Merge pull request #1446 from bunnei/fast_fermi_copy

gl_rasterizer: Implement accelerated Fermi2D copies.

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -617,14 +617,10 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
     InvalidateRegion(addr, size);
 }
 
-bool RasterizerOpenGL::AccelerateDisplayTransfer(const void* config) {
+bool RasterizerOpenGL::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src,
+                                             const Tegra::Engines::Fermi2D::Regs::Surface& dst) {
     MICROPROFILE_SCOPE(OpenGL_Blits);
-    UNREACHABLE();
-    return true;
-}
-
-bool RasterizerOpenGL::AccelerateTextureCopy(const void* config) {
-    UNREACHABLE();
+    res_cache.FermiCopySurface(src, dst);
     return true;
 }
 

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -52,8 +52,8 @@ public:
     void FlushRegion(VAddr addr, u64 size) override;
     void InvalidateRegion(VAddr addr, u64 size) override;
     void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
-    bool AccelerateDisplayTransfer(const void* config) override;
-    bool AccelerateTextureCopy(const void* config) override;
+    bool AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src,
+                               const Tegra::Engines::Fermi2D::Regs::Surface& dst) override;
     bool AccelerateFill(const void* config) override;
     bool AccelerateDisplay(const Tegra::FramebufferConfig& config, VAddr framebuffer_addr,
                            u32 pixel_stride) override;

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -143,6 +143,28 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
     return params;
 }
 
+/*static*/ SurfaceParams SurfaceParams::CreateForFermiCopySurface(
+    const Tegra::Engines::Fermi2D::Regs::Surface& config) {
+    SurfaceParams params{};
+    params.addr = TryGetCpuAddr(config.Address());
+    params.is_tiled = !config.linear;
+    params.block_height = params.is_tiled ? config.BlockHeight() : 0,
+    params.pixel_format = PixelFormatFromRenderTargetFormat(config.format);
+    params.component_type = ComponentTypeFromRenderTarget(config.format);
+    params.type = GetFormatType(params.pixel_format);
+    params.width = config.width;
+    params.height = config.height;
+    params.unaligned_height = config.height;
+    params.target = SurfaceTarget::Texture2D;
+    params.depth = 1;
+    params.size_in_bytes_total = params.SizeInBytesTotal();
+    params.size_in_bytes_2d = params.SizeInBytes2D();
+    params.max_mip_level = 0;
+    params.rt = {};
+
+    return params;
+}
+
 static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_format_tuples = {{
     {GL_RGBA8, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, ComponentType::UNorm, false}, // ABGR8U
     {GL_RGBA8, GL_RGBA, GL_BYTE, ComponentType::SNorm, false},                     // ABGR8S
@@ -559,6 +581,18 @@ static bool BlitSurface(const Surface& src_surface, const Surface& dst_surface,
     return true;
 }
 
+static void FastCopySurface(const Surface& src_surface, const Surface& dst_surface) {
+    const auto& src_params{src_surface->GetSurfaceParams()};
+    const auto& dst_params{dst_surface->GetSurfaceParams()};
+
+    const u32 width{std::min(src_params.width, dst_params.width)};
+    const u32 height{std::min(src_params.height, dst_params.height)};
+
+    glCopyImageSubData(src_surface->Texture().handle, SurfaceTargetToGL(src_params.target), 0, 0, 0,
+                       0, dst_surface->Texture().handle, SurfaceTargetToGL(dst_params.target), 0, 0,
+                       0, 0, width, height, 1);
+}
+
 static void CopySurface(const Surface& src_surface, const Surface& dst_surface,
                         GLuint copy_pbo_handle, GLenum src_attachment = 0,
                         GLenum dst_attachment = 0, std::size_t cubemap_face = 0) {
@@ -1033,6 +1067,26 @@ Surface RasterizerCacheOpenGL::GetUncachedSurface(const SurfaceParams& params) {
     return surface;
 }
 
+void RasterizerCacheOpenGL::FermiCopySurface(
+    const Tegra::Engines::Fermi2D::Regs::Surface& src_config,
+    const Tegra::Engines::Fermi2D::Regs::Surface& dst_config) {
+
+    const auto& src_params = SurfaceParams::CreateForFermiCopySurface(src_config);
+    const auto& dst_params = SurfaceParams::CreateForFermiCopySurface(dst_config);
+
+    ASSERT(src_params.width == dst_params.width);
+    ASSERT(src_params.height == dst_params.height);
+    ASSERT(src_params.pixel_format == dst_params.pixel_format);
+    ASSERT(src_params.block_height == dst_params.block_height);
+    ASSERT(src_params.is_tiled == dst_params.is_tiled);
+    ASSERT(src_params.depth == dst_params.depth);
+    ASSERT(src_params.depth == 1); // Currently, FastCopySurface only works with 2D surfaces
+    ASSERT(src_params.target == dst_params.target);
+    ASSERT(src_params.rt.index == dst_params.rt.index);
+
+    FastCopySurface(GetSurface(src_params, true), GetSurface(dst_params, false));
+}
+
 Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
                                                const SurfaceParams& new_params) {
     // Verify surface is compatible for blitting
@@ -1041,6 +1095,15 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
     // Get a new surface with the new parameters, and blit the previous surface to it
     Surface new_surface{GetUncachedSurface(new_params)};
 
+    // For compatible surfaces, we can just do fast glCopyImageSubData based copy
+    if (old_params.target == new_params.target && old_params.type == new_params.type &&
+        old_params.depth == new_params.depth && old_params.depth == 1 &&
+        SurfaceParams::GetFormatBpp(old_params.pixel_format) ==
+            SurfaceParams::GetFormatBpp(new_params.pixel_format)) {
+        FastCopySurface(old_surface, new_surface);
+        return new_surface;
+    }
+
     // If the format is the same, just do a framebuffer blit. This is significantly faster than
     // using PBOs. The is also likely less accurate, as textures will be converted rather than
     // reinterpreted. When use_accurate_framebuffers setting is enabled, perform a more accurate

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -13,6 +13,7 @@
 #include "common/common_types.h"
 #include "common/hash.h"
 #include "common/math_util.h"
+#include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
@@ -719,6 +720,10 @@ struct SurfaceParams {
                                               Tegra::GPUVAddr zeta_address,
                                               Tegra::DepthFormat format);
 
+    /// Creates SurfaceParams for a Fermi2D surface copy
+    static SurfaceParams CreateForFermiCopySurface(
+        const Tegra::Engines::Fermi2D::Regs::Surface& config);
+
     /// Checks if surfaces are compatible for caching
     bool IsCompatibleSurface(const SurfaceParams& other) const {
         return std::tie(pixel_format, type, width, height, target, depth) ==
@@ -837,6 +842,10 @@ public:
     /// Tries to find a framebuffer using on the provided CPU address
     Surface TryFindFramebufferSurface(VAddr addr) const;
 
+    /// Copies the contents of one surface to another
+    void FermiCopySurface(const Tegra::Engines::Fermi2D::Regs::Surface& src_config,
+                          const Tegra::Engines::Fermi2D::Regs::Surface& dst_config);
+
 private:
     void LoadSurface(const Surface& surface);
     Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true);