Merge pull request #483 from bunnei/sonic

Several GPU fixes to boot Sonic Mania

src/core/hle/service/nvdrv/devices/nvhost_ctrl.cpp

@@ -16,7 +16,11 @@ u32 nvhost_ctrl::ioctl(Ioctl command, const std::vector<u8>& input, std::vector<
     case IoctlCommand::IocGetConfigCommand:
         return NvOsGetConfigU32(input, output);
     case IoctlCommand::IocCtrlEventWaitCommand:
-        return IocCtrlEventWait(input, output);
+        return IocCtrlEventWait(input, output, false);
+    case IoctlCommand::IocCtrlEventWaitAsyncCommand:
+        return IocCtrlEventWait(input, output, true);
+    case IoctlCommand::IocCtrlEventRegisterCommand:
+        return IocCtrlEventRegister(input, output);
     }
     UNIMPLEMENTED_MSG("Unimplemented ioctl");
     return 0;
@@ -45,11 +49,13 @@ u32 nvhost_ctrl::NvOsGetConfigU32(const std::vector<u8>& input, std::vector<u8>&
     return 0;
 }
 
-u32 nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output) {
+u32 nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output,
+                                  bool is_async) {
     IocCtrlEventWaitParams params{};
     std::memcpy(&params, input.data(), sizeof(params));
-    NGLOG_WARNING(Service_NVDRV, "(STUBBED) called, syncpt_id={} threshold={} timeout={}",
+    NGLOG_WARNING(Service_NVDRV,
-                  params.syncpt_id, params.threshold, params.timeout);
+                  "(STUBBED) called, syncpt_id={}, threshold={}, timeout={}, is_async={}",
+                  params.syncpt_id, params.threshold, params.timeout, is_async);
 
     // TODO(Subv): Implement actual syncpt waiting.
     params.value = 0;
@@ -57,4 +63,10 @@ u32 nvhost_ctrl::IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>&
     return 0;
 }
 
+u32 nvhost_ctrl::IocCtrlEventRegister(const std::vector<u8>& input, std::vector<u8>& output) {
+    NGLOG_WARNING(Service_NVDRV, "(STUBBED) called");
+    // TODO(bunnei): Implement this.
+    return 0;
+}
+
 } // namespace Service::Nvidia::Devices

src/core/hle/service/nvdrv/devices/nvhost_ctrl.h

@@ -134,7 +134,9 @@ private:
 
     u32 NvOsGetConfigU32(const std::vector<u8>& input, std::vector<u8>& output);
 
-    u32 IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output);
+    u32 IocCtrlEventWait(const std::vector<u8>& input, std::vector<u8>& output, bool is_async);
+
+    u32 IocCtrlEventRegister(const std::vector<u8>& input, std::vector<u8>& output);
 };
 
 } // namespace Service::Nvidia::Devices

src/video_core/engines/shader_bytecode.h

@@ -456,9 +456,9 @@ private:
             INST("00011110--------", Id::FMUL32_IMM, Type::Arithmetic, "FMUL32_IMM"),
             INST("0101000010000---", Id::MUFU, Type::Arithmetic, "MUFU"),
             INST("0101110010010---", Id::RRO, Type::Arithmetic, "RRO"),
-            INST("0100110010101---", Id::F2F_C, Type::Arithmetic, "F2F_C"),
+            INST("0100110010101---", Id::F2F_C, Type::Conversion, "F2F_C"),
-            INST("0101110010101---", Id::F2F_R, Type::Arithmetic, "F2F_R"),
+            INST("0101110010101---", Id::F2F_R, Type::Conversion, "F2F_R"),
-            INST("0011100-10101---", Id::F2F_IMM, Type::Arithmetic, "F2F_IMM"),
+            INST("0011100-10101---", Id::F2F_IMM, Type::Conversion, "F2F_IMM"),
             INST("0100110010110---", Id::F2I_C, Type::Arithmetic, "F2I_C"),
             INST("0101110010110---", Id::F2I_R, Type::Arithmetic, "F2I_R"),
             INST("0011100-10110---", Id::F2I_IMM, Type::Arithmetic, "F2I_IMM"),

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -45,7 +45,7 @@ struct FormatTuple {
 
 static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_format_tuples = {{
     {GL_RGBA8, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV, false},                    // ABGR8
-    {GL_RGB, GL_RGB, GL_UNSIGNED_SHORT_5_6_5_REV, false},                       // B5G6R5
+    {GL_RGB, GL_RGB, GL_UNSIGNED_SHORT_5_6_5, false},                           // B5G6R5
     {GL_RGB10_A2, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, false},              // A2B10G10R10
     {GL_RGB5_A1, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, false},                // A1B5G5R5
     {GL_R8, GL_RED, GL_UNSIGNED_BYTE, false},                                   // R8

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -834,13 +834,14 @@ private:
         }
         case OpCode::Type::Conversion: {
             ASSERT_MSG(instr.conversion.size == Register::Size::Word, "Unimplemented");
-            ASSERT_MSG(!instr.conversion.selector, "Unimplemented");
             ASSERT_MSG(!instr.conversion.negate_a, "Unimplemented");
             ASSERT_MSG(!instr.conversion.saturate_a, "Unimplemented");
 
             switch (opcode->GetId()) {
             case OpCode::Id::I2I_R:
             case OpCode::Id::I2F_R: {
+                ASSERT_MSG(!instr.conversion.selector, "Unimplemented");
+
                 std::string op_a =
                     regs.GetRegisterAsInteger(instr.gpr20, 0, instr.conversion.is_signed);
 
@@ -851,6 +852,16 @@ private:
                 regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_signed, 0, op_a, 1, 1);
                 break;
             }
+            case OpCode::Id::F2F_R: {
+                std::string op_a = regs.GetRegisterAsFloat(instr.gpr20);
+
+                if (instr.conversion.abs_a) {
+                    op_a = "abs(" + op_a + ')';
+                }
+
+                regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1);
+                break;
+            }
             default: {
                 NGLOG_CRITICAL(HW_GPU, "Unhandled conversion instruction: {}", opcode->GetName());
                 UNREACHABLE();