texture_cache: detiler: added support for block coded 64bit images

src/video_core/host_shaders/CMakeLists.txt

@@ -4,6 +4,7 @@
 set(SHADER_FILES
     detile_m8x1.comp
     detile_m32x1.comp
+    detile_m32x2.comp
     detile_m32x4.comp
 )
 

src/video_core/host_shaders/detile_m32x2.comp

@@ -0,0 +1,48 @@
+// 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(rg32ui, binding = 1) uniform writeonly uimage2D output_img;
+
+layout(push_constant) uniform image_info {
+    uint pitch;
+} info;
+
+// Inverse morton LUT, small enough to fit into K$
+uint rmort[16] = {
+    0x11011000, 0x31213020,
+    0x13031202, 0x33233222,
+    0x51415040, 0x71617060,
+    0x53435242, 0x73637262,
+
+    0x15051404, 0x35253424,
+    0x17071606, 0x37273626,
+    0x55455444, 0x75657464,
+    0x57475646, 0x77677666,
+};
+
+#define MICRO_TILE_DIM      (8)
+
+void main() {
+    uint bit_ofs = 8 * (gl_LocalInvocationID.x % 4);
+    uint packed_pos = rmort[gl_LocalInvocationID.x >> 2] >> bit_ofs;
+    uint col = bitfieldExtract(packed_pos, 4, 4);
+    uint row = bitfieldExtract(packed_pos, 0, 4);
+
+    uint block_ofs = 2 * gl_GlobalInvocationID.x;
+    uint p0 = in_data[block_ofs + 0];
+    uint p1 = in_data[block_ofs + 1];
+
+    uint tiles_per_pitch = (info.pitch >> 3) >> 2; // log2(MICRO_TILE_DIM) / 4
+    ivec2 img_pos = MICRO_TILE_DIM * ivec2(
+        gl_WorkGroupID.x % tiles_per_pitch,
+        gl_WorkGroupID.x / tiles_per_pitch
+    );
+    imageStore(output_img, img_pos + ivec2(col, row), uvec4(p0, p1, 0, 0));
+}