From a5bff8e9b3eed5419bbb38abaaa642ed11b235db Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Sun, 16 Jan 2022 15:52:34 -0500
Subject: astc_decoder: Combine FastReplicate functions to work around new NV
 driver bug

The new Nvidia drivers have a bug where the FastReplicateTo6 function produces a lookup into the REPLICATE_TO_8 table rather than the REPLICATE_TO_6 table.

This seems to be an optimization gone wrong. Combining the logic of the FastReplicate functions seems to address the bug.
---
 src/video_core/host_shaders/astc_decoder.comp | 80 +++++++++++++++------------
 1 file changed, 46 insertions(+), 34 deletions(-)

(limited to 'src')

diff --git a/src/video_core/host_shaders/astc_decoder.comp b/src/video_core/host_shaders/astc_decoder.comp
index f34c5f5d9..3a10578cb 100644
--- 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) {
-- 
cgit v1.2.3