early-access version 3342

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 3341.
+This is the source code for early-access 3342.
 
 ## Legal Notice
 

src/input_common/helpers/joycon_protocol/calibration.cpp

@@ -13,34 +13,34 @@ CalibrationProtocol::CalibrationProtocol(std::shared_ptr<JoyconHandle> handle)
 
 DriverResult CalibrationProtocol::GetLeftJoyStickCalibration(JoyStickCalibration& calibration) {
     ScopedSetBlocking sb(this);
-    std::vector<u8> buffer;
     DriverResult result{DriverResult::Success};
+    JoystickLeftSpiCalibration spi_calibration{};
+    bool has_user_calibration = false;
     calibration = {};
 
-    result = ReadSPI(CalAddr::USER_LEFT_MAGIC, sizeof(u16), buffer);
-
     if (result == DriverResult::Success) {
-        const bool has_user_calibration = buffer[0] == 0xB2 && buffer[1] == 0xA1;
+        result = HasUserCalibration(SpiAddress::USER_LEFT_MAGIC, has_user_calibration);
-        if (has_user_calibration) {
-            result = ReadSPI(CalAddr::USER_LEFT_DATA, 9, buffer);
-        } else {
-            result = ReadSPI(CalAddr::FACT_LEFT_DATA, 9, buffer);
     }
+
+    // Read User defined calibration
+    if (result == DriverResult::Success && has_user_calibration) {
+        result = ReadSPI(SpiAddress::USER_LEFT_DATA, spi_calibration);
+    }
+
+    // Read Factory calibration
+    if (result == DriverResult::Success && !has_user_calibration) {
+        result = ReadSPI(SpiAddress::FACT_LEFT_DATA, spi_calibration);
     }
 
     if (result == DriverResult::Success) {
-        calibration.x.max = static_cast<u16>(((buffer[1] & 0x0F) << 8) | buffer[0]);
+        calibration.x.center = GetXAxisCalibrationValue(spi_calibration.center);
-        calibration.y.max = static_cast<u16>((buffer[2] << 4) | (buffer[1] >> 4));
+        calibration.y.center = GetYAxisCalibrationValue(spi_calibration.center);
-        calibration.x.center = static_cast<u16>(((buffer[4] & 0x0F) << 8) | buffer[3]);
+        calibration.x.min = GetXAxisCalibrationValue(spi_calibration.min);
-        calibration.y.center = static_cast<u16>((buffer[5] << 4) | (buffer[4] >> 4));
+        calibration.y.min = GetYAxisCalibrationValue(spi_calibration.min);
-        calibration.x.min = static_cast<u16>(((buffer[7] & 0x0F) << 8) | buffer[6]);
+        calibration.x.max = GetXAxisCalibrationValue(spi_calibration.max);
-        calibration.y.min = static_cast<u16>((buffer[8] << 4) | (buffer[7] >> 4));
+        calibration.y.max = GetYAxisCalibrationValue(spi_calibration.max);
     }
 
-    // Nintendo fix for drifting stick
-    // result = ReadSPI(0x60, 0x86 ,buffer, 16);
-    // calibration.deadzone = (u16)((buffer[4] << 8) & 0xF00 | buffer[3]);
-
     // Set a valid default calibration if data is missing
     ValidateCalibration(calibration);
 
@@ -49,34 +49,34 @@ DriverResult CalibrationProtocol::GetLeftJoyStickCalibration(JoyStickCalibration
 
 DriverResult CalibrationProtocol::GetRightJoyStickCalibration(JoyStickCalibration& calibration) {
     ScopedSetBlocking sb(this);
-    std::vector<u8> buffer;
     DriverResult result{DriverResult::Success};
+    JoystickRightSpiCalibration spi_calibration{};
+    bool has_user_calibration = false;
     calibration = {};
 
-    result = ReadSPI(CalAddr::USER_RIGHT_MAGIC, sizeof(u16), buffer);
-
     if (result == DriverResult::Success) {
-        const bool has_user_calibration = buffer[0] == 0xB2 && buffer[1] == 0xA1;
+        result = HasUserCalibration(SpiAddress::USER_RIGHT_MAGIC, has_user_calibration);
-        if (has_user_calibration) {
-            result = ReadSPI(CalAddr::USER_RIGHT_DATA, 9, buffer);
-        } else {
-            result = ReadSPI(CalAddr::FACT_RIGHT_DATA, 9, buffer);
     }
+
+    // Read User defined calibration
+    if (result == DriverResult::Success && has_user_calibration) {
+        result = ReadSPI(SpiAddress::USER_RIGHT_DATA, spi_calibration);
+    }
+
+    // Read Factory calibration
+    if (result == DriverResult::Success && !has_user_calibration) {
+        result = ReadSPI(SpiAddress::FACT_RIGHT_DATA, spi_calibration);
     }
 
     if (result == DriverResult::Success) {
-        calibration.x.center = static_cast<u16>(((buffer[1] & 0x0F) << 8) | buffer[0]);
+        calibration.x.center = GetXAxisCalibrationValue(spi_calibration.center);
-        calibration.y.center = static_cast<u16>((buffer[2] << 4) | (buffer[1] >> 4));
+        calibration.y.center = GetYAxisCalibrationValue(spi_calibration.center);
-        calibration.x.min = static_cast<u16>(((buffer[4] & 0x0F) << 8) | buffer[3]);
+        calibration.x.min = GetXAxisCalibrationValue(spi_calibration.min);
-        calibration.y.min = static_cast<u16>((buffer[5] << 4) | (buffer[4] >> 4));
+        calibration.y.min = GetYAxisCalibrationValue(spi_calibration.min);
-        calibration.x.max = static_cast<u16>(((buffer[7] & 0x0F) << 8) | buffer[6]);
+        calibration.x.max = GetXAxisCalibrationValue(spi_calibration.max);
-        calibration.y.max = static_cast<u16>((buffer[8] << 4) | (buffer[7] >> 4));
+        calibration.y.max = GetYAxisCalibrationValue(spi_calibration.max);
     }
 
-    // Nintendo fix for drifting stick
-    // buffer = ReadSPI(0x60, 0x98 , 16);
-    // joystick.deadzone = (u16)((buffer[4] << 8) & 0xF00 | buffer[3]);
-
     // Set a valid default calibration if data is missing
     ValidateCalibration(calibration);
 
@@ -85,39 +85,41 @@ DriverResult CalibrationProtocol::GetRightJoyStickCalibration(JoyStickCalibratio
 
 DriverResult CalibrationProtocol::GetImuCalibration(MotionCalibration& calibration) {
     ScopedSetBlocking sb(this);
-    std::vector<u8> buffer;
     DriverResult result{DriverResult::Success};
+    ImuSpiCalibration spi_calibration{};
+    bool has_user_calibration = false;
     calibration = {};
 
-    result = ReadSPI(CalAddr::USER_IMU_MAGIC, sizeof(u16), buffer);
-
     if (result == DriverResult::Success) {
-        const bool has_user_calibration = buffer[0] == 0xB2 && buffer[1] == 0xA1;
+        result = HasUserCalibration(SpiAddress::USER_IMU_MAGIC, has_user_calibration);
-        if (has_user_calibration) {
-            result = ReadSPI(CalAddr::USER_IMU_DATA, sizeof(IMUCalibration), buffer);
-        } else {
-            result = ReadSPI(CalAddr::FACT_IMU_DATA, sizeof(IMUCalibration), buffer);
     }
+
+    // Read User defined calibration
+    if (result == DriverResult::Success && has_user_calibration) {
+        result = ReadSPI(SpiAddress::USER_IMU_DATA, spi_calibration);
+    }
+
+    // Read Factory calibration
+    if (result == DriverResult::Success && !has_user_calibration) {
+        result = ReadSPI(SpiAddress::FACT_IMU_DATA, spi_calibration);
     }
 
     if (result == DriverResult::Success) {
-        IMUCalibration device_calibration{};
+        calibration.accelerometer[0].offset = spi_calibration.accelerometer_offset[0];
-        memcpy(&device_calibration, buffer.data(), sizeof(IMUCalibration));
+        calibration.accelerometer[1].offset = spi_calibration.accelerometer_offset[1];
-        calibration.accelerometer[0].offset = device_calibration.accelerometer_offset[0];
+        calibration.accelerometer[2].offset = spi_calibration.accelerometer_offset[2];
-        calibration.accelerometer[1].offset = device_calibration.accelerometer_offset[1];
-        calibration.accelerometer[2].offset = device_calibration.accelerometer_offset[2];
 
-        calibration.accelerometer[0].scale = device_calibration.accelerometer_scale[0];
+        calibration.accelerometer[0].scale = spi_calibration.accelerometer_scale[0];
-        calibration.accelerometer[1].scale = device_calibration.accelerometer_scale[1];
+        calibration.accelerometer[1].scale = spi_calibration.accelerometer_scale[1];
-        calibration.accelerometer[2].scale = device_calibration.accelerometer_scale[2];
+        calibration.accelerometer[2].scale = spi_calibration.accelerometer_scale[2];
 
-        calibration.gyro[0].offset = device_calibration.gyroscope_offset[0];
+        calibration.gyro[0].offset = spi_calibration.gyroscope_offset[0];
-        calibration.gyro[1].offset = device_calibration.gyroscope_offset[1];
+        calibration.gyro[1].offset = spi_calibration.gyroscope_offset[1];
-        calibration.gyro[2].offset = device_calibration.gyroscope_offset[2];
+        calibration.gyro[2].offset = spi_calibration.gyroscope_offset[2];
 
-        calibration.gyro[0].scale = device_calibration.gyroscope_scale[0];
+        calibration.gyro[0].scale = spi_calibration.gyroscope_scale[0];
-        calibration.gyro[1].scale = device_calibration.gyroscope_scale[1];
+        calibration.gyro[1].scale = spi_calibration.gyroscope_scale[1];
-        calibration.gyro[2].scale = device_calibration.gyroscope_scale[2];
+        calibration.gyro[2].scale = spi_calibration.gyroscope_scale[2];
     }
 
     ValidateCalibration(calibration);
@@ -127,10 +129,12 @@ DriverResult CalibrationProtocol::GetImuCalibration(MotionCalibration& calibrati
 
 DriverResult CalibrationProtocol::GetRingCalibration(RingCalibration& calibration,
                                                      s16 current_value) {
+    constexpr s16 DefaultRingRange{800};
+
     // TODO: Get default calibration form ring itself
     if (ring_data_max == 0 && ring_data_min == 0) {
-        ring_data_max = current_value + 800;
+        ring_data_max = current_value + DefaultRingRange;
-        ring_data_min = current_value - 800;
+        ring_data_min = current_value - DefaultRingRange;
         ring_data_default = current_value;
     }
     ring_data_max = std::max(ring_data_max, current_value);
@@ -143,42 +147,72 @@ DriverResult CalibrationProtocol::GetRingCalibration(RingCalibration& calibratio
     return DriverResult::Success;
 }
 
+DriverResult CalibrationProtocol::HasUserCalibration(SpiAddress address,
+                                                     bool& has_user_calibration) {
+    MagicSpiCalibration spi_magic{};
+    const DriverResult result{ReadSPI(address, spi_magic)};
+    has_user_calibration = false;
+    if (result == DriverResult::Success) {
+        has_user_calibration = spi_magic.first == CalibrationMagic::USR_MAGIC_0 &&
+                               spi_magic.second == CalibrationMagic::USR_MAGIC_1;
+    }
+    return result;
+}
+
+u16 CalibrationProtocol::GetXAxisCalibrationValue(std::span<u8> block) const {
+    return static_cast<u16>(((block[1] & 0x0F) << 8) | block[0]);
+}
+
+u16 CalibrationProtocol::GetYAxisCalibrationValue(std::span<u8> block) const {
+    return static_cast<u16>((block[2] << 4) | (block[1] >> 4));
+}
+
 void CalibrationProtocol::ValidateCalibration(JoyStickCalibration& calibration) {
-    constexpr u16 DefaultStickCenter{2048};
+    constexpr u16 DefaultStickCenter{0x800};
-    constexpr u16 DefaultStickRange{1740};
+    constexpr u16 DefaultStickRange{0x6cc};
 
-    if (calibration.x.center == 0xFFF || calibration.x.center == 0) {
+    calibration.x.center = ValidateValue(calibration.x.center, DefaultStickCenter);
-        calibration.x.center = DefaultStickCenter;
+    calibration.x.max = ValidateValue(calibration.x.max, DefaultStickRange);
-    }
+    calibration.x.min = ValidateValue(calibration.x.min, DefaultStickRange);
-    if (calibration.x.max == 0xFFF || calibration.x.max == 0) {
-        calibration.x.max = DefaultStickRange;
-    }
-    if (calibration.x.min == 0xFFF || calibration.x.min == 0) {
-        calibration.x.min = DefaultStickRange;
-    }
 
-    if (calibration.y.center == 0xFFF || calibration.y.center == 0) {
+    calibration.y.center = ValidateValue(calibration.y.center, DefaultStickCenter);
-        calibration.y.center = DefaultStickCenter;
+    calibration.y.max = ValidateValue(calibration.y.max, DefaultStickRange);
-    }
+    calibration.y.min = ValidateValue(calibration.y.min, DefaultStickRange);
-    if (calibration.y.max == 0xFFF || calibration.y.max == 0) {
-        calibration.y.max = DefaultStickRange;
-    }
-    if (calibration.y.min == 0xFFF || calibration.y.min == 0) {
-        calibration.y.min = DefaultStickRange;
-    }
 }
 
 void CalibrationProtocol::ValidateCalibration(MotionCalibration& calibration) {
+    constexpr s16 DefaultAccelerometerScale{0x4000};
+    constexpr s16 DefaultGyroScale{0x3be7};
+    constexpr s16 DefaultOffset{0};
+
     for (auto& sensor : calibration.accelerometer) {
-        if (sensor.scale == 0) {
+        sensor.scale = ValidateValue(sensor.scale, DefaultAccelerometerScale);
-            sensor.scale = 0x4000;
+        sensor.offset = ValidateValue(sensor.offset, DefaultOffset);
-        }
     }
     for (auto& sensor : calibration.gyro) {
-        if (sensor.scale == 0) {
+        sensor.scale = ValidateValue(sensor.scale, DefaultGyroScale);
-            sensor.scale = 0x3be7;
+        sensor.offset = ValidateValue(sensor.offset, DefaultOffset);
-        }
     }
 }
 
+u16 CalibrationProtocol::ValidateValue(u16 value, u16 default_value) const {
+    if (value == 0) {
+        return default_value;
+    }
+    if (value == 0xFFF) {
+        return default_value;
+    }
+    return value;
+}
+
+s16 CalibrationProtocol::ValidateValue(s16 value, s16 default_value) const {
+    if (value == 0) {
+        return default_value;
+    }
+    if (value == 0xFFF) {
+        return default_value;
+    }
+    return value;
+}
+
 } // namespace InputCommon::Joycon

src/input_common/helpers/joycon_protocol/calibration.h

@@ -53,9 +53,27 @@ public:
     DriverResult GetRingCalibration(RingCalibration& calibration, s16 current_value);
 
 private:
+    /// Returns true if the specified address corresponds to the magic value of user calibration
+    DriverResult HasUserCalibration(SpiAddress address, bool& has_user_calibration);
+
+    /// Converts a raw calibration block to an u16 value containing the x axis value
+    u16 GetXAxisCalibrationValue(std::span<u8> block) const;
+
+    /// Converts a raw calibration block to an u16 value containing the y axis value
+    u16 GetYAxisCalibrationValue(std::span<u8> block) const;
+
+    /// Ensures that all joystick calibration values are set
     void ValidateCalibration(JoyStickCalibration& calibration);
+
+    /// Ensures that all motion calibration values are set
     void ValidateCalibration(MotionCalibration& calibration);
 
+    /// Returns the default value if the value is either zero or 0xFFF
+    u16 ValidateValue(u16 value, u16 default_value) const;
+
+    /// Returns the default value if the value is either zero or 0xFFF
+    s16 ValidateValue(s16 value, s16 default_value) const;
+
     s16 ring_data_max = 0;
     s16 ring_data_default = 0;
     s16 ring_data_min = 0;

src/input_common/helpers/joycon_protocol/common_protocol.cpp

@@ -22,8 +22,8 @@ void JoyconCommonProtocol::SetNonBlocking() {
 }
 
 DriverResult JoyconCommonProtocol::GetDeviceType(ControllerType& controller_type) {
-    std::vector<u8> buffer;
+    std::array<u8, 1> buffer{};
-    const auto result = ReadSPI(CalAddr::DEVICE_TYPE, 1, buffer);
+    const auto result = ReadRawSPI(SpiAddress::DEVICE_TYPE, buffer);
     controller_type = ControllerType::None;
 
     if (result == DriverResult::Success) {
@@ -148,11 +148,13 @@ DriverResult JoyconCommonProtocol::SendVibrationReport(std::span<const u8> buffe
     return SendData(local_buffer);
 }
 
-DriverResult JoyconCommonProtocol::ReadSPI(CalAddr addr, u8 size, std::vector<u8>& output) {
+DriverResult JoyconCommonProtocol::ReadRawSPI(SpiAddress addr, std::span<u8> output) {
+    constexpr std::size_t HeaderSize = 20;
     constexpr std::size_t MaxTries = 10;
+    const auto size = output.size();
     std::size_t tries = 0;
-    std::array<u8, 5> buffer = {0x00, 0x00, 0x00, 0x00, size};
+    std::array<u8, 5> buffer = {0x00, 0x00, 0x00, 0x00, static_cast<u8>(size)};
-    std::vector<u8> local_buffer(size + 20);
+    std::vector<u8> local_buffer{};
 
     buffer[0] = static_cast<u8>(static_cast<u16>(addr) & 0x00FF);
     buffer[1] = static_cast<u8>((static_cast<u16>(addr) & 0xFF00) >> 8);
@@ -167,8 +169,12 @@ DriverResult JoyconCommonProtocol::ReadSPI(CalAddr addr, u8 size, std::vector<u8
         }
     } while (local_buffer[15] != buffer[0] || local_buffer[16] != buffer[1]);
 
+    if (local_buffer.size() < size + HeaderSize) {
+        return DriverResult::WrongReply;
+    }
+
     // Remove header from output
-    output = std::vector<u8>(local_buffer.begin() + 20, local_buffer.begin() + 20 + size);
+    memcpy(output.data(), local_buffer.data() + HeaderSize, size);
     return DriverResult::Success;
 }
 

src/input_common/helpers/joycon_protocol/common_protocol.h

@@ -97,10 +97,29 @@ public:
     /**
      * Reads the SPI memory stored on the joycon
      * @param Initial address location
-     * @param size in bytes to be read
      * @returns output buffer containing the responce
      */
-    DriverResult ReadSPI(CalAddr addr, u8 size, std::vector<u8>& output);
+    DriverResult ReadRawSPI(SpiAddress addr, std::span<u8> output);
+
+    /**
+     * Reads the SPI memory stored on the joycon
+     * @param Initial address location
+     * @returns output object containing the responce
+     */
+    template <typename Output>
+    requires std::is_trivially_copyable_v<Output> DriverResult ReadSPI(SpiAddress addr,
+                                                                       Output& output) {
+        std::array<u8, sizeof(Output)> buffer;
+        output = {};
+
+        const auto result = ReadRawSPI(addr, buffer);
+        if (result != DriverResult::Success) {
+            return result;
+        }
+
+        std::memcpy(&output, buffer.data(), sizeof(Output));
+        return DriverResult::Success;
+    }
 
     /**
      * Enables MCU chip on the joycon

src/input_common/helpers/joycon_protocol/generic_functions.cpp

@@ -71,8 +71,8 @@ DriverResult GenericProtocol::GetBattery(u32& battery_level) {
 
 DriverResult GenericProtocol::GetColor(Color& color) {
     ScopedSetBlocking sb(this);
-    std::vector<u8> buffer;
+    std::array<u8, 12> buffer{};
-    const auto result = ReadSPI(CalAddr::COLOR_DATA, 12, buffer);
+    const auto result = ReadRawSPI(SpiAddress::COLOR_DATA, buffer);
 
     color = {};
     if (result == DriverResult::Success) {
@@ -87,8 +87,8 @@ DriverResult GenericProtocol::GetColor(Color& color) {
 
 DriverResult GenericProtocol::GetSerialNumber(SerialNumber& serial_number) {
     ScopedSetBlocking sb(this);
-    std::vector<u8> buffer;
+    std::array<u8, 16> buffer{};
-    const auto result = ReadSPI(CalAddr::SERIAL_NUMBER, 16, buffer);
+    const auto result = ReadRawSPI(SpiAddress::SERIAL_NUMBER, buffer);
 
     serial_number = {};
     if (result == DriverResult::Success) {

src/input_common/helpers/joycon_protocol/joycon_types.h

@@ -159,13 +159,12 @@ enum class UsbSubCommand : u8 {
     SEND_UART = 0x92,
 };
 
-enum class CalMagic : u8 {
+enum class CalibrationMagic : u8 {
     USR_MAGIC_0 = 0xB2,
     USR_MAGIC_1 = 0xA1,
-    USRR_MAGI_SIZE = 2,
 };
 
-enum class CalAddr {
+enum class SpiAddress {
     SERIAL_NUMBER = 0X6000,
     DEVICE_TYPE = 0X6012,
     COLOR_EXIST = 0X601B,
@@ -396,10 +395,35 @@ struct MotionData {
     u64 delta_timestamp{};
 };
 
+// Output from SPI read command containing user calibration magic
+struct MagicSpiCalibration {
+    CalibrationMagic first;
+    CalibrationMagic second;
+};
+static_assert(sizeof(MagicSpiCalibration) == 0x2, "MagicSpiCalibration is an invalid size");
+
+// Output from SPI read command containing left joystick calibration
+struct JoystickLeftSpiCalibration {
+    std::array<u8, 3> max;
+    std::array<u8, 3> center;
+    std::array<u8, 3> min;
+};
+static_assert(sizeof(JoystickLeftSpiCalibration) == 0x9,
+              "JoystickLeftSpiCalibration is an invalid size");
+
+// Output from SPI read command containing right joystick calibration
+struct JoystickRightSpiCalibration {
+    std::array<u8, 3> center;
+    std::array<u8, 3> min;
+    std::array<u8, 3> max;
+};
+static_assert(sizeof(JoystickRightSpiCalibration) == 0x9,
+              "JoystickRightSpiCalibration is an invalid size");
+
 struct JoyStickAxisCalibration {
-    u16 max{1};
+    u16 max;
-    u16 min{1};
+    u16 min;
-    u16 center{0};
+    u16 center;
 };
 
 struct JoyStickCalibration {
@@ -407,6 +431,14 @@ struct JoyStickCalibration {
     JoyStickAxisCalibration y;
 };
 
+struct ImuSpiCalibration {
+    std::array<s16, 3> accelerometer_offset;
+    std::array<s16, 3> accelerometer_scale;
+    std::array<s16, 3> gyroscope_offset;
+    std::array<s16, 3> gyroscope_scale;
+};
+static_assert(sizeof(ImuSpiCalibration) == 0x18, "ImuSpiCalibration is an invalid size");
+
 struct RingCalibration {
     s16 default_value;
     s16 max_value;
@@ -488,14 +520,6 @@ struct InputReportNfcIr {
 static_assert(sizeof(InputReportNfcIr) == 0x29, "InputReportNfcIr is an invalid size");
 #pragma pack(pop)
 
-struct IMUCalibration {
-    std::array<s16, 3> accelerometer_offset;
-    std::array<s16, 3> accelerometer_scale;
-    std::array<s16, 3> gyroscope_offset;
-    std::array<s16, 3> gyroscope_scale;
-};
-static_assert(sizeof(IMUCalibration) == 0x18, "IMUCalibration is an invalid size");
-
 struct NFCReadBlock {
     u8 start;
     u8 end;

src/shader_recompiler/backend/glasm/emit_glasm_image.cpp

@@ -532,7 +532,7 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
 }
 
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              ScalarS32 lod) {
+                              ScalarS32 lod, [[maybe_unused]] const IR::Value& skip_mips) {
     const auto info{inst.Flags<IR::TextureInstInfo>()};
     const std::string texture{Texture(ctx, info, index)};
     const std::string_view type{TextureType(info)};

src/shader_recompiler/backend/glasm/emit_glasm_instructions.h

@@ -581,7 +581,7 @@ void EmitImageGatherDref(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
 void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                     const IR::Value& coord, const IR::Value& offset, ScalarS32 lod, ScalarS32 ms);
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              ScalarS32 lod);
+                              ScalarS32 lod, const IR::Value& skip_mips);
 void EmitImageQueryLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index, Register coord);
 void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                        const IR::Value& coord, const IR::Value& derivatives,

src/shader_recompiler/backend/glsl/emit_glsl_image.cpp

@@ -414,7 +414,7 @@ void EmitImageGatherDref(EmitContext& ctx, IR::Inst& inst, const IR::Value& inde
 
 void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                     std::string_view coords, std::string_view offset, std::string_view lod,
-                    [[maybe_unused]] std::string_view ms) {
+                    std::string_view ms) {
     const auto info{inst.Flags<IR::TextureInstInfo>()};
     if (info.has_bias) {
         throw NotImplementedException("EmitImageFetch Bias texture samples");
@@ -431,19 +431,24 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
         ctx.AddU1("{}=true;", *sparse_inst);
     }
     if (!sparse_inst || !supports_sparse) {
-        if (!offset.empty()) {
+        const auto int_coords{CoordsCastToInt(coords, info)};
-            ctx.Add("{}=texelFetchOffset({},{},int({}),{});", texel, texture,
+        if (!ms.empty()) {
-                    CoordsCastToInt(coords, info), lod, CoordsCastToInt(offset, info));
+            ctx.Add("{}=texelFetch({},{},int({}));", texel, texture, int_coords, ms);
+        } else if (!offset.empty()) {
+            ctx.Add("{}=texelFetchOffset({},{},int({}),{});", texel, texture, int_coords, lod,
+                    CoordsCastToInt(offset, info));
         } else {
             if (info.type == TextureType::Buffer) {
                 ctx.Add("{}=texelFetch({},int({}));", texel, texture, coords);
             } else {
-                ctx.Add("{}=texelFetch({},{},int({}));", texel, texture,
+                ctx.Add("{}=texelFetch({},{},int({}));", texel, texture, int_coords, lod);
-                        CoordsCastToInt(coords, info), lod);
             }
         }
         return;
     }
+    if (!ms.empty()) {
+        throw NotImplementedException("EmitImageFetch Sparse MSAA samples");
+    }
     if (!offset.empty()) {
         ctx.AddU1("{}=sparseTexelsResidentARB(sparseTexelFetchOffsetARB({},{},int({}),{},{}));",
                   *sparse_inst, texture, CastToIntVec(coords, info), lod,
@@ -455,27 +460,27 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
 }
 
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              std::string_view lod) {
+                              std::string_view lod, const IR::Value& skip_mips_val) {
     const auto info{inst.Flags<IR::TextureInstInfo>()};
     const auto texture{Texture(ctx, info, index)};
+    const bool skip_mips{skip_mips_val.U1()};
+    const auto mips{
+        [&] { return skip_mips ? "0u" : fmt::format("uint(textureQueryLevels({}))", texture); }};
     switch (info.type) {
     case TextureType::Color1D:
-        return ctx.AddU32x4(
+        return ctx.AddU32x4("{}=uvec4(uint(textureSize({},int({}))),0u,0u,{});", inst, texture, lod,
-            "{}=uvec4(uint(textureSize({},int({}))),0u,0u,uint(textureQueryLevels({})));", inst,
+                            mips());
-            texture, lod, texture);
     case TextureType::ColorArray1D:
     case TextureType::Color2D:
     case TextureType::ColorCube:
     case TextureType::Color2DRect:
-        return ctx.AddU32x4(
+        return ctx.AddU32x4("{}=uvec4(uvec2(textureSize({},int({}))),0u,{});", inst, texture, lod,
-            "{}=uvec4(uvec2(textureSize({},int({}))),0u,uint(textureQueryLevels({})));", inst,
+                            mips());
-            texture, lod, texture);
     case TextureType::ColorArray2D:
     case TextureType::Color3D:
     case TextureType::ColorArrayCube:
-        return ctx.AddU32x4(
+        return ctx.AddU32x4("{}=uvec4(uvec3(textureSize({},int({}))),{});", inst, texture, lod,
-            "{}=uvec4(uvec3(textureSize({},int({}))),uint(textureQueryLevels({})));", inst, texture,
+                            mips());
-            lod, texture);
     case TextureType::Buffer:
         throw NotImplementedException("EmitImageQueryDimensions Texture buffers");
     }

src/shader_recompiler/backend/glsl/emit_glsl_instructions.h

@@ -654,7 +654,7 @@ void EmitImageFetch(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                     std::string_view coords, std::string_view offset, std::string_view lod,
                     std::string_view ms);
 void EmitImageQueryDimensions(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
-                              std::string_view lod);
+                              std::string_view lod, const IR::Value& skip_mips);
 void EmitImageQueryLod(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,
                        std::string_view coords);
 void EmitImageGradient(EmitContext& ctx, IR::Inst& inst, const IR::Value& index,

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -445,11 +445,13 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id c
                 TextureImage(ctx, info, index), coords, operands.MaskOptional(), operands.Span());
 }
 
-Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod) {
+Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod,
+                            const IR::Value& skip_mips_val) {
     const auto info{inst->Flags<IR::TextureInstInfo>()};
     const Id image{TextureImage(ctx, info, index)};
     const Id zero{ctx.u32_zero_value};
-    const auto mips{[&] { return ctx.OpImageQueryLevels(ctx.U32[1], image); }};
+    const bool skip_mips{skip_mips_val.U1()};
+    const auto mips{[&] { return skip_mips ? zero : ctx.OpImageQueryLevels(ctx.U32[1], image); }};
     switch (info.type) {
     case TextureType::Color1D:
         return ctx.OpCompositeConstruct(ctx.U32[4], ctx.OpImageQuerySizeLod(ctx.U32[1], image, lod),

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -539,7 +539,8 @@ Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, const IR::Value& index,
                        const IR::Value& offset, const IR::Value& offset2, Id dref);
 Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords, Id offset,
                   Id lod, Id ms);
-Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod);
+Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id lod,
+                            const IR::Value& skip_mips);
 Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
 Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
                      Id derivates, Id offset, Id lod_clamp);

src/shader_recompiler/frontend/ir/ir_emitter.cpp

@@ -1846,15 +1846,16 @@ Value IREmitter::ImageFetch(const Value& handle, const Value& coords, const Valu
     return Inst(op, Flags{info}, handle, coords, offset, lod, multisampling);
 }
 
-Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod) {
+Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
+                                     const IR::U1& skip_mips) {
     const Opcode op{handle.IsImmediate() ? Opcode::BoundImageQueryDimensions
                                          : Opcode::BindlessImageQueryDimensions};
-    return Inst(op, handle, lod);
+    return Inst(op, handle, lod, skip_mips);
 }
 
 Value IREmitter::ImageQueryDimension(const Value& handle, const IR::U32& lod,
-                                     TextureInstInfo info) {
+                                     const IR::U1& skip_mips, TextureInstInfo info) {
-    return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod);
+    return Inst(Opcode::ImageQueryDimensions, Flags{info}, handle, lod, skip_mips);
 }
 
 Value IREmitter::ImageQueryLod(const Value& handle, const Value& coords, TextureInstInfo info) {

src/shader_recompiler/frontend/ir/ir_emitter.h

@@ -320,9 +320,10 @@ public:
     [[nodiscard]] F32 ImageSampleDrefExplicitLod(const Value& handle, const Value& coords,
                                                  const F32& dref, const F32& lod,
                                                  const Value& offset, TextureInstInfo info);
-    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod);
     [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
-                                            TextureInstInfo info);
+                                            const IR::U1& skip_mips);
+    [[nodiscard]] Value ImageQueryDimension(const Value& handle, const IR::U32& lod,
+                                            const IR::U1& skip_mips, TextureInstInfo info);
 
     [[nodiscard]] Value ImageQueryLod(const Value& handle, const Value& coords,
                                       TextureInstInfo info);

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -482,7 +482,7 @@ OPCODE(BindlessImageSampleDrefExplicitLod,                  F32,            U32,
 OPCODE(BindlessImageGather,                                 F32x4,          U32,            Opaque,         Opaque,         Opaque,                         )
 OPCODE(BindlessImageGatherDref,                             F32x4,          U32,            Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(BindlessImageFetch,                                  F32x4,          U32,            Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(BindlessImageQueryDimensions,                        U32x4,          U32,            U32,                                                            )
+OPCODE(BindlessImageQueryDimensions,                        U32x4,          U32,            U32,            U1,                                             )
 OPCODE(BindlessImageQueryLod,                               F32x4,          U32,            Opaque,                                                         )
 OPCODE(BindlessImageGradient,                               F32x4,          U32,            Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(BindlessImageRead,                                   U32x4,          U32,            Opaque,                                                         )
@@ -495,7 +495,7 @@ OPCODE(BoundImageSampleDrefExplicitLod,                     F32,            U32,
 OPCODE(BoundImageGather,                                    F32x4,          U32,            Opaque,         Opaque,         Opaque,                         )
 OPCODE(BoundImageGatherDref,                                F32x4,          U32,            Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(BoundImageFetch,                                     F32x4,          U32,            Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(BoundImageQueryDimensions,                           U32x4,          U32,            U32,                                                            )
+OPCODE(BoundImageQueryDimensions,                           U32x4,          U32,            U32,            U1,                                             )
 OPCODE(BoundImageQueryLod,                                  F32x4,          U32,            Opaque,                                                         )
 OPCODE(BoundImageGradient,                                  F32x4,          U32,            Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(BoundImageRead,                                      U32x4,          U32,            Opaque,                                                         )
@@ -508,7 +508,7 @@ OPCODE(ImageSampleDrefExplicitLod,                          F32,            Opaq
 OPCODE(ImageGather,                                         F32x4,          Opaque,         Opaque,         Opaque,         Opaque,                         )
 OPCODE(ImageGatherDref,                                     F32x4,          Opaque,         Opaque,         Opaque,         Opaque,         F32,            )
 OPCODE(ImageFetch,                                          F32x4,          Opaque,         Opaque,         Opaque,         U32,            Opaque,         )
-OPCODE(ImageQueryDimensions,                                U32x4,          Opaque,         U32,                                                            )
+OPCODE(ImageQueryDimensions,                                U32x4,          Opaque,         U32,            U1,                                             )
 OPCODE(ImageQueryLod,                                       F32x4,          Opaque,         Opaque,                                                         )
 OPCODE(ImageGradient,                                       F32x4,          Opaque,         Opaque,         Opaque,         Opaque,         Opaque,         )
 OPCODE(ImageRead,                                           U32x4,          Opaque,         Opaque,                                                         )

src/shader_recompiler/frontend/maxwell/translate/impl/texture_query.cpp

@@ -15,11 +15,13 @@ enum class Mode : u64 {
     SamplePos = 5,
 };
 
-IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg) {
+IR::Value Query(TranslatorVisitor& v, const IR::U32& handle, Mode mode, IR::Reg src_reg, u64 mask) {
     switch (mode) {
     case Mode::Dimension: {
+        const bool needs_num_mips{((mask >> 3) & 1) != 0};
+        const IR::U1 skip_mips{v.ir.Imm1(!needs_num_mips)};
         const IR::U32 lod{v.X(src_reg)};
-        return v.ir.ImageQueryDimension(handle, lod);
+        return v.ir.ImageQueryDimension(handle, lod, skip_mips);
     }
     case Mode::TextureType:
     case Mode::SamplePos:
@@ -46,7 +48,7 @@ void Impl(TranslatorVisitor& v, u64 insn, std::optional<u32> cbuf_offset) {
         handle = v.X(src_reg);
         ++src_reg;
     }
-    const IR::Value query{Query(v, handle, txq.mode, src_reg)};
+    const IR::Value query{Query(v, handle, txq.mode, src_reg, txq.mask)};
     IR::Reg dest_reg{txq.dest_reg};
     for (int element = 0; element < 4; ++element) {
         if (((txq.mask >> element) & 1) == 0) {

src/shader_recompiler/ir_opt/texture_pass.cpp

@@ -452,7 +452,8 @@ void PatchImageSampleImplicitLod(IR::Block& block, IR::Inst& inst) {
     const IR::Value coord(inst.Arg(1));
     const IR::Value handle(ir.Imm32(0));
     const IR::U32 lod{ir.Imm32(0)};
-    const IR::Value texture_size = ir.ImageQueryDimension(handle, lod, info);
+    const IR::U1 skip_mips{ir.Imm1(true)};
+    const IR::Value texture_size = ir.ImageQueryDimension(handle, lod, skip_mips, info);
     inst.SetArg(
         1, ir.CompositeConstruct(
                ir.FPMul(IR::F32(ir.CompositeExtract(coord, 0)),