From 5ab80535118e593ef3add3ce2b5935437e1dc1d3 Mon Sep 17 00:00:00 2001
From: ameerj <52414509+ameerj@users.noreply.github.com>
Date: Sat, 31 Jul 2021 22:24:15 -0400
Subject: astc_decoder: Compute offset swizzles in-shader

Alleviates the dependency on the swizzle table and a uniform which is constant for all ASTC texture sizes.
---
 src/video_core/host_shaders/astc_decoder.comp | 46 ++++++++-------------------
 1 file changed, 13 insertions(+), 33 deletions(-)

(limited to 'src/video_core/host_shaders')

diff --git a/src/video_core/host_shaders/astc_decoder.comp b/src/video_core/host_shaders/astc_decoder.comp
index 392f09c68..74ce058a9 100644
--- a/src/video_core/host_shaders/astc_decoder.comp
+++ b/src/video_core/host_shaders/astc_decoder.comp
@@ -10,8 +10,7 @@
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)
 #define BINDING_INPUT_BUFFER 0
-#define BINDING_SWIZZLE_BUFFER 1
-#define BINDING_OUTPUT_IMAGE 2
+#define BINDING_OUTPUT_IMAGE 1
 
 #else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
 
@@ -19,7 +18,6 @@
 #define END_PUSH_CONSTANTS
 #define UNIFORM(n) layout(location = n) uniform
 #define BINDING_INPUT_BUFFER 0
-#define BINDING_SWIZZLE_BUFFER 1
 #define BINDING_OUTPUT_IMAGE 0
 
 #endif
@@ -28,13 +26,11 @@ layout(local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
 
 BEGIN_PUSH_CONSTANTS
 UNIFORM(1) uvec2 block_dims;
-
-UNIFORM(2) uint bytes_per_block_log2;
-UNIFORM(3) uint layer_stride;
-UNIFORM(4) uint block_size;
-UNIFORM(5) uint x_shift;
-UNIFORM(6) uint block_height;
-UNIFORM(7) uint block_height_mask;
+UNIFORM(2) uint layer_stride;
+UNIFORM(3) uint block_size;
+UNIFORM(4) uint x_shift;
+UNIFORM(5) uint block_height;
+UNIFORM(6) uint block_height_mask;
 END_PUSH_CONSTANTS
 
 struct EncodingData {
@@ -53,35 +49,17 @@ struct TexelWeightParams {
     bool void_extent_hdr;
 };
 
-// Swizzle data
-layout(binding = BINDING_SWIZZLE_BUFFER, std430) readonly buffer SwizzleTable {
-    uint swizzle_table[];
-};
-
 layout(binding = BINDING_INPUT_BUFFER, std430) readonly buffer InputBufferU32 {
     uvec4 astc_data[];
 };
 
 layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2DArray dest_image;
 
-const uint GOB_SIZE_X = 64;
-const uint GOB_SIZE_Y = 8;
-const uint GOB_SIZE_Z = 1;
-const uint GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;
-
 const uint GOB_SIZE_X_SHIFT = 6;
 const uint GOB_SIZE_Y_SHIFT = 3;
-const uint GOB_SIZE_Z_SHIFT = 0;
-const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;
-
-const uvec2 SWIZZLE_MASK = uvec2(GOB_SIZE_X - 1, GOB_SIZE_Y - 1);
-
-const int BLOCK_SIZE_IN_BYTES = 16;
+const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT;
 
-const int BLOCK_INFO_ERROR = 0;
-const int BLOCK_INFO_VOID_EXTENT_HDR = 1;
-const int BLOCK_INFO_VOID_EXTENT_LDR = 2;
-const int BLOCK_INFO_NORMAL = 3;
+const uint BYTES_PER_BLOCK_LOG2 = 4;
 
 const int JUST_BITS = 0;
 const int QUINT = 1;
@@ -168,8 +146,10 @@ int texel_vector_index = 0;
 uint unquantized_texel_weights[2][144];
 
 uint SwizzleOffset(uvec2 pos) {
-    pos = pos & SWIZZLE_MASK;
-    return swizzle_table[pos.y * 64 + pos.x];
+    uint x = pos.x;
+    uint y = pos.y;
+    return ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
+                          (y % 2) * 16 + (x % 16);
 }
 
 // Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
@@ -1253,7 +1233,7 @@ void DecompressBlock(ivec3 coord) {
 
 void main() {
     uvec3 pos = gl_GlobalInvocationID;
-    pos.x <<= bytes_per_block_log2;
+    pos.x <<= BYTES_PER_BLOCK_LOG2;
 
     // Read as soon as possible due to its latency
     const uint swizzle = SwizzleOffset(pos.xy);
-- 
cgit v1.2.3