texture_cache: 32bpp and 64bpp macro detilers (#1852)

* added 32bpp macro detiler

* added 64bpp macro detiler

* consider 3d depth alignment in size calculations

src/video_core/host_shaders/CMakeLists.txt

@@ -7,6 +7,8 @@ set(SHADER_FILES
     detile_m32x1.comp
     detile_m32x2.comp
     detile_m32x4.comp
+    detile_macro32x1.comp
+    detile_macro32x2.comp
     fs_tri.vert
     post_process.frag
 )

src/video_core/host_shaders/detile_m32x1.comp

@@ -15,6 +15,7 @@ layout(std430, binding = 1) buffer output_buf {
 layout(push_constant) uniform image_info {
     uint num_levels;
     uint pitch;
+    uint height;
     uint sizes[14];
 } info;
 

src/video_core/host_shaders/detile_m32x2.comp

@@ -15,6 +15,7 @@ layout(std430, binding = 1) buffer output_buf {
 layout(push_constant) uniform image_info {
     uint num_levels;
     uint pitch;
+    uint height;
     uint sizes[14];
 } info;
 

src/video_core/host_shaders/detile_m32x4.comp

@@ -15,6 +15,7 @@ layout(std430, binding = 1) buffer output_buf {
 layout(push_constant) uniform image_info {
     uint num_levels;
     uint pitch;
+    uint height;
     uint sizes[14];
 } info;
 

src/video_core/host_shaders/detile_m8x1.comp

@@ -18,6 +18,7 @@ layout(std430, binding = 1) buffer output_buf {
 layout(push_constant) uniform image_info {
     uint num_levels;
     uint pitch;
+    uint height;
     uint sizes[14];
 } info;
 

src/video_core/host_shaders/detile_m8x2.comp

@@ -17,6 +17,7 @@ layout(std430, binding = 1) buffer output_buf {
 layout(push_constant) uniform image_info {
     uint num_levels;
     uint pitch;
+    uint height;
     uint sizes[14];
 } info;
 

src/video_core/host_shaders/detile_macro32x1.comp

@@ -0,0 +1,90 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 450
+
+layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout(std430, binding = 0) buffer input_buf {
+    uint in_data[];
+};
+layout(std430, binding = 1) buffer output_buf {
+    uint out_data[];
+};
+
+layout(push_constant) uniform image_info {
+    uint num_levels;
+    uint pitch;
+    uint height;
+    uint c0;
+    uint c1;
+} info;
+
+// Each LUT is 64 bytes, so should fit into K$ given tiled slices locality
+const uint lut_32bpp[][64] = {
+    {
+        0x00, 0x01, 0x04, 0x05, 0x40, 0x41, 0x44, 0x45,
+        0x02, 0x03, 0x06, 0x07, 0x42, 0x43, 0x46, 0x47,
+        0x10, 0x11, 0x14, 0x15, 0x50, 0x51, 0x54, 0x55,
+        0x12, 0x13, 0x16, 0x17, 0x52, 0x53, 0x56, 0x57,
+        0x80, 0x81, 0x84, 0x85, 0xc0, 0xc1, 0xc4, 0xc5,
+        0x82, 0x83, 0x86, 0x87, 0xc2, 0xc3, 0xc6, 0xc7,
+        0x90, 0x91, 0x94, 0x95, 0xd0, 0xd1, 0xd4, 0xd5,
+        0x92, 0x93, 0x96, 0x97, 0xd2, 0xd3, 0xd6, 0xd7,
+    },
+    {
+        0x08, 0x09, 0x0c, 0x0d, 0x48, 0x49, 0x4c, 0x4d,
+        0x0a, 0x0b, 0x0e, 0x0f, 0x4a, 0x4b, 0x4e, 0x4f,
+        0x18, 0x19, 0x1c, 0x1d, 0x58, 0x59, 0x5c, 0x5d,
+        0x1a, 0x1b, 0x1e, 0x1f, 0x5a, 0x5b, 0x5e, 0x5f,
+        0x88, 0x89, 0x8c, 0x8d, 0xc8, 0xc9, 0xcc, 0xcd,
+        0x8a, 0x8b, 0x8e, 0x8f, 0xca, 0xcb, 0xce, 0xcf,
+        0x98, 0x99, 0x9c, 0x9d, 0xd8, 0xd9, 0xdc, 0xdd,
+        0x9a, 0x9b, 0x9e, 0x9f, 0xda, 0xdb, 0xde, 0xdf,
+    },
+    {
+        0x20, 0x21, 0x24, 0x25, 0x60, 0x61, 0x64, 0x65,
+        0x22, 0x23, 0x26, 0x27, 0x62, 0x63, 0x66, 0x67,
+        0x30, 0x31, 0x34, 0x35, 0x70, 0x71, 0x74, 0x75,
+        0x32, 0x33, 0x36, 0x37, 0x72, 0x73, 0x76, 0x77,
+        0xa0, 0xa1, 0xa4, 0xa5, 0xe0, 0xe1, 0xe4, 0xe5,
+        0xa2, 0xa3, 0xa6, 0xa7, 0xe2, 0xe3, 0xe6, 0xe7,
+        0xb0, 0xb1, 0xb4, 0xb5, 0xf0, 0xf1, 0xf4, 0xf5,
+        0xb2, 0xb3, 0xb6, 0xb7, 0xf2, 0xf3, 0xf6, 0xf7,
+    },
+    {
+        0x28, 0x29, 0x2c, 0x2d, 0x68, 0x69, 0x6c, 0x6d,
+        0x2a, 0x2b, 0x2e, 0x2f, 0x6a, 0x6b, 0x6e, 0x6f,
+        0x38, 0x39, 0x3c, 0x3d, 0x78, 0x79, 0x7c, 0x7d,
+        0x3a, 0x3b, 0x3e, 0x3f, 0x7a, 0x7b, 0x7e, 0x7f,
+        0xa8, 0xa9, 0xac, 0xad, 0xe8, 0xe9, 0xec, 0xed,
+        0xaa, 0xab, 0xae, 0xaf, 0xea, 0xeb, 0xee, 0xef,
+        0xb8, 0xb9, 0xbc, 0xbd, 0xf8, 0xf9, 0xfc, 0xfd,
+        0xba, 0xbb, 0xbe, 0xbf, 0xfa, 0xfb, 0xfe, 0xff,
+    }
+};
+
+#define MICRO_TILE_DIM      (8)
+#define MICRO_TILE_SZ       (1024)
+#define TEXELS_PER_ELEMENT  (1)
+#define BPP                 (32)
+
+void main() {
+    uint x =  gl_GlobalInvocationID.x % info.pitch;
+    uint y = (gl_GlobalInvocationID.x / info.pitch) % info.height;
+    uint z =  gl_GlobalInvocationID.x / (info.pitch * info.height);
+
+    uint col = bitfieldExtract(x, 0, 3);
+    uint row = bitfieldExtract(y, 0, 3);
+    uint lut = bitfieldExtract(z, 0, 2);
+    uint idx = lut_32bpp[lut][col + row * MICRO_TILE_DIM];
+
+    uint slice_offs = (z >> 2u) * info.c1 * MICRO_TILE_SZ; 
+    uint tile_row = y / MICRO_TILE_DIM;
+    uint tile_column = x / MICRO_TILE_DIM;
+    uint tile_offs = ((tile_row * info.c0) + tile_column) * MICRO_TILE_SZ;
+    uint offs = slice_offs + tile_offs + (idx * BPP / 8);
+
+    uint p0 = in_data[offs >> 2u];
+    out_data[gl_GlobalInvocationID.x] = p0;
+}

src/video_core/host_shaders/detile_macro32x2.comp

@@ -0,0 +1,91 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 450
+
+layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout(std430, binding = 0) buffer input_buf {
+    uint in_data[];
+};
+layout(std430, binding = 1) buffer output_buf {
+    uint out_data[];
+};
+
+layout(push_constant) uniform image_info {
+    uint num_levels;
+    uint pitch;
+    uint height;
+    uint c0;
+    uint c1;
+} info;
+
+const uint lut_64bpp[][64] = {
+    {
+        0x00, 0x01, 0x08, 0x09, 0x40, 0x41, 0x48, 0x49,
+        0x02, 0x03, 0x0a, 0x0b, 0x42, 0x43, 0x4a, 0x4b,
+        0x10, 0x11, 0x18, 0x19, 0x50, 0x51, 0x58, 0x59,
+        0x12, 0x13, 0x1a, 0x1b, 0x52, 0x53, 0x5a, 0x5b,
+        0x80, 0x81, 0x88, 0x89, 0xc0, 0xc1, 0xc8, 0xc9,
+        0x82, 0x83, 0x8a, 0x8b, 0xc2, 0xc3, 0xca, 0xcb,
+        0x90, 0x91, 0x98, 0x99, 0xd0, 0xd1, 0xd8, 0xd9,
+        0x92, 0x93, 0x9a, 0x9b, 0xd2, 0xd3, 0xda, 0xdb,
+    },
+    {
+        0x04, 0x05, 0x0c, 0x0d, 0x44, 0x45, 0x4c, 0x4d,
+        0x06, 0x07, 0x0e, 0x0f, 0x46, 0x47, 0x4e, 0x4f,
+        0x14, 0x15, 0x1c, 0x1d, 0x54, 0x55, 0x5c, 0x5d,
+        0x16, 0x17, 0x1e, 0x1f, 0x56, 0x57, 0x5e, 0x5f,
+        0x84, 0x85, 0x8c, 0x8d, 0xc4, 0xc5, 0xcc, 0xcd,
+        0x86, 0x87, 0x8e, 0x8f, 0xc6, 0xc7, 0xce, 0xcf,
+        0x94, 0x95, 0x9c, 0x9d, 0xd4, 0xd5, 0xdc, 0xdd,
+        0x96, 0x97, 0x9e, 0x9f, 0xd6, 0xd7, 0xde, 0xdf,
+    },
+    {
+        0x20, 0x21, 0x28, 0x29, 0x60, 0x61, 0x68, 0x69,
+        0x22, 0x23, 0x2a, 0x2b, 0x62, 0x63, 0x6a, 0x6b,
+        0x30, 0x31, 0x38, 0x39, 0x70, 0x71, 0x78, 0x79,
+        0x32, 0x33, 0x3a, 0x3b, 0x72, 0x73, 0x7a, 0x7b,
+        0xa0, 0xa1, 0xa8, 0xa9, 0xe0, 0xe1, 0xe8, 0xe9,
+        0xa2, 0xa3, 0xaa, 0xab, 0xe2, 0xe3, 0xea, 0xeb,
+        0xb0, 0xb1, 0xb8, 0xb9, 0xf0, 0xf1, 0xf8, 0xf9,
+        0xb2, 0xb3, 0xba, 0xbb, 0xf2, 0xf3, 0xfa, 0xfb,
+    },
+    {
+        0x24, 0x25, 0x2c, 0x2d, 0x64, 0x65, 0x6c, 0x6d,
+        0x26, 0x27, 0x2e, 0x2f, 0x66, 0x67, 0x6e, 0x6f,
+        0x34, 0x35, 0x3c, 0x3d, 0x74, 0x75, 0x7c, 0x7d,
+        0x36, 0x37, 0x3e, 0x3f, 0x76, 0x77, 0x7e, 0x7f,
+        0xa4, 0xa5, 0xac, 0xad, 0xe4, 0xe5, 0xec, 0xed,
+        0xa6, 0xa7, 0xae, 0xaf, 0xe6, 0xe7, 0xee, 0xef,
+        0xb4, 0xb5, 0xbc, 0xbd, 0xf4, 0xf5, 0xfc, 0xfd,
+        0xb6, 0xb7, 0xbe, 0xbf, 0xf6, 0xf7, 0xfe, 0xff,
+    },
+};
+
+#define MICRO_TILE_DIM      (8)
+#define MICRO_TILE_SZ       (2048)
+#define TEXELS_PER_ELEMENT  (1)
+#define BPP                 (64)
+
+void main() {
+    uint x =  gl_GlobalInvocationID.x % info.pitch;
+    uint y = (gl_GlobalInvocationID.x / info.pitch) % info.height;
+    uint z =  gl_GlobalInvocationID.x / (info.pitch * info.height);
+
+    uint col = bitfieldExtract(x, 0, 3);
+    uint row = bitfieldExtract(y, 0, 3);
+    uint lut = bitfieldExtract(z, 0, 2);
+    uint idx = lut_64bpp[lut][col + row * MICRO_TILE_DIM];
+
+    uint slice_offs = (z >> 2u) * info.c1 * MICRO_TILE_SZ; 
+    uint tile_row = y / MICRO_TILE_DIM;
+    uint tile_column = x / MICRO_TILE_DIM;
+    uint tile_offs = ((tile_row * info.c0) + tile_column) * MICRO_TILE_SZ;
+    uint offs = slice_offs + tile_offs + (idx * BPP / 8);
+
+    uint p0 = in_data[(offs >> 2) + 0];
+    uint p1 = in_data[(offs >> 2) + 1];    
+    out_data[2 * gl_GlobalInvocationID.x + 0] = p0;
+    out_data[2 * gl_GlobalInvocationID.x + 1] = p1;    
+}