video_core: Add vulkan shaders (#6619)

src/video_core/host_shaders/CMakeLists.txt

@@ -4,13 +4,12 @@
 
 set(SHADER_FILES
     format_reinterpreter/d24s8_to_rgba8.frag
-    format_reinterpreter/fullscreen_quad.vert
     format_reinterpreter/rgba4_to_rgb5a1.frag
+    format_reinterpreter/vulkan_d24s8_to_rgba8.comp
     texture_filtering/bicubic.frag
     texture_filtering/nearest_neighbor.frag
     texture_filtering/refine.frag
     texture_filtering/scale_force.frag
-    texture_filtering/tex_coord.vert
     texture_filtering/xbrz_freescale.frag
     texture_filtering/mmpx.frag
     texture_filtering/x_gradient.frag
@@ -20,8 +19,22 @@ set(SHADER_FILES
     opengl_present.vert
     opengl_present_anaglyph.frag
     opengl_present_interlaced.frag
+    vulkan_depth_to_buffer.comp
+    vulkan_present.frag
+    vulkan_present.vert
+    vulkan_present_anaglyph.frag
+    vulkan_present_interlaced.frag
+    vulkan_blit_depth_stencil.frag
 )
 
+find_program(GLSLANGVALIDATOR "glslangValidator")
+if ("${GLSLANGVALIDATOR}" STREQUAL "GLSLANGVALIDATOR-NOTFOUND")
+    message(FATAL_ERROR "Required program `glslangValidator` not found.")
+endif()
+
+set(MACROS "-Dgl_VertexID=gl_VertexIndex")
+set(QUIET_FLAG "--quiet")
+
 set(SHADER_INCLUDE ${CMAKE_CURRENT_BINARY_DIR}/include)
 set(SHADER_DIR ${SHADER_INCLUDE}/video_core/host_shaders)
 set(HOST_SHADERS_INCLUDE ${SHADER_INCLUDE} PARENT_SCOPE)
@@ -29,6 +42,23 @@ set(HOST_SHADERS_INCLUDE ${SHADER_INCLUDE} PARENT_SCOPE)
 set(INPUT_FILE ${CMAKE_CURRENT_SOURCE_DIR}/source_shader.h.in)
 set(HEADER_GENERATOR ${CMAKE_CURRENT_SOURCE_DIR}/StringShaderHeader.cmake)
 
+# Check if `--quiet` is available on host's glslangValidator version
+# glslangValidator prints to STDERR iff an unrecognized flag is passed to it
+execute_process(
+    COMMAND
+        ${GLSLANGVALIDATOR} ${QUIET_FLAG}
+    ERROR_VARIABLE
+        GLSLANG_ERROR
+    # STDOUT variable defined to silence unnecessary output during CMake configuration
+    OUTPUT_VARIABLE
+        GLSLANG_OUTPUT
+)
+
+if (NOT GLSLANG_ERROR STREQUAL "")
+    message(WARNING "Refusing to use unavailable flag `${QUIET_FLAG}` on `${GLSLANGVALIDATOR}`")
+    set(QUIET_FLAG "")
+endif()
+
 foreach(FILENAME IN ITEMS ${SHADER_FILES})
     string(REPLACE "." "_" SHADER_NAME ${FILENAME})
     set(SOURCE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/${FILENAME})
@@ -48,6 +78,21 @@ foreach(FILENAME IN ITEMS ${SHADER_FILES})
         )
         set(SHADER_HEADERS ${SHADER_HEADERS} ${SOURCE_HEADER_FILE})
     endif()
+    # Skip compiling to SPIR-V OpenGL exclusive files
+    if (NOT ${FILENAME} MATCHES "opengl.*")
+        get_filename_component(FILE_NAME ${SHADER_NAME} NAME)
+        string(TOUPPER ${FILE_NAME}_SPV SPIRV_VARIABLE_NAME)
+        set(SPIRV_HEADER_FILE ${SHADER_DIR}/${SHADER_NAME}_spv.h)
+        add_custom_command(
+            OUTPUT
+                ${SPIRV_HEADER_FILE}
+            COMMAND
+                ${GLSLANGVALIDATOR} --target-env vulkan1.1 --glsl-version 450 ${QUIET_FLAG} ${MACROS} --variable-name ${SPIRV_VARIABLE_NAME} -o ${SPIRV_HEADER_FILE} ${SOURCE_FILE}
+            MAIN_DEPENDENCY
+                ${SOURCE_FILE}
+        )
+        set(SHADER_HEADERS ${SHADER_HEADERS} ${SPIRV_HEADER_FILE})
+    endif()
 endforeach()
 
 set(SHADER_SOURCES ${SHADER_FILES})

src/video_core/host_shaders/format_reinterpreter/vulkan_d24s8_to_rgba8.comp

@@ -0,0 +1,26 @@
+// Copyright 2023 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_EXT_samplerless_texture_functions : require
+
+layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
+layout(set = 0, binding = 0) uniform highp texture2D depth;
+layout(set = 0, binding = 1) uniform lowp utexture2D stencil;
+layout(set = 0, binding = 2, rgba8) uniform highp writeonly image2D color;
+
+layout(push_constant, std140) uniform ComputeInfo {
+    mediump ivec2 src_offset;
+    mediump ivec2 extent;
+};
+
+void main() {
+    ivec2 tex_coord = src_offset + ivec2(gl_GlobalInvocationID.xy);
+    highp uint depth_val =
+        uint(texelFetch(depth, tex_coord, 0).x * (exp2(32.0) - 1.0));
+    lowp uint stencil_val = texelFetch(stencil, tex_coord, 0).x;
+    highp uvec4 components =
+        uvec4(stencil_val, (uvec3(depth_val) >> uvec3(24u, 16u, 8u)) & 0x000000FFu);
+    imageStore(color, tex_coord, vec4(components) / (exp2(8.0) - 1.0));
+}

src/video_core/host_shaders/full_screen_triangle.vert

@@ -9,8 +9,20 @@ out gl_PerVertex {
 
 layout(location = 0) out vec2 texcoord;
 
-layout (location = 0) uniform vec2 tex_scale;
+#ifdef VULKAN
-layout (location = 1) uniform vec2 tex_offset;
+#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
+#define END_PUSH_CONSTANTS };
+#define UNIFORM(n)
+#else // if OpenGL
+#define BEGIN_PUSH_CONSTANTS
+#define END_PUSH_CONSTANTS
+#define UNIFORM(n) layout (location = n) uniform
+#endif
+
+BEGIN_PUSH_CONSTANTS
+UNIFORM(0) vec2 tex_scale;
+UNIFORM(1) vec2 tex_offset;
+END_PUSH_CONSTANTS
 
 void main() {
     float x = float((gl_VertexID & 1) << 2);

src/video_core/host_shaders/texture_filtering/xbrz_freescale.frag

@@ -9,7 +9,14 @@ layout(location = 0) in vec2 tex_coord;
 layout(location = 0) out vec4 frag_color;
 
 layout(binding = 0) uniform sampler2D tex;
-layout(location = 2) uniform lowp float scale;
+
+#ifdef VULKAN
+layout(push_constant, std140) uniform XbrzInfo {
+    float scale;
+};
+#else
+layout(location = 2) uniform float scale;
+#endif
 
 const int BLEND_NONE = 0;
 const int BLEND_NORMAL = 1;

src/video_core/host_shaders/vulkan_blit_depth_stencil.frag

@@ -0,0 +1,16 @@
+// Copyright 2022 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_ARB_shader_stencil_export : require
+
+layout(binding = 0) uniform sampler2D depth_tex;
+layout(binding = 1) uniform usampler2D stencil_tex;
+
+layout(location = 0) in vec2 texcoord;
+
+void main() {
+    gl_FragDepth = textureLod(depth_tex, texcoord, 0).r;
+    gl_FragStencilRefARB = int(textureLod(stencil_tex, texcoord, 0).r);
+}

src/video_core/host_shaders/vulkan_depth_to_buffer.comp

@@ -0,0 +1,29 @@
+// Copyright 2023 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+
+layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
+layout(binding = 0) uniform highp sampler2D depth;
+layout(binding = 1) uniform lowp usampler2D stencil;
+
+layout(binding = 2) writeonly buffer OutputBuffer{
+    uint pixels[];
+} staging;
+
+layout(push_constant, std140) uniform ComputeInfo {
+    mediump ivec2 src_offset;
+    mediump ivec2 extent;
+};
+
+void main() {
+    ivec2 rect = ivec2(gl_NumWorkGroups.xy) * ivec2(8);
+    ivec2 dst_coord = ivec2(gl_GlobalInvocationID.xy);
+    ivec2 tex_icoord = src_offset + dst_coord;
+    highp vec2 tex_coord = vec2(tex_icoord) / vec2(extent);
+    highp uint depth_val = uint(texture(depth, tex_coord).x * (exp2(24.0) - 1.0));
+    lowp uint stencil_val = texture(stencil, tex_coord).x;
+    highp uint value = stencil_val | (depth_val << 8);
+    staging.pixels[dst_coord.y * rect.x + dst_coord.x] = value;
+}

src/video_core/host_shaders/vulkan_present.frag

@@ -0,0 +1,25 @@
+// Copyright 2022 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_ARB_separate_shader_objects : enable
+
+layout (location = 0) in vec2 frag_tex_coord;
+layout (location = 0) out vec4 color;
+
+layout (push_constant, std140) uniform DrawInfo {
+    mat4 modelview_matrix;
+    vec4 i_resolution;
+    vec4 o_resolution;
+    int screen_id_l;
+    int screen_id_r;
+    int layer;
+    int reverse_interlaced;
+};
+
+layout (set = 0, binding = 0) uniform sampler2D screen_textures[3];
+
+void main() {
+    color = texture(screen_textures[screen_id_l], frag_tex_coord);
+}

src/video_core/host_shaders/vulkan_present.vert

@@ -0,0 +1,25 @@
+// Copyright 2022 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_ARB_separate_shader_objects : enable
+
+layout (location = 0) in vec2 vert_position;
+layout (location = 1) in vec2 vert_tex_coord;
+layout (location = 0) out vec2 frag_tex_coord;
+
+layout (push_constant, std140) uniform DrawInfo {
+    mat4 modelview_matrix;
+    vec4 i_resolution;
+    vec4 o_resolution;
+    int screen_id_l;
+    int screen_id_r;
+    int layer;
+};
+
+void main() {
+    vec4 position = vec4(vert_position, 0.0, 1.0) * modelview_matrix;
+    gl_Position = vec4(position.x, position.y, 0.0, 1.0);
+    frag_tex_coord = vert_tex_coord;
+}

src/video_core/host_shaders/vulkan_present_anaglyph.frag

@@ -0,0 +1,39 @@
+// Copyright 2022 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_ARB_separate_shader_objects : enable
+
+layout (location = 0) in vec2 frag_tex_coord;
+layout (location = 0) out vec4 color;
+
+// Anaglyph Red-Cyan shader based on Dubois algorithm
+// Constants taken from the paper:
+// "Conversion of a Stereo Pair to Anaglyph with
+// the Least-Squares Projection Method"
+// Eric Dubois, March 2009
+const mat3 l = mat3( 0.437, 0.449, 0.164,
+              -0.062,-0.062,-0.024,
+              -0.048,-0.050,-0.017);
+const mat3 r = mat3(-0.011,-0.032,-0.007,
+               0.377, 0.761, 0.009,
+              -0.026,-0.093, 1.234);
+
+layout (push_constant, std140) uniform DrawInfo {
+    mat4 modelview_matrix;
+    vec4 i_resolution;
+    vec4 o_resolution;
+    int screen_id_l;
+    int screen_id_r;
+    int layer;
+    int reverse_interlaced;
+};
+
+layout (set = 0, binding = 0) uniform sampler2D screen_textures[3];
+
+void main() {
+    vec4 color_tex_l = texture(screen_textures[screen_id_l], frag_tex_coord);
+    vec4 color_tex_r = texture(screen_textures[screen_id_r], frag_tex_coord);
+    color = vec4(color_tex_l.rgb*l+color_tex_r.rgb*r, color_tex_l.a);
+}

src/video_core/host_shaders/vulkan_present_interlaced.frag

@@ -0,0 +1,29 @@
+// Copyright 2022 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#version 450 core
+#extension GL_ARB_separate_shader_objects : enable
+
+layout (location = 0) in vec2 frag_tex_coord;
+layout (location = 0) out vec4 color;
+
+layout (push_constant, std140) uniform DrawInfo {
+    mat4 modelview_matrix;
+    vec4 i_resolution;
+    vec4 o_resolution;
+    int screen_id_l;
+    int screen_id_r;
+    int layer;
+    int reverse_interlaced;
+};
+
+layout (set = 0, binding = 0) uniform sampler2D screen_textures[3];
+
+void main() {
+    float screen_row = o_resolution.x * frag_tex_coord.x;
+    if (int(screen_row) % 2 == reverse_interlaced)
+        color = texture(screen_textures[screen_id_l], frag_tex_coord);
+    else
+        color = texture(screen_textures[screen_id_r], frag_tex_coord);
+}

src/video_core/rasterizer_cache/rasterizer_cache.h

@@ -1000,7 +1000,7 @@ void RasterizerCache<T>::UploadSurface(Surface& surface, SurfaceInterval interva
     }
 
     const BufferTextureCopy upload = {
-        .buffer_offset = 0,
+        .buffer_offset = staging.offset,
         .buffer_size = staging.size,
         .texture_rect = surface.GetSubRect(load_info),
         .texture_level = surface.LevelOf(load_info.addr),
@@ -1080,7 +1080,7 @@ void RasterizerCache<T>::DownloadSurface(Surface& surface, SurfaceInterval inter
         flush_info.width * flush_info.height * surface.GetInternalBytesPerPixel(), false);
 
     const BufferTextureCopy download = {
-        .buffer_offset = 0,
+        .buffer_offset = staging.offset,
         .buffer_size = staging.size,
         .texture_rect = surface.GetSubRect(flush_info),
         .texture_level = surface.LevelOf(flush_start),
@@ -1137,7 +1137,7 @@ bool RasterizerCache<T>::ValidateByReinterpretation(Surface& surface, SurfacePar
     if (reinterpret_id) {
         Surface& src_surface = slot_surfaces[reinterpret_id];
         const SurfaceInterval copy_interval = src_surface.GetCopyableInterval(params);
-        if (boost::icl::is_empty(copy_interval)) {
+        if (boost::icl::is_empty(copy_interval & interval)) {
             return false;
         }
         const PAddr addr = boost::icl::lower(interval);

src/video_core/rasterizer_cache/utils.h

@@ -75,8 +75,8 @@ struct BufferTextureCopy {
 
 struct StagingData {
     u32 size;
+    u32 offset;
     std::span<u8> mapped;
-    u64 buffer_offset;
 };
 
 struct TextureCubeConfig {

src/video_core/renderer_opengl/gl_texture_runtime.cpp

@@ -144,8 +144,8 @@ VideoCore::StagingData TextureRuntime::FindStaging(u32 size, bool upload) {
     }
     return VideoCore::StagingData{
         .size = size,
+        .offset = 0,
         .mapped = std::span{staging_buffer.data(), size},
-        .buffer_offset = 0,
     };
 }