add macro 16bpp implementation

src/video_core/host_shaders/CMakeLists.txt

@@ -3,6 +3,7 @@
 
 set(SHADER_FILES
     detilers/display_micro_64bpp.comp
+    detilers/macro_16bpp.comp
     detilers/macro_32bpp.comp
     detilers/macro_64bpp.comp
     detilers/macro_8bpp.comp

src/video_core/host_shaders/detilers/macro_16bpp.comp

@@ -0,0 +1,73 @@
+// 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;
+
+#define MICRO_TILE_DIM  8u
+#define BPP             16u        
+#define MICRO_TILE_SZ   512u        
+shared uint scratch[32];            
+
+uint calc_micro_idx(uint col, uint row, uint slice2)
+{
+    uint idx = 0u;
+    idx |= (col & 1u) << 0u;
+    idx |= (row & 1u) << 1u;
+    idx |= (col & 2u) << 1u;
+    idx |= (row & 2u) << 2u;
+    idx |= (col & 4u) << 2u;
+    idx |= (row & 4u) << 3u;
+    idx |=  slice2     << 6u;
+    return idx;
+}
+
+void main()
+{
+    uint lane      = gl_LocalInvocationID.x;   
+    uint slot      = lane >> 1u;              
+    uint halfIdx   = lane & 1u;               
+
+    if (halfIdx == 0u) scratch[slot] = 0u;
+    barrier();                         
+
+    uint linear = gl_GlobalInvocationID.x;
+    uint x =  linear % info.pitch;
+    uint y = (linear / info.pitch) % info.height;
+    uint z =  linear / (info.pitch * info.height);
+
+    uint col = x & 7u;
+    uint row = y & 7u;
+    uint idx = calc_micro_idx(col, row, z & 3u);
+
+    uint slice_offs = (z >> 2u) * info.c1 * MICRO_TILE_SZ;
+    uint tile_row   = y / MICRO_TILE_DIM;
+    uint tile_col   = x / MICRO_TILE_DIM;
+    uint tile_offs  = ((tile_row * info.c0) + tile_col) * MICRO_TILE_SZ;
+    uint offs       = slice_offs + tile_offs + idx * 2u;   // 2 B/Texel
+
+    uint word   = in_data[offs >> 2u];
+    int  shift  = int((offs & 2u) * 8u);      // 0 oder 16
+    uint tex16  = bitfieldExtract(word, shift, 16);
+
+
+    uint mask = tex16 << (halfIdx * 16u);
+    scratch[slot] |= mask;
+
+    barrier();                               
+
+    if (halfIdx == 0u)
+        out_data[linear >> 1u] = scratch[slot];
+}

src/video_core/texture_cache/tile_manager.cpp

@@ -9,6 +9,7 @@
 #include "video_core/texture_cache/tile_manager.h"
 
 #include "video_core/host_shaders/detilers/display_micro_64bpp_comp.h"
+#include "video_core/host_shaders/detilers/macro_16bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_32bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_64bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_8bpp_comp.h"
@@ -46,6 +47,10 @@ const DetilerContext* TileManager::GetDetiler(const ImageInfo& info) const {
         switch (bpp) {
         case 8:
             return &detilers[DetilerType::Macro8];
+        case 16:
+            LOG_INFO(Render_Vulkan, "Using 16bpp detiler for volume texture: {} ({})",
+                     vk::to_string(info.pixel_format), NameOf(info.tiling_mode));
+            return &detilers[DetilerType::Macro16];
         case 32:
             return &detilers[DetilerType::Macro32];
         case 64:
@@ -77,11 +82,11 @@ struct DetilerParams {
 TileManager::TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler)
     : instance{instance}, scheduler{scheduler} {
     static const std::array detiler_shaders{
-        HostShaders::MICRO_8BPP_COMP,          HostShaders::MICRO_16BPP_COMP,
+        HostShaders::MICRO_8BPP_COMP,   HostShaders::MICRO_16BPP_COMP,
-        HostShaders::MICRO_32BPP_COMP,         HostShaders::MICRO_64BPP_COMP,
+        HostShaders::MICRO_32BPP_COMP,  HostShaders::MICRO_64BPP_COMP,
-        HostShaders::MICRO_128BPP_COMP,        HostShaders::MACRO_8BPP_COMP,
+        HostShaders::MICRO_128BPP_COMP, HostShaders::MACRO_8BPP_COMP,
-        HostShaders::MACRO_32BPP_COMP,         HostShaders::MACRO_64BPP_COMP,
+        HostShaders::MACRO_16BPP_COMP,  HostShaders::MACRO_32BPP_COMP,
-        HostShaders::DISPLAY_MICRO_64BPP_COMP,
+        HostShaders::MACRO_64BPP_COMP,  HostShaders::DISPLAY_MICRO_64BPP_COMP,
     };
 
     boost::container::static_vector<vk::DescriptorSetLayoutBinding, 2> bindings{

src/video_core/texture_cache/tile_manager.h

@@ -19,6 +19,7 @@ enum DetilerType : u32 {
     Micro128,
 
     Macro8,
+    Macro16,
     Macro32,
     Macro64,