Populate the zbuffer for broader ReShade compatability

src/core/rdp.h

@@ -68,6 +68,7 @@ struct rdp_config
 struct rdp_frame_buffer
 {
     uint32_t* pixels;
+    uint32_t* depth;
     uint32_t width;
     uint32_t height;
     uint32_t pitch;

src/core/rdp/vi.c

@@ -84,6 +84,7 @@ static uint32_t tvfadeoutstate[PRESCALE_HEIGHT];
 
 // prescale buffer
 static uint32_t prescale[PRESCALE_WIDTH * PRESCALE_HEIGHT];
+static uint32_t prescale_depth[PRESCALE_WIDTH * PRESCALE_HEIGHT];
 static uint32_t prescale_ptr;
 static int32_t linecount;
 
@@ -115,6 +116,7 @@ static void vi_init(void)
     vi_restore_init();
 
     memset(prescale, 0, sizeof(prescale));
+    memset(prescale_depth, 0, sizeof(prescale_depth));
 
     prevvicurrent = 0;
     emucontrolsvicurrent = -1;
@@ -527,6 +529,7 @@ static void vi_process_end(void)
     struct rdp_frame_buffer fb;
     fb.pixels = prescale;
     fb.pitch = PRESCALE_WIDTH;
+    fb.depth = prescale_depth;
 
     int32_t output_height;
 
@@ -596,12 +599,14 @@ static void vi_process_fast(uint32_t worker_id)
         int32_t x;
         int32_t line = y * vi_width_low;
         uint32_t* dst = prescale + y * hres_raw;
+        uint32_t* d = prescale_depth + y * hres_raw;
 
         for (x = 0; x < hres_raw; x++) {
-            uint32_t r, g, b;
+            uint32_t r, g, b, zr, zg, zb;
 
             switch (config.vi.mode) {
                 case VI_MODE_COLOR:
+                    zr = zg = zb = rdram_read_idx16((rdp_states[worker_id].zb_address >> 1) + line + x) >> 8;
                     switch (ctrl.type) {
                         case VI_TYPE_RGBA5551: {
                             uint16_t pix = rdram_read_idx16((frame_buffer >> 1) + line + x);
@@ -625,7 +630,7 @@ static void vi_process_fast(uint32_t worker_id)
                     break;
 
                 case VI_MODE_DEPTH: {
-                    r = g = b = rdram_read_idx16((rdp_states[0].zb_address >> 1) + line + x) >> 8;
+                    r = g = b = zr = zg = zb = rdram_read_idx16((rdp_states[0].zb_address >> 1) + line + x) >> 8;
                     break;
                 }
 
@@ -635,6 +640,7 @@ static void vi_process_fast(uint32_t worker_id)
                     uint16_t pix;
                     rdram_read_pair16(&pix, &hval, (frame_buffer >> 1) + line + x);
                     r = g = b = (((pix & 1) << 2) | hval) << 5;
+                    zr = zg = zb = rdram_read_idx16((rdp_states[0].zb_address >> 1) + line + x) >> 8;
                     break;
                 }
 
@@ -643,8 +649,10 @@ static void vi_process_fast(uint32_t worker_id)
             }
 
             gamma_filters(&r, &g, &b, ctrl, &rdp_states[worker_id].rand_vi);
+            gamma_filters(&zr, &zg, &zb, ctrl, &rdp_states[worker_id].rand_vi);
 
             dst[x] = (b << 16) | (g << 8) | r;
+            d[x] = (zb << 16) | (zg << 8) | zr;
         }
     }
 }
@@ -656,6 +664,7 @@ static void vi_process_end_fast(void)
     fb.width = hres_raw;
     fb.height = vres_raw;
     fb.pitch = hres_raw;
+    fb.depth = prescale_depth;
 
     int32_t filtered_height = (vres << 1) * V_SYNC_NTSC / v_sync;
     int32_t output_height = hres_raw * filtered_height / hres;

src/plugin/common/gl_screen.c

@@ -18,7 +18,11 @@
 
 static GLuint program;
 static GLuint vao;
-static GLuint texture;
+static GLuint texture = 1;
+static GLuint depth_texture = 2;
+
+static GLint colorValueTextureLocation;
+static GLint depthValueTextureLocation;
 
 static int32_t tex_width;
 static int32_t tex_height;
@@ -136,39 +140,63 @@ void gl_screen_init(struct rdp_config* config)
         SHADER_HEADER
         "in vec2 uv;\n"
         "layout(location = 0) out vec4 color;\n"
-        "uniform sampler2D tex0;\n"
+
+        "uniform sampler2D ColorValueTexture;\n"
+        "uniform sampler2D DepthValueTexture;\n"
+
         "void main(void) {\n"
 #ifdef GLES
-        "    color = texture(tex0, uv);\n"
+        "    color = texture(ColorValueTexture, uv);\n"
 #else
-        "    color.bgra = texture(tex0, uv);\n"
+        "    color.bgra = texture(ColorValueTexture, uv);\n"
 #endif
+        "    gl_FragDepth = texture(DepthValueTexture, uv).r;\n"
         "}\n";
 
     // compile and link OpenGL program
     GLuint vert = gl_shader_compile(GL_VERTEX_SHADER, vert_shader);
     GLuint frag = gl_shader_compile(GL_FRAGMENT_SHADER, frag_shader);
     program = gl_shader_link(vert, frag);
+
+    // get the uniform variables location
+    depthValueTextureLocation = glGetUniformLocation(program, "DepthValueTexture");
+    colorValueTextureLocation = glGetUniformLocation(program, "ColorValueTexture");
+
+    // specify the shader program to use
     glUseProgram(program);
 
+    // bind the uniform variables locations
+    glUniform1i(depthValueTextureLocation, 0);
+    glUniform1i(colorValueTextureLocation, 1);
+
     // prepare dummy VAO
     glGenVertexArrays(1, &vao);
     glBindVertexArray(vao);
 
-    // prepare texture
-    glGenTextures(1, &texture);
-    glBindTexture(GL_TEXTURE_2D, texture);
-
     // select interpolation method
     GLint filter;
     switch (config->vi.interp) {
-        case VI_INTERP_LINEAR:
-            filter = GL_LINEAR;
-            break;
-        case VI_INTERP_NEAREST:
-        default:
-            filter = GL_NEAREST;
+    case VI_INTERP_LINEAR:
+        filter = GL_LINEAR;
+        break;
+    case VI_INTERP_NEAREST:
+    default:
+        filter = GL_NEAREST;
     }
+
+    // prepare depth texture
+    glActiveTexture(GL_TEXTURE0 + 0);
+    glGenTextures(1, &depth_texture);
+    glBindTexture(GL_TEXTURE_2D, depth_texture);
+    // configure interpolation method
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, filter);
+
+    // prepare color texture
+    glActiveTexture(GL_TEXTURE0 + 1);
+    glGenTextures(1, &texture);
+    glBindTexture(GL_TEXTURE_2D, texture);
+    // configure interpolation method
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, filter);
 
@@ -185,18 +213,51 @@ bool gl_screen_write(struct rdp_frame_buffer* fb, int32_t output_height)
         tex_width = fb->width;
         tex_height = fb->height;
 
+        // select the color value binding
+        glActiveTexture(GL_TEXTURE0 + 1);
+        glBindTexture(GL_TEXTURE_2D, texture);
+
         // set pitch for all unpacking operations
         glPixelStorei(GL_UNPACK_ROW_LENGTH, fb->pitch);
+        // reallocate texture buffer on GPU
+        glDepthMask(false);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, tex_width, tex_height, 0, TEX_FORMAT, TEX_TYPE, fb->pixels);
+        glDepthMask(true);
 
+        // select the depth value binding
+        glActiveTexture(GL_TEXTURE0 + 0);
+        glBindTexture(GL_TEXTURE_2D, depth_texture);
+
+        // set pitch for all unpacking operations
+        glPixelStorei(GL_UNPACK_ROW_LENGTH, fb->pitch);
         // reallocate texture buffer on GPU
-        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, tex_width,
-            tex_height, 0, TEX_FORMAT, TEX_TYPE, fb->pixels);
+        glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+        glEnable(GL_DEPTH_TEST);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, tex_width, tex_height, 0, TEX_FORMAT, TEX_TYPE, fb->depth);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glDisable(GL_DEPTH_TEST);
 
         msg_debug("%s: resized framebuffer texture: %dx%d", __FUNCTION__, tex_width, tex_height);
     } else {
+        // select the color value binding
+        glActiveTexture(GL_TEXTURE0 + 1);
+        glBindTexture(GL_TEXTURE_2D, texture);
+
+        // copy local buffer to GPU texture buffer
+        glDepthMask(false);
+        glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, tex_width, tex_height, TEX_FORMAT, TEX_TYPE, fb->pixels);
+        glDepthMask(true);
+
+        // select the depth value binding
+        glActiveTexture(GL_TEXTURE0 + 0);
+        glBindTexture(GL_TEXTURE_2D, depth_texture);
+
         // copy local buffer to GPU texture buffer
-        glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, tex_width, tex_height,
-            TEX_FORMAT, TEX_TYPE, fb->pixels);
+        glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
+        glEnable(GL_DEPTH_TEST);
+        glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, tex_width, tex_height, TEX_FORMAT, TEX_TYPE, fb->depth);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glDisable(GL_DEPTH_TEST);
     }
 
     // update output size
@@ -239,8 +300,21 @@ void gl_screen_render(int32_t win_width, int32_t win_height, int32_t win_x, int3
     // configure viewport
     glViewport(win_x, win_y, win_width, win_height);
 
-    // draw fullscreen triangle
+    // select the color value binding
+    glActiveTexture(GL_TEXTURE0 + 1);
+    glBindTexture(GL_TEXTURE_2D, texture);
+
+    // draw
+    glDrawArrays(GL_TRIANGLES, 0, 3);
+
+    // select the depth value binding
+    glActiveTexture(GL_TEXTURE0 + 0);
+    glBindTexture(GL_TEXTURE_2D, depth_texture);
+
+    // draw
+    glEnable(GL_DEPTH_TEST);
     glDrawArrays(GL_TRIANGLES, 0, 3);
+    glDisable(GL_DEPTH_TEST);
 
     // check if there was an error when using any of the commands above
     gl_check_errors();