture_buffer) {
         glActiveTexture(TextureUnits::ProcTexColorMap.Enum());
-        glBindTexture(GL_TEXTURE_1D, proctex_color_map.texture_1d);
+        glBindTexture(GL_TEXTURE_BUFFER, proctex_color_map.texture_buffer);
     }
 
     // ProcTex Alpha Map
-    if (proctex_alpha_map.texture_1d != cur_state.proctex_alpha_map.texture_1d) {
+    if (proctex_alpha_map.texture_buffer != cur_state.proctex_alpha_map.texture_buffer) {
         glActiveTexture(TextureUnits::ProcTexAlphaMap.Enum());
-        glBindTexture(GL_TEXTURE_1D, proctex_alpha_map.texture_1d);
+        glBindTexture(GL_TEXTURE_BUFFER, proctex_alpha_map.texture_buffer);
     }
 
     // ProcTex LUT
-    if (proctex_lut.texture_1d != cur_state.proctex_lut.texture_1d) {
+    if (proctex_lut.texture_buffer != cur_state.proctex_lut.texture_buffer) {
         glActiveTexture(TextureUnits::ProcTexLUT.Enum());
-        glBindTexture(GL_TEXTURE_1D, proctex_lut.texture_1d);
+        glBindTexture(GL_TEXTURE_BUFFER, proctex_lut.texture_buffer);
     }
 
     // ProcTex Diff LUT
-    if (proctex_diff_lut.texture_1d != cur_state.proctex_diff_lut.texture_1d) {
+    if (proctex_diff_lut.texture_buffer != cur_state.proctex_diff_lut.texture_buffer) {
         glActiveTexture(TextureUnits::ProcTexDiffLUT.Enum());
-        glBindTexture(GL_TEXTURE_1D, proctex_diff_lut.texture_1d);
+        glBindTexture(GL_TEXTURE_BUFFER, proctex_diff_lut.texture_buffer);
     }
 
     // Framebuffer
@@ -274,16 +274,16 @@ void OpenGLState::ResetTexture(GLuint handle) {
         cur_state.lighting_lut.texture_buffer = 0;
     if (cur_state.fog_lut.texture_buffer == handle)
         cur_state.fog_lut.texture_buffer = 0;
-    if (cur_state.proctex_noise_lut.texture_1d == handle)
-        cur_state.proctex_noise_lut.texture_1d = 0;
-    if (cur_state.proctex_color_map.texture_1d == handle)
-        cur_state.proctex_color_map.texture_1d = 0;
-    if (cur_state.proctex_alpha_map.texture_1d == handle)
-        cur_state.proctex_alpha_map.texture_1d = 0;
-    if (cur_state.proctex_lut.texture_1d == handle)
-        cur_state.proctex_lut.texture_1d = 0;
-    if (cur_state.proctex_diff_lut.texture_1d == handle)
-        cur_state.proctex_diff_lut.texture_1d = 0;
+    if (cur_state.proctex_noise_lut.texture_buffer == handle)
+        cur_state.proctex_noise_lut.texture_buffer = 0;
+    if (cur_state.proctex_color_map.texture_buffer == handle)
+        cur_state.proctex_color_map.texture_buffer = 0;
+    if (cur_state.proctex_alpha_map.texture_buffer == handle)
+        cur_state.proctex_alpha_map.texture_buffer = 0;
+    if (cur_state.proctex_lut.texture_buffer == handle)
+        cur_state.proctex_lut.texture_buffer = 0;
+    if (cur_state.proctex_diff_lut.texture_buffer == handle)
+        cur_state.proctex_diff_lut.texture_buffer = 0;
 }
 
 void OpenGLState::ResetSampler(GLuint handle) {
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 1efcf0811..745a74479 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -95,23 +95,23 @@ public:
     } fog_lut;
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } proctex_noise_lut;
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } proctex_color_map;
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } proctex_alpha_map;
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } proctex_lut;
 
     struct {
-        GLuint texture_1d; // GL_TEXTURE_BINDING_1D
+        GLuint texture_buffer; // GL_TEXTURE_BINDING_BUFFER
     } proctex_diff_lut;
 
     struct {