From 9fa6f1f2610ca847b39406d70a7ac81a1c892d0e Mon Sep 17 00:00:00 2001
From: pineappleEA <pineaea@gmail.com>
Date: Mon, 17 Jan 2022 00:57:42 +0100
Subject: [PATCH] early-access version 2404

---
 README.md                                     |  2 +-
 src/CMakeLists.txt                            |  5 +-
 src/video_core/host_shaders/astc_decoder.comp | 80 +++++++++++--------
 3 files changed, 49 insertions(+), 38 deletions(-)

diff --git a/README.md b/README.md
index ca4098df6..2447f57f4 100755
--- a/README.md
+++ b/README.md
@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 2402.
+This is the source code for early-access 2404.
 
 ## Legal Notice
 
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 19f7ba339..4ee98beba 100755
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -23,7 +23,6 @@ if (MSVC)
 
     # /W3                 - Level 3 warnings
     # /MP                 - Multi-threaded compilation
-    # /Zf                 - Improves PDB generation time in parallel builds
     # /Zi                 - Output debugging information
     # /Zm                 - Specifies the precompiled header memory allocation limit
     # /Zo                 - Enhanced debug info for optimized builds
@@ -37,9 +36,9 @@ if (MSVC)
     # /GT                 - Supports fiber safety for data allocated using static thread-local storage
     add_compile_options(
         /MP
-        /Zf
         /Zi
-        /Zm400
+        /Zm300
+        /Zf
         /Zo
         /permissive-
         /EHsc
diff --git a/src/video_core/host_shaders/astc_decoder.comp b/src/video_core/host_shaders/astc_decoder.comp
index f34c5f5d9..3a10578cb 100755
--- a/src/video_core/host_shaders/astc_decoder.comp
+++ b/src/video_core/host_shaders/astc_decoder.comp
@@ -155,9 +155,6 @@ uint SwizzleOffset(uvec2 pos) {
 // Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
 // is the same as [(num_bits - 1):0] and repeats all the way down.
 uint Replicate(uint val, uint num_bits, uint to_bit) {
-    if (num_bits == 0 || to_bit == 0) {
-        return 0;
-    }
     const uint v = val & uint((1 << num_bits) - 1);
     uint res = v;
     uint reslen = num_bits;
@@ -187,42 +184,57 @@ uint ReplicateBitTo9(uint value) {
     return REPLICATE_1_BIT_TO_9_TABLE[value];
 }
 
-uint FastReplicateTo8(uint value, uint num_bits) {
-    switch (num_bits) {
-    case 1:
-        return REPLICATE_1_BIT_TO_8_TABLE[value];
-    case 2:
-        return REPLICATE_2_BIT_TO_8_TABLE[value];
-    case 3:
-        return REPLICATE_3_BIT_TO_8_TABLE[value];
-    case 4:
-        return REPLICATE_4_BIT_TO_8_TABLE[value];
-    case 5:
-        return REPLICATE_5_BIT_TO_8_TABLE[value];
-    case 6:
-        return REPLICATE_6_BIT_TO_8_TABLE[value];
-    case 7:
-        return REPLICATE_7_BIT_TO_8_TABLE[value];
-    case 8:
+uint FastReplicate(uint value, uint num_bits, uint to_bit) {
+    if (num_bits == 0) {
+        return 0;
+    }
+    if (num_bits == to_bit) {
         return value;
     }
-    return Replicate(value, num_bits, 8);
+    if (to_bit == 6) {
+        switch (num_bits) {
+        case 1:
+            return REPLICATE_1_BIT_TO_6_TABLE[value];
+        case 2:
+            return REPLICATE_2_BIT_TO_6_TABLE[value];
+        case 3:
+            return REPLICATE_3_BIT_TO_6_TABLE[value];
+        case 4:
+            return REPLICATE_4_BIT_TO_6_TABLE[value];
+        case 5:
+            return REPLICATE_5_BIT_TO_6_TABLE[value];
+        default:
+            break;
+        }
+    } else { /* if (to_bit == 8) */
+        switch (num_bits) {
+        case 1:
+            return REPLICATE_1_BIT_TO_8_TABLE[value];
+        case 2:
+            return REPLICATE_2_BIT_TO_8_TABLE[value];
+        case 3:
+            return REPLICATE_3_BIT_TO_8_TABLE[value];
+        case 4:
+            return REPLICATE_4_BIT_TO_8_TABLE[value];
+        case 5:
+            return REPLICATE_5_BIT_TO_8_TABLE[value];
+        case 6:
+            return REPLICATE_6_BIT_TO_8_TABLE[value];
+        case 7:
+            return REPLICATE_7_BIT_TO_8_TABLE[value];
+        default:
+            break;
+        }
+    }
+    return Replicate(value, num_bits, to_bit);
+}
+
+uint FastReplicateTo8(uint value, uint num_bits) {
+    return FastReplicate(value, num_bits, 8);
 }
 
 uint FastReplicateTo6(uint value, uint num_bits) {
-    switch (num_bits) {
-    case 1:
-        return REPLICATE_1_BIT_TO_6_TABLE[value];
-    case 2:
-        return REPLICATE_2_BIT_TO_6_TABLE[value];
-    case 3:
-        return REPLICATE_3_BIT_TO_6_TABLE[value];
-    case 4:
-        return REPLICATE_4_BIT_TO_6_TABLE[value];
-    case 5:
-        return REPLICATE_5_BIT_TO_6_TABLE[value];
-    }
-    return Replicate(value, num_bits, 6);
+    return FastReplicate(value, num_bits, 6);
 }
 
 uint Div3Floor(uint v) {