port: better menu blur

TODO: make it render the blurred texture to the fb once instead of rendering it blurred every frame

port/fast3d/gfx_cc.cpp

@@ -17,6 +17,7 @@ void gfx_cc_get_features(uint64_t shader_id0, uint32_t shader_id1, struct CCFeat
     cc_features->opt_alpha_threshold = (shader_id1 & SHADER_OPT_ALPHA_THRESHOLD) != 0;
     cc_features->opt_invisible = (shader_id1 & SHADER_OPT_INVISIBLE) != 0;
     cc_features->opt_grayscale = (shader_id1 & SHADER_OPT_GRAYSCALE) != 0;
+    cc_features->opt_blur = (shader_id1 & SHADER_OPT_BLUR) != 0;
 
     cc_features->clamp[0][0] = (shader_id1 & SHADER_OPT_TEXEL0_CLAMP_S);
     cc_features->clamp[0][1] = (shader_id1 & SHADER_OPT_TEXEL0_CLAMP_T);

port/fast3d/gfx_cc.h

@@ -53,6 +53,7 @@ enum {
 #define SHADER_OPT_TEXEL1_MASK (1 << 13)
 #define SHADER_OPT_TEXEL0_BLEND (1 << 14)
 #define SHADER_OPT_TEXEL1_BLEND (1 << 15)
+#define SHADER_OPT_BLUR (1 << 16)
 
 struct ColorCombinerKey {
     uint64_t combine_mode;
@@ -75,6 +76,7 @@ struct CCFeatures {
     bool opt_alpha_threshold;
     bool opt_invisible;
     bool opt_grayscale;
+    bool opt_blur;
     bool used_textures[2];
     bool used_masks[2];
     bool used_blend[2];

port/fast3d/gfx_opengl.cpp

@@ -394,7 +394,30 @@ static struct ShaderProgram* gfx_opengl_create_and_load_new_shader(uint64_t shad
     append_line(fs_buf, &fs_len, "    return fract(sin(random) * 143758.5453);");
     append_line(fs_buf, &fs_len, "}");
 
-    if (current_filter_mode == FILTER_THREE_POINT) {
+    if (cc_features.opt_blur) {
+        // blur filter, used for menu backgrounds
+        append_line(fs_buf, &fs_len, R"(
+            const vec4 blurOffsets = vec4(-1.5, -0.5, +0.5, +1.5);
+            const mat4 blurWeights = mat4(
+                0.009947, 0.009641, 0.008778, 0.007509,
+                0.009641, 0.009345, 0.008508, 0.007278,
+                0.008778, 0.008508, 0.007747, 0.006626,
+                0.007509, 0.007278, 0.006626, 0.005668
+            );
+            lowp vec4 hookTexture2D(in sampler2D tex, in vec2 texCoord, in vec2 texSize) {
+                lowp vec4 color = vec4(0.0);
+                lowp float wacc = 0.0;
+                for (int y = 0; y < 4; ++y) {
+                    for (int x = 0; x < 4; ++x) {
+                        vec2 ofs = vec2(blurOffsets[x], blurOffsets[y]) / texSize;
+                        wacc += blurWeights[x][y];
+                        color += texture2D(tex, texCoord + ofs) * blurWeights[x][y];
+                    }
+                }
+                return vec4(color.rgb / wacc, 1.0);
+            })"
+        );
+    } else if (current_filter_mode == FILTER_THREE_POINT) {
 #if __APPLE__
         append_line(fs_buf, &fs_len, "#define TEX_OFFSET(off) texture(tex, texCoord - (off)/texSize)");
 #else

port/fast3d/gfx_pc.cpp

@@ -1288,6 +1288,7 @@ static void gfx_sp_tri1(uint8_t vtx1_idx, uint8_t vtx2_idx, uint8_t vtx3_idx, bo
         (rdp.other_mode_l & (3 << 24)) == (G_BL_0 << 24) && (rdp.other_mode_l & (3 << 20)) == (G_BL_CLR_MEM << 20);
     bool use_grayscale = rdp.grayscale;
     bool use_modulate = use_alpha && (rsp.extra_geometry_mode & G_MODULATE_EXT) != 0;
+    bool use_blur = (rdp.other_mode_h & (3U << G_MDSFT_TEXTFILT)) == G_TF_BLUR_EXT;
 
     if (texture_edge) {
         use_alpha = true;
@@ -1317,6 +1318,9 @@ static void gfx_sp_tri1(uint8_t vtx1_idx, uint8_t vtx2_idx, uint8_t vtx3_idx, bo
     if (use_grayscale) {
         cc_options |= (uint64_t)SHADER_OPT_GRAYSCALE;
     }
+    if (use_blur) {
+        cc_options |= (uint64_t)SHADER_OPT_BLUR;
+    }
 
     // If we are not using alpha, clear the alpha components of the combiner as they have no effect
     if (!use_alpha) {

src/game/menu.c

@@ -50,6 +50,9 @@
 #include "types.h"
 #ifndef PLATFORM_N64
 #include "video.h"
+#define BLUR_OFS 10
+#else
+#define BLUR_OFS 30
 #endif
 
 #if VERSION >= VERSION_PAL_FINAL
@@ -5084,23 +5087,9 @@ Gfx *menuRenderBackgroundLayer1(Gfx *gdl, u8 bg, f32 frac)
 			// Render the blurred background texture with full alpha
 			gdl = menugfxRenderBgBlur(gdl, 0xffffff00 | alpha, 0, 0);
 
-#ifdef PLATFORM_N64
 			// Render it twice more with half alpha and offset
-			gdl = menugfxRenderBgBlur(gdl, 0xffffff00 | alpha >> 1, -30, -30);
-			gdl = menugfxRenderBgBlur(gdl, 0xffffff00 | alpha >> 1, 30, 30);
-#else
-			// HACK: we don't actually blur anything on the CPU, so
-			// render it offset to all cardinal and ordinal directions to achieve crappy box blur
-			const s32 card[2] = { -30, 30 };
-			const s32 ord[2] = { -21, 21 }; // 30 / sqrt(2)
-			const u32 rgba = 0xffffff00 | alpha >> 3;
-			for (s32 i = 0; i < 2; ++i) {
-				gdl = menugfxRenderBgBlur(gdl, rgba, 0, card[i]);
-				gdl = menugfxRenderBgBlur(gdl, rgba, card[i], 0);
-				gdl = menugfxRenderBgBlur(gdl, rgba, ord[0], ord[i]);
-				gdl = menugfxRenderBgBlur(gdl, rgba, ord[i], ord[0]);
-			}
-#endif
+			gdl = menugfxRenderBgBlur(gdl, 0xffffff00 | alpha >> 1, -BLUR_OFS, -BLUR_OFS);
+			gdl = menugfxRenderBgBlur(gdl, 0xffffff00 | alpha >> 1, BLUR_OFS, BLUR_OFS);
 		}
 		break;
 	case MENUBG_BLACK:

src/game/menugfx.c

@@ -164,7 +164,13 @@ Gfx *menugfxRenderBgBlur(Gfx *gdl, u32 colour, s16 arg2, s16 arg3)
 	gDPSetAlphaCompare(gdl++, G_AC_NONE);
 	gDPSetCombineMode(gdl++, G_CC_MODULATEI, G_CC_MODULATEI);
 	gSPClearGeometryMode(gdl++, G_CULL_BOTH);
+
+#ifdef PLATFORM_N64
 	gDPSetTextureFilter(gdl++, G_TF_BILERP);
+#else
+	// enable blur filter for the texture
+	gDPSetTextureFilter(gdl++, G_TF_BLUR_EXT);
+#endif
 
 	gDPSetRenderMode(gdl++, G_RM_XLU_SURF, G_RM_XLU_SURF2);
 
@@ -240,6 +246,10 @@ Gfx *menugfxRenderBgBlur(Gfx *gdl, u32 colour, s16 arg2, s16 arg3)
 
 	gSPTri2(gdl++, 0, 1, 2, 2, 3, 0);
 
+#ifndef PLATFORM_N64
+	gDPSetTextureFilter(gdl++, G_TF_BILERP);
+#endif
+
 	return gdl;
 }
 

src/include/gbiex.h

@@ -207,6 +207,10 @@
 #define G_NO_CLIPPING_EXT        0x00000100
 #define G_MODULATE_EXT           0x00000200 // this should really go into OTHERMODE_H, but for some reason I can't get it to work
 
+/* Extra texture filtering mode */
+
+#define G_TF_BLUR_EXT (1 << G_MDSFT_TEXTFILT)
+
 /* Extended command macros */
 
 #define gDPSetFramebufferTargetEXT(pkt, f, s, w, i) \