Fix 3

port/fast3d/gfx_opengl.cpp

@@ -34,7 +34,9 @@ struct ShaderProgram {
     GLint noise_scale_location;
     GLint three_point_filter_locations[2];
     GLint decal_mode_location;
+    GLint decal_scale_location;
     int decal_mode;
+    float decal_scale;
 };
 
 struct Framebuffer {
@@ -64,8 +66,10 @@ static MipmapFilteringMode current_mipmap_filter_mode = MIPMAP_LINEAR;
 static bool current_textures_linear_filter[2] = {false, false};
 static struct ShaderProgram* current_shader_program;
 
-// Decals are clipped to the geometry they lie on, using a copy of the depth buffer
+// N64 decal depth mode: pixels only pass when close to the stored depth, using a copy of the depth buffer
 #define DECAL_DEPTH_TEXTURE_UNIT 2
+// Height of the N64 frame; the RDP's depth slope tolerance is measured in N64 pixels
+#define DECAL_N64_SCREEN_HEIGHT 240.0f
 static bool depth_copy_supported;
 static bool depth_copy_dirty = true;
 static size_t depth_copy_source;
@@ -399,6 +403,7 @@ static struct ShaderProgram* gfx_opengl_create_and_load_new_shader(uint64_t shad
     append_line(fs_buf, &fs_len, "uniform int frame_count;");
     append_line(fs_buf, &fs_len, "uniform float noise_scale;");
     append_line(fs_buf, &fs_len, "uniform int uDecalMode;");
+    append_line(fs_buf, &fs_len, "uniform highp float uDecalPixelScale;");
     append_line(fs_buf, &fs_len, "uniform highp sampler2D uDepthTex;");
 
     append_line(fs_buf, &fs_len, "float random(in vec3 value) {");
@@ -460,12 +465,13 @@ static struct ShaderProgram* gfx_opengl_create_and_load_new_shader(uint64_t shad
 
     append_line(fs_buf, &fs_len, "void main() {");
 
-    // Discard decal pixels hanging past an edge, i.e. where the stored surface is clearly further away than the decal.
-    // (1 - z) is roughly proportional to 1 / distance, so this allows a gap of about 3% of the distance, plus slope.
+    // RDP decal Z mode: pass only if |z - stored z| <= dz (the per-pixel depth slope), and never over cleared depth.
+    // The slope is computed before anything can discard, since derivatives are undefined after a neighbor discards.
     append_line(fs_buf, &fs_len, "    highp float decalSlope = fwidth(gl_FragCoord.z);");
     append_line(fs_buf, &fs_len, "    if (uDecalMode != 0) {");
     append_line(fs_buf, &fs_len, "        highp float storedZ = texelFetch(uDepthTex, ivec2(gl_FragCoord.xy), 0).r;");
-    append_line(fs_buf, &fs_len, "        if (storedZ - gl_FragCoord.z > 0.03 * (1.0 - gl_FragCoord.z) + 3.0 * decalSlope) discard;");
+    append_line(fs_buf, &fs_len, "        highp float dz = max(decalSlope * uDecalPixelScale, 1.0 / 262144.0);");
+    append_line(fs_buf, &fs_len, "        if (storedZ >= 1.0 || abs(gl_FragCoord.z - storedZ) > dz) discard;");
     append_line(fs_buf, &fs_len, "    }");
 
     for (int i = 0; i < 2; i++) {
@@ -679,7 +685,9 @@ static struct ShaderProgram* gfx_opengl_create_and_load_new_shader(uint64_t shad
     prg->three_point_filter_locations[0] = glGetUniformLocation(shader_program, "three_point_filter0");
     prg->three_point_filter_locations[1] = glGetUniformLocation(shader_program, "three_point_filter1");
     prg->decal_mode_location = glGetUniformLocation(shader_program, "uDecalMode");
+    prg->decal_scale_location = glGetUniformLocation(shader_program, "uDecalPixelScale");
     prg->decal_mode = 0;
+    prg->decal_scale = 0.0f;
     if (prg->decal_mode_location >= 0) {
         glUniform1i(prg->decal_mode_location, 0);
         glUniform1i(glGetUniformLocation(shader_program, "uDepthTex"), DECAL_DEPTH_TEXTURE_UNIT);
@@ -926,11 +934,29 @@ static void gfx_opengl_set_decal_uniforms(struct ShaderProgram* prg, bool enable
         glUniform1i(prg->decal_mode_location, mode);
         prg->decal_mode = mode;
     }
+
+    const float scale = framebuffers[current_framebuffer].height / DECAL_N64_SCREEN_HEIGHT;
+    if (enabled && prg->decal_scale != scale) {
+        glUniform1f(prg->decal_scale_location, scale);
+        prg->decal_scale = scale;
+    }
 }
 
 static void gfx_opengl_draw_triangles(float buf_vbo[], size_t buf_vbo_len, size_t buf_vbo_num_tris) {
     // printf("flushing %d tris\n", buf_vbo_num_tris);
-    gfx_opengl_set_decal_uniforms(current_shader_program, current_depth_decal && gfx_opengl_update_depth_copy());
+    if (current_depth_decal) {
+        // The shader replaces the depth test like the RDP does; fall back to polygon offset if depth can't be copied
+        const bool decal = gfx_opengl_update_depth_copy();
+        glDepthFunc(decal ? GL_ALWAYS : GL_LEQUAL);
+        if (decal) {
+            glDisable(GL_POLYGON_OFFSET_FILL);
+        } else {
+            glEnable(GL_POLYGON_OFFSET_FILL);
+        }
+        gfx_opengl_set_decal_uniforms(current_shader_program, decal);
+    } else {
+        gfx_opengl_set_decal_uniforms(current_shader_program, false);
+    }
 
     glBufferData(GL_ARRAY_BUFFER, sizeof(float) * buf_vbo_len, buf_vbo, GL_STREAM_DRAW);
     glDrawArrays(GL_TRIANGLES, 0, 3 * buf_vbo_num_tris);