GLCache: Use the full uncompressed size when blitting from one texture to another.

This avoids the problem of only copying a tiny piece of the textures when they are compressed.

src/video_core/engines/shader_bytecode.h

@@ -390,9 +390,6 @@ class OpCode {
 public:
     enum class Id {
         KIL,
-        BFE_C,
-        BFE_R,
-        BFE_IMM,
         BRA,
         LD_A,
         LD_C,
@@ -447,9 +444,6 @@ public:
         FMNMX_C,
         FMNMX_R,
         FMNMX_IMM,
-        IMNMX_C,
-        IMNMX_R,
-        IMNMX_IMM,
         FSETP_C, // Set Predicate
         FSETP_R,
         FSETP_IMM,
@@ -460,10 +454,6 @@ public:
         ISETP_IMM,
         ISETP_R,
         PSETP,
-        XMAD_IMM,
-        XMAD_CR,
-        XMAD_RC,
-        XMAD_RR,
     };
 
     enum class Type {
@@ -575,9 +565,6 @@ private:
         std::vector<Matcher> table = {
 #define INST(bitstring, op, type, name) Detail::GetMatcher(bitstring, op, type, name)
             INST("111000110011----", Id::KIL, Type::Flow, "KIL"),
-            INST("0100110000000---", Id::BFE_C, Type::Flow, "BFE_C"),
-            INST("0101110000000---", Id::BFE_R, Type::Flow, "BFE_R"),
-            INST("0011100-00000---", Id::BFE_IMM, Type::Flow, "BFE_IMM"),
             INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111110010---", Id::LD_C, Type::Memory, "LD_C"),
@@ -619,9 +606,6 @@ private:
             INST("0100110001100---", Id::FMNMX_C, Type::Arithmetic, "FMNMX_C"),
             INST("0101110001100---", Id::FMNMX_R, Type::Arithmetic, "FMNMX_R"),
             INST("0011100-01100---", Id::FMNMX_IMM, Type::Arithmetic, "FMNMX_IMM"),
-            INST("0100110000100---", Id::IMNMX_C, Type::Arithmetic, "FMNMX_IMM"),
-            INST("0101110000100---", Id::IMNMX_R, Type::Arithmetic, "FMNMX_IMM"),
-            INST("0011100-00100---", Id::IMNMX_IMM, Type::Arithmetic, "FMNMX_IMM"),
             INST("000001----------", Id::LOP32I, Type::Logic, "LOP32I"),
             INST("0100110001001---", Id::SHL_C, Type::Shift, "SHL_C"),
             INST("0101110001001---", Id::SHL_R, Type::Shift, "SHL_R"),
@@ -645,10 +629,6 @@ private:
             INST("010110110110----", Id::ISETP_R, Type::IntegerSetPredicate, "ISETP_R"),
             INST("0011011-0110----", Id::ISETP_IMM, Type::IntegerSetPredicate, "ISETP_IMM"),
             INST("0101000010010---", Id::PSETP, Type::PredicateSetPredicate, "PSETP"),
-            INST("0011011-00------", Id::XMAD_IMM, Type::Arithmetic, "XMAD_IMM"),
-            INST("0100111---------", Id::XMAD_CR, Type::Arithmetic, "XMAD_CR"),
-            INST("010100010-------", Id::XMAD_RC, Type::Arithmetic, "XMAD_RC"),
-            INST("0101101100------", Id::XMAD_RR, Type::Arithmetic, "XMAD_RR"),
         };
 #undef INST
         std::stable_sort(table.begin(), table.end(), [](const auto& a, const auto& b) {

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -169,60 +169,10 @@ static void AllocateSurfaceTexture(GLuint texture, const FormatTuple& format_tup
 static bool BlitTextures(GLuint src_tex, const MathUtil::Rectangle<u32>& src_rect, GLuint dst_tex,
                          const MathUtil::Rectangle<u32>& dst_rect, SurfaceType type,
                          GLuint read_fb_handle, GLuint draw_fb_handle) {
-    OpenGLState state = OpenGLState::GetCurState();
-
-    OpenGLState prev_state = state;
-    SCOPE_EXIT({ prev_state.Apply(); });
-
-    // Make sure textures aren't bound to texture units, since going to bind them to framebuffer
-    // components
-    state.ResetTexture(src_tex);
-    state.ResetTexture(dst_tex);
-
-    state.draw.read_framebuffer = read_fb_handle;
-    state.draw.draw_framebuffer = draw_fb_handle;
-    state.Apply();
-
-    u32 buffers = 0;
-
-    if (type == SurfaceType::ColorTexture) {
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, src_tex,
-                               0);
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0,
-                               0);
-
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, dst_tex,
-                               0);
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0,
-                               0);
-
-        buffers = GL_COLOR_BUFFER_BIT;
-    } else if (type == SurfaceType::Depth) {
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, src_tex, 0);
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
-
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, dst_tex, 0);
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
-
-        buffers = GL_DEPTH_BUFFER_BIT;
-    } else if (type == SurfaceType::DepthStencil) {
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D,
-                               src_tex, 0);
-
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D,
-                               dst_tex, 0);
-
-        buffers = GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT;
-    }
-
-    glBlitFramebuffer(src_rect.left, src_rect.bottom, src_rect.right, src_rect.top, dst_rect.left,
-                      dst_rect.bottom, dst_rect.right, dst_rect.top, buffers,
-                      buffers == GL_COLOR_BUFFER_BIT ? GL_LINEAR : GL_NEAREST);
 
+    glCopyImageSubData(src_tex, GL_TEXTURE_2D, 0, src_rect.left, src_rect.bottom, 0, dst_tex,
+                       GL_TEXTURE_2D, 0, dst_rect.left, dst_rect.bottom, 0, src_rect.GetWidth(),
+                       src_rect.GetHeight(), 0);
     return true;
 }
 
@@ -1102,16 +1052,19 @@ Surface RasterizerCacheOpenGL::GetTextureSurface(const Tegra::Texture::FullTextu
 
     params.UpdateParams();
 
-    if (config.tic.Width() % 8 != 0 || config.tic.Height() % 8 != 0 ||
+    if (params.GetActualWidth() % 8 != 0 || params.GetActualHeight() % 8 != 0 ||
         params.stride != params.width) {
         Surface src_surface;
         MathUtil::Rectangle<u32> rect;
         std::tie(src_surface, rect) = GetSurfaceSubRect(params, ScaleMatch::Ignore, true);
 
+        rect = rect.Scale(params.GetCompresssionFactor());
+
         params.res_scale = src_surface->res_scale;
         Surface tmp_surface = CreateSurface(params);
-        BlitTextures(src_surface->texture.handle, rect, tmp_surface->texture.handle,
-                     tmp_surface->GetScaledRect(),
+
+        auto dst_rect = tmp_surface->GetScaledRect().Scale(params.GetCompresssionFactor());
+        BlitTextures(src_surface->texture.handle, rect, tmp_surface->texture.handle, dst_rect,
                      SurfaceParams::GetFormatType(params.pixel_format), read_framebuffer.handle,
                      draw_framebuffer.handle);
 

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -1112,11 +1112,13 @@ private:
             break;
         }
         case OpCode::Type::Memory: {
+            const Attribute::Index attribute = instr.attribute.fmt20.index;
+
             switch (opcode->GetId()) {
             case OpCode::Id::LD_A: {
                 ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
                 regs.SetRegisterToInputAttibute(instr.gpr0, instr.attribute.fmt20.element,
-                                                instr.attribute.fmt20.index);
+                                                attribute);
                 break;
             }
             case OpCode::Id::LD_C: {
@@ -1148,11 +1150,12 @@ private:
             }
             case OpCode::Id::ST_A: {
                 ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
-                regs.SetOutputAttributeToRegister(instr.attribute.fmt20.index,
-                                                  instr.attribute.fmt20.element, instr.gpr0);
+                regs.SetOutputAttributeToRegister(attribute, instr.attribute.fmt20.element,
+                                                  instr.gpr0);
                 break;
             }
             case OpCode::Id::TEX: {
+                ASSERT_MSG(instr.attribute.fmt20.size == 4, "untested");
                 const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
                 const std::string op_b = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
                 const std::string sampler = GetSampler(instr.sampler);
@@ -1165,7 +1168,7 @@ private:
                 const std::string texture = "texture(" + sampler + ", coords)";
 
                 size_t dest_elem{};
-                for (size_t elem = 0; elem < 4; ++elem) {
+                for (size_t elem = 0; elem < instr.attribute.fmt20.size; ++elem) {
                     if (!instr.tex.IsComponentEnabled(elem)) {
                         // Skip disabled components
                         continue;
@@ -1178,6 +1181,7 @@ private:
                 break;
             }
             case OpCode::Id::TEXS: {
+                ASSERT_MSG(instr.attribute.fmt20.size == 4, "untested");
                 const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
                 const std::string op_b = regs.GetRegisterAsFloat(instr.gpr20);
                 const std::string sampler = GetSampler(instr.sampler);