bullshit3

src/video_core/host_shaders/CMakeLists.txt

@@ -90,7 +90,7 @@ if ("${GLSLANGVALIDATOR}" STREQUAL "GLSLANGVALIDATOR-NOTFOUND")
     message(FATAL_ERROR "Required program `glslangValidator` not found.")
 endif()
 
-set(GLSL_FLAGS "")
+set(GLSL_FLAGS "-Os")
 set(SPIR_V_VERSION "spirv1.3")
 set(QUIET_FLAG "--quiet")
 

src/video_core/host_shaders/astc_decoder.comp

@@ -784,34 +784,30 @@ bool IsError(uint mode) {
 }
 
 int FindLayout(uint mode) {
-    if ((mode & 3) != 0) {
-        if ((mode & 8) != 0) {
-            if ((mode & 4) != 0) {
-                if ((mode & 0x100) != 0) {
-                    return 4;
-                }
-                return 3;
-            }
-            return 2;
-        }
-        if ((mode & 4) != 0) {
-            return 1;
-        }
-        return 0;
-    }
-    if ((mode & 0x100) != 0) {
-        if ((mode & 0x80) != 0) {
-            if ((mode & 0x20) != 0) {
-                return 8;
-            }
-            return 7;
-        }
-        return 9;
-    }
-    if ((mode & 0x80) != 0) {
-        return 6;
-    }
-    return 5;
+    // Possible (Relevant), x = dont care, (in hex)
+    // Before shift: 02x, 08x, 10x, 00x, 0Ax, 12x, 18x, 1Ax
+    // After shift:   01,  04,  08,  00,  05,  09,  0c,  0d
+    // Reassemble and remove middle 0x040 (stupid fucking table)
+    // 0000 -> 0    0020 -> 1   0040 -> 0   0060 -> 1
+    // 0080 -> 2    00a0 -> 3   00c0 -> 2   00e0 -> 3
+    // 0100 -> 4    0120 -> 5   0140 -> 4   0160 -> 5
+    // 0180 -> 6    01a0 -> 7   01c0 -> 6   01e0 -> 7
+    // Key ranges: 01a0 -> 7, 0180 -> 6, 0100 -> 4, 0080 -> 2
+    // 8>>2 = 2, 4>>2 = 1 :: () = 0, (4) = 1, (8) = 2, (8,4) = 3
+    const uint mask = 0 - uint((mode & 3) != 0);
+    const uint sh3_mode = (mode >> 2) & 3;
+    const uint sh0_mode = ((mode >> 6) & 6) | ((mode >> 5) & 1);
+    const uint fl_const_table = 0
+        | ((1) << (2 * 3)) //0080 -> 2, 1 + 5 = 6
+        | ((1) << (3 * 3))
+        | ((4) << (4 * 3)) //0100 -> 4, 4 + 5 = 9
+        | ((4) << (5 * 3))
+        | ((2) << (6 * 3)) //0180 -> 6, 2 + 5 = 7
+        | ((3) << (7 * 3)) //01a0 -> 7, 3 + 5 = 8
+    ;
+    const uint if_mode3_t = sh3_mode + uint((mode & 0x10c) == 0x10c);
+    const uint if_mode3_f = 5 + ((fl_const_table >> (sh0_mode * 3)) & 7);
+    return int((if_mode3_t & mask) | (if_mode3_f & ~mask));
 }
 
 uvec2 DecodeBlockSize(uint mode_layout, uint mode) {