Add more NVML and cuBLAS coverage (#481)

Cargo.lock

@@ -455,6 +455,7 @@ dependencies = [
  "cuda_macros",
  "hip_runtime-sys",
  "rocblas-sys",
+ "rocm_smi-sys",
 ]
 
 [[package]]
@@ -2737,6 +2738,10 @@ dependencies = [
  "hip_runtime-sys",
 ]
 
+[[package]]
+name = "rocm_smi-sys"
+version = "0.0.0"
+
 [[package]]
 name = "rustc-hash"
 version = "1.1.0"
@@ -3750,6 +3755,7 @@ version = "0.0.0"
 dependencies = [
  "cuda_macros",
  "cuda_types",
+ "hip_runtime-sys",
  "rocblas-sys",
  "zluda_common",
 ]
@@ -3817,6 +3823,8 @@ version = "0.0.0"
 dependencies = [
  "cuda_macros",
  "cuda_types",
+ "rocm_smi-sys",
+ "zluda_common",
 ]
 
 [[package]]

cuda_macros/src/nvml.rs

@@ -8498,4 +8498,125 @@ returned in \a sessionCount
         device: cuda_types::nvml::nvmlDevice_t,
         state: *mut cuda_types::nvml::nvmlPowerSmoothingState_t,
     ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlInit() -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetCount(
+        deviceCount: *mut ::core::ffi::c_uint,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetHandleByIndex(
+        index: ::core::ffi::c_uint,
+        device: *mut cuda_types::nvml::nvmlDevice_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetHandleByPciBusId(
+        pciBusId: *const ::core::ffi::c_char,
+        device: *mut cuda_types::nvml::nvmlDevice_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetPciInfo(
+        device: cuda_types::nvml::nvmlDevice_t,
+        pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetPciInfo_v2(
+        device: cuda_types::nvml::nvmlDevice_t,
+        pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetNvLinkRemotePciInfo(
+        device: cuda_types::nvml::nvmlDevice_t,
+        link: ::core::ffi::c_uint,
+        pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGridLicensableFeatures(
+        device: cuda_types::nvml::nvmlDevice_t,
+        pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGridLicensableFeatures_v2(
+        device: cuda_types::nvml::nvmlDevice_t,
+        pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGridLicensableFeatures_v3(
+        device: cuda_types::nvml::nvmlDevice_t,
+        pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceRemoveGpu(
+        pciInfo: *mut cuda_types::nvml::nvmlPciInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlEventSetWait(
+        set: cuda_types::nvml::nvmlEventSet_t,
+        data: *mut cuda_types::nvml::nvmlEventData_t,
+        timeoutms: ::core::ffi::c_uint,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetAttributes(
+        device: cuda_types::nvml::nvmlDevice_t,
+        attributes: *mut cuda_types::nvml::nvmlDeviceAttributes_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlComputeInstanceGetInfo(
+        computeInstance: cuda_types::nvml::nvmlComputeInstance_t,
+        info: *mut cuda_types::nvml::nvmlComputeInstanceInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetComputeRunningProcesses(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetComputeRunningProcesses_v2(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGraphicsRunningProcesses(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGraphicsRunningProcesses_v2(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetMPSComputeRunningProcesses(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetMPSComputeRunningProcesses_v2(
+        device: cuda_types::nvml::nvmlDevice_t,
+        infoCount: *mut ::core::ffi::c_uint,
+        infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetGpuInstancePossiblePlacements(
+        device: cuda_types::nvml::nvmlDevice_t,
+        profileId: ::core::ffi::c_uint,
+        placements: *mut cuda_types::nvml::nvmlGpuInstancePlacement_t,
+        count: *mut ::core::ffi::c_uint,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlVgpuInstanceGetLicenseInfo(
+        vgpuInstance: cuda_types::nvml::nvmlVgpuInstance_t,
+        licenseInfo: *mut cuda_types::nvml::nvmlVgpuLicenseInfo_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
+    #[must_use]
+    fn nvmlDeviceGetDriverModel(
+        device: cuda_types::nvml::nvmlDevice_t,
+        current: *mut cuda_types::nvml::nvmlDriverModel_t,
+        pending: *mut cuda_types::nvml::nvmlDriverModel_t,
+    ) -> cuda_types::nvml::nvmlReturn_t;
 }

cuda_types/Cargo.toml

@@ -9,3 +9,6 @@ cuda_macros = { path = "../cuda_macros" }
 hip_runtime-sys = { path = "../ext/hip_runtime-sys" }
 bitflags = "2.9.1"
 rocblas-sys = { path = "../ext/rocblas-sys" }
+
+[target.'cfg(unix)'.dependencies]
+rocm_smi-sys = { path = "../ext/rocm_smi-sys" }

cuda_types/src/nvml.rs

@@ -4879,3 +4879,20 @@ pub type nvmlReturn_t = ::core::result::Result<(), nvmlError_t>;
 const _: fn() = || {
     let _ = std::mem::transmute::<nvmlReturn_t, u32>;
 };
+#[cfg(unix)]
+impl From<rocm_smi_sys::rsmi_error> for nvmlError_t {
+    fn from(error: rocm_smi_sys::rsmi_error) -> Self {
+        match error {
+            rocm_smi_sys::rsmi_error::INVALID_ARGS => nvmlError_t::from(nvmlError_t::INVALID_ARGUMENT),
+            rocm_smi_sys::rsmi_error::NOT_SUPPORTED => nvmlError_t::from(nvmlError_t::NOT_SUPPORTED),
+            rocm_smi_sys::rsmi_error::PERMISSION => nvmlError_t::from(nvmlError_t::NO_PERMISSION),
+            rocm_smi_sys::rsmi_error::INPUT_OUT_OF_BOUNDS => nvmlError_t::from(nvmlError_t::INVALID_ARGUMENT),
+            rocm_smi_sys::rsmi_error::INIT_ERROR => nvmlError_t::from(nvmlError_t::UNINITIALIZED),
+            rocm_smi_sys::rsmi_error::NOT_FOUND => nvmlError_t::from(nvmlError_t::GPU_NOT_FOUND),
+            rocm_smi_sys::rsmi_error::INSUFFICIENT_SIZE => nvmlError_t::from(nvmlError_t::INSUFFICIENT_SIZE),
+            rocm_smi_sys::rsmi_error::INTERRUPT => nvmlError_t::from(nvmlError_t::IRQ_ISSUE),
+            rocm_smi_sys::rsmi_error::NO_DATA => nvmlError_t::from(nvmlError_t::NO_DATA),
+            _ => nvmlError_t::from(nvmlError_t::UNKNOWN),
+        }
+    }
+}

ext/rocm_smi-sys/Cargo.toml

@@ -0,0 +1,7 @@
+[package]
+name = "rocm_smi-sys"
+version = "0.0.0"
+authors = ["Andrzej Janik <vosen@vosen.pl>"]
+edition = "2021"
+
+[lib]

format/src/format_generated_nvml.rs

@@ -13103,6 +13103,566 @@ pub fn write_nvmlDevicePowerSmoothingSetState(
     )?;
     writer.write_all(b")")
 }
+pub fn write_nvmlInit(
+    writer: &mut (impl std::io::Write + ?Sized),
+) -> std::io::Result<()> {
+    writer.write_all(b"()")
+}
+pub fn write_nvmlDeviceGetCount(
+    writer: &mut (impl std::io::Write + ?Sized),
+    deviceCount: *mut ::core::ffi::c_uint,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(deviceCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&deviceCount, "nvmlDeviceGetCount", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetHandleByIndex(
+    writer: &mut (impl std::io::Write + ?Sized),
+    index: ::core::ffi::c_uint,
+    device: *mut cuda_types::nvml::nvmlDevice_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(index), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&index, "nvmlDeviceGetHandleByIndex", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&device, "nvmlDeviceGetHandleByIndex", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetHandleByPciBusId(
+    writer: &mut (impl std::io::Write + ?Sized),
+    pciBusId: *const ::core::ffi::c_char,
+    device: *mut cuda_types::nvml::nvmlDevice_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(pciBusId), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &pciBusId,
+        "nvmlDeviceGetHandleByPciBusId",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetHandleByPciBusId",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetPciInfo(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&device, "nvmlDeviceGetPciInfo", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pci), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&pci, "nvmlDeviceGetPciInfo", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetPciInfo_v2(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&device, "nvmlDeviceGetPciInfo_v2", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pci), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&pci, "nvmlDeviceGetPciInfo_v2", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetNvLinkRemotePciInfo(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    link: ::core::ffi::c_uint,
+    pci: *mut cuda_types::nvml::nvmlPciInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetNvLinkRemotePciInfo",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(link), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &link,
+        "nvmlDeviceGetNvLinkRemotePciInfo",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pci), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &pci,
+        "nvmlDeviceGetNvLinkRemotePciInfo",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGridLicensableFeatures(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGridLicensableFeatures",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pGridLicensableFeatures), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &pGridLicensableFeatures,
+        "nvmlDeviceGetGridLicensableFeatures",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGridLicensableFeatures_v2(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGridLicensableFeatures_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pGridLicensableFeatures), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &pGridLicensableFeatures,
+        "nvmlDeviceGetGridLicensableFeatures_v2",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGridLicensableFeatures_v3(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    pGridLicensableFeatures: *mut cuda_types::nvml::nvmlGridLicensableFeatures_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGridLicensableFeatures_v3",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pGridLicensableFeatures), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &pGridLicensableFeatures,
+        "nvmlDeviceGetGridLicensableFeatures_v3",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceRemoveGpu(
+    writer: &mut (impl std::io::Write + ?Sized),
+    pciInfo: *mut cuda_types::nvml::nvmlPciInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(pciInfo), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&pciInfo, "nvmlDeviceRemoveGpu", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlEventSetWait(
+    writer: &mut (impl std::io::Write + ?Sized),
+    set: cuda_types::nvml::nvmlEventSet_t,
+    data: *mut cuda_types::nvml::nvmlEventData_t,
+    timeoutms: ::core::ffi::c_uint,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(set), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&set, "nvmlEventSetWait", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(data), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&data, "nvmlEventSetWait", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(timeoutms), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&timeoutms, "nvmlEventSetWait", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetAttributes(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    attributes: *mut cuda_types::nvml::nvmlDeviceAttributes_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&device, "nvmlDeviceGetAttributes", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(attributes), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&attributes, "nvmlDeviceGetAttributes", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlComputeInstanceGetInfo(
+    writer: &mut (impl std::io::Write + ?Sized),
+    computeInstance: cuda_types::nvml::nvmlComputeInstance_t,
+    info: *mut cuda_types::nvml::nvmlComputeInstanceInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(computeInstance), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &computeInstance,
+        "nvmlComputeInstanceGetInfo",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(info), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&info, "nvmlComputeInstanceGetInfo", arg_idx, writer)?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetComputeRunningProcesses(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetComputeRunningProcesses_v2(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGraphicsRunningProcesses(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGraphicsRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetGraphicsRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetGraphicsRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGraphicsRunningProcesses_v2(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGraphicsRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetGraphicsRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetGraphicsRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetMPSComputeRunningProcesses(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v1_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetMPSComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetMPSComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetMPSComputeRunningProcesses",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetMPSComputeRunningProcesses_v2(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    infoCount: *mut ::core::ffi::c_uint,
+    infos: *mut cuda_types::nvml::nvmlProcessInfo_v2_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetMPSComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infoCount), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infoCount,
+        "nvmlDeviceGetMPSComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(infos), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &infos,
+        "nvmlDeviceGetMPSComputeRunningProcesses_v2",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetGpuInstancePossiblePlacements(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    profileId: ::core::ffi::c_uint,
+    placements: *mut cuda_types::nvml::nvmlGpuInstancePlacement_t,
+    count: *mut ::core::ffi::c_uint,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &device,
+        "nvmlDeviceGetGpuInstancePossiblePlacements",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(profileId), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &profileId,
+        "nvmlDeviceGetGpuInstancePossiblePlacements",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(placements), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &placements,
+        "nvmlDeviceGetGpuInstancePossiblePlacements",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(count), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &count,
+        "nvmlDeviceGetGpuInstancePossiblePlacements",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlVgpuInstanceGetLicenseInfo(
+    writer: &mut (impl std::io::Write + ?Sized),
+    vgpuInstance: cuda_types::nvml::nvmlVgpuInstance_t,
+    licenseInfo: *mut cuda_types::nvml::nvmlVgpuLicenseInfo_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(vgpuInstance), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &vgpuInstance,
+        "nvmlVgpuInstanceGetLicenseInfo",
+        arg_idx,
+        writer,
+    )?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(licenseInfo), ": ").as_bytes())?;
+    crate::CudaDisplay::write(
+        &licenseInfo,
+        "nvmlVgpuInstanceGetLicenseInfo",
+        arg_idx,
+        writer,
+    )?;
+    writer.write_all(b")")
+}
+pub fn write_nvmlDeviceGetDriverModel(
+    writer: &mut (impl std::io::Write + ?Sized),
+    device: cuda_types::nvml::nvmlDevice_t,
+    current: *mut cuda_types::nvml::nvmlDriverModel_t,
+    pending: *mut cuda_types::nvml::nvmlDriverModel_t,
+) -> std::io::Result<()> {
+    let mut arg_idx = 0usize;
+    writer.write_all(b"(")?;
+    writer.write_all(concat!(stringify!(device), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&device, "nvmlDeviceGetDriverModel", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(current), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&current, "nvmlDeviceGetDriverModel", arg_idx, writer)?;
+    arg_idx += 1;
+    writer.write_all(b", ")?;
+    writer.write_all(concat!(stringify!(pending), ": ").as_bytes())?;
+    crate::CudaDisplay::write(&pending, "nvmlDeviceGetDriverModel", arg_idx, writer)?;
+    writer.write_all(b")")
+}
 impl crate::CudaDisplay for cuda_types::nvml::nvmlReturn_t {
     fn write(
         &self,

zluda_bindgen/src/main.rs

@@ -29,6 +29,7 @@ fn main() {
         &["..", "ext", "hip_runtime-sys", "src", "lib.rs"],
     );
     generate_rocblas(&crate_root, &["..", "ext", "rocblas-sys", "src", "lib.rs"]);
+    generate_rocm_smi(&crate_root, &["..", "ext", "rocm_smi-sys", "src", "lib.rs"]);
     let cuda_functions = generate_cuda(&crate_root);
     generate_process_address_table(&crate_root, cuda_functions);
     generate_ml(&crate_root);
@@ -179,6 +180,7 @@ fn generate_cufft(crate_root: &PathBuf) {
         &crate_root,
         &["..", "cuda_types", "src", "cufft.rs"],
         &module,
+        None,
     );
     generate_display_perflib(
         Some(&result_options),
@@ -245,6 +247,7 @@ fn generate_cusparse(crate_root: &PathBuf) {
         &crate_root,
         &["..", "cuda_types", "src", "cusparse.rs"],
         &module,
+        None,
     );
     generate_display_perflib(
         Some(&result_options),
@@ -685,6 +688,7 @@ fn generate_cublas(crate_root: &PathBuf) {
         &crate_root,
         &["..", "cuda_types", "src", "cublas.rs"],
         &module,
+        None,
     );
     generate_display_perflib(
         Some(&result_options),
@@ -759,6 +763,7 @@ fn generate_cublaslt(crate_root: &PathBuf) {
         &crate_root,
         &["..", "cuda_types", "src", "cublaslt.rs"],
         &module_blas,
+        None,
     );
     generate_display_perflib(
         None,
@@ -837,7 +842,10 @@ fn generate_ml(crate_root: &PathBuf) {
         .allowlist_var("^NVML.*")
         .must_use_type("nvmlReturn_t")
         .constified_enum("nvmlReturn_enum")
-        .clang_args(["-I/usr/local/cuda/include"])
+        .clang_args([
+            "-I/usr/local/cuda/include",
+            "-DNVML_NO_UNVERSIONED_FUNC_DEFS",
+        ])
         .generate()
         .unwrap()
         .to_string();
@@ -856,12 +864,31 @@ fn generate_ml(crate_root: &PathBuf) {
         success: ("NVML_SUCCESS", "SUCCESS"),
         hip_type: None,
     };
+    let suffix =
+"#[cfg(unix)]
+impl From<rocm_smi_sys::rsmi_error> for nvmlError_t {
+    fn from(error: rocm_smi_sys::rsmi_error) -> Self {
+        match error {
+            rocm_smi_sys::rsmi_error::INVALID_ARGS => nvmlError_t::from(nvmlError_t::INVALID_ARGUMENT),
+            rocm_smi_sys::rsmi_error::NOT_SUPPORTED => nvmlError_t::from(nvmlError_t::NOT_SUPPORTED),
+            rocm_smi_sys::rsmi_error::PERMISSION => nvmlError_t::from(nvmlError_t::NO_PERMISSION),
+            rocm_smi_sys::rsmi_error::INPUT_OUT_OF_BOUNDS => nvmlError_t::from(nvmlError_t::INVALID_ARGUMENT),
+            rocm_smi_sys::rsmi_error::INIT_ERROR => nvmlError_t::from(nvmlError_t::UNINITIALIZED),
+            rocm_smi_sys::rsmi_error::NOT_FOUND => nvmlError_t::from(nvmlError_t::GPU_NOT_FOUND),
+            rocm_smi_sys::rsmi_error::INSUFFICIENT_SIZE => nvmlError_t::from(nvmlError_t::INSUFFICIENT_SIZE),
+            rocm_smi_sys::rsmi_error::INTERRUPT => nvmlError_t::from(nvmlError_t::IRQ_ISSUE),
+            rocm_smi_sys::rsmi_error::NO_DATA => nvmlError_t::from(nvmlError_t::NO_DATA),
+            _ => nvmlError_t::from(nvmlError_t::UNKNOWN),
+        }
+    }
+}";
     generate_types_library(
         Some(&result_options),
         None,
         &crate_root,
         &["..", "cuda_types", "src", "nvml.rs"],
         &module,
+        Some(suffix),
     );
     generate_display_perflib(
         Some(&result_options),
@@ -878,6 +905,7 @@ fn generate_types_library(
     crate_root: &PathBuf,
     path: &[&str],
     module: &syn::File,
+    suffix: Option<&str>,
 ) {
     let module = generate_types_library_impl(result_options, module);
     let mut output = crate_root.clone();
@@ -895,6 +923,9 @@ fn generate_types_library(
                 .replace(" cuDoubleComplex", " super::cuda::cuDoubleComplex");
         }
     }
+    if let Some(suffix) = suffix {
+        text.push_str(suffix);
+    }
     write_rust_to_file(output, &text)
 }
 
@@ -1038,6 +1069,35 @@ fn generate_rocblas(output: &PathBuf, path: &[&str]) {
     write_rust_to_file(output, text)
 }
 
+fn generate_rocm_smi(output: &PathBuf, path: &[&str]) {
+    let rocm_smi_header = new_builder()
+        .header("/opt/rocm/include/rocm_smi/rocm_smi.h")
+        .allowlist_type("^rsmi.*")
+        .allowlist_function("^rsmi.*")
+        .allowlist_var("^RSMI_.*")
+        .must_use_type("rsmi_status_t")
+        .constified_enum("rsmi_status_t")
+        .clang_args(["-I/opt/rocm/include"])
+        .generate()
+        .unwrap()
+        .to_string();
+    let mut module: syn::File = syn::parse_str(&rocm_smi_header).unwrap();
+    let result_options = ConvertIntoRustResultOptions {
+        type_: "rsmi_status_t",
+        underlying_type: "rsmi_status_t",
+        new_error_type: "rsmi_error",
+        error_prefix: ("RSMI_STATUS_", "ERROR_"),
+        success: ("RSMI_STATUS_SUCCESS", "SUCCESS"),
+        hip_type: None,
+    };
+    let mut converter = ConvertIntoRustResult::new(result_options);
+    module.items = converter.convert(module.items).collect();
+    converter.flush(&mut module.items);
+    let mut output = output.clone();
+    output.extend(path);
+    write_rust_to_file(output, &prettyplease::unparse(&module))
+}
+
 fn add_send_sync(items: &mut Vec<Item>, arg: &[&str]) {
     for type_ in arg {
         let type_ = Ident::new(type_, Span::call_site());

zluda_blas/Cargo.toml

@@ -12,6 +12,7 @@ cuda_macros = { path = "../cuda_macros" }
 cuda_types = { path = "../cuda_types" }
 zluda_common = { path = "../zluda_common" }
 rocblas-sys = { path = "../ext/rocblas-sys" }
+hip_runtime-sys = { path = "../ext/hip_runtime-sys" }
 
 [package.metadata.zluda]
 linux_symlinks = [

zluda_blas/src/impl.rs

@@ -1,9 +1,8 @@
-use std::mem;
-
 use cuda_types::cublas::*;
-use zluda_common::{from_cuda_object, ZludaObject};
+use hip_runtime_sys::*;
-
 use rocblas_sys::*;
+use std::mem;
+use zluda_common::{from_cuda_object, ZludaObject};
 
 pub struct Handle {
     handle: rocblas_handle,
@@ -153,3 +152,71 @@ pub(crate) fn sgemm_v2(
 pub(crate) fn destroy_v2(handle: cublasHandle_t) -> cublasStatus_t {
     zluda_common::drop_checked::<Handle>(handle)
 }
+
+pub(crate) unsafe fn set_stream_v2(handle: &Handle, stream: hipStream_t) -> rocblas_status {
+    rocblas_set_stream(handle.handle, stream)
+}
+
+pub(crate) unsafe fn set_workspace_v2(
+    handle: &Handle,
+    workspace: *mut ::core::ffi::c_void,
+    size: usize,
+) -> rocblas_status {
+    rocblas_set_workspace(handle.handle, workspace, size)
+}
+
+pub(crate) unsafe fn get_math_mode(handle: &Handle, mode: &mut cublasMath_t) -> rocblas_status {
+    let mut roc_mode = mem::zeroed();
+    rocblas_get_math_mode(handle.handle, &mut roc_mode)?;
+    *mode = zluda_common::FromCuda::from_cuda(&roc_mode)?;
+    Ok(())
+}
+
+pub(crate) unsafe fn gemm_ex(
+    handle: &Handle,
+    transa: rocblas_operation,
+    transb: rocblas_operation,
+    m: ::core::ffi::c_int,
+    n: ::core::ffi::c_int,
+    k: ::core::ffi::c_int,
+    alpha: *const ::core::ffi::c_void,
+    a: *const ::core::ffi::c_void,
+    a_type: rocblas_datatype,
+    lda: ::core::ffi::c_int,
+    b: *const ::core::ffi::c_void,
+    b_type: rocblas_datatype,
+    ldb: ::core::ffi::c_int,
+    beta: *const ::core::ffi::c_void,
+    c: *mut ::core::ffi::c_void,
+    c_type: rocblas_datatype,
+    ldc: ::core::ffi::c_int,
+    compute_type: rocblas_datatype,
+    algo: rocblas_gemm_algo,
+) -> rocblas_status {
+    rocblas_gemm_ex(
+        handle.handle,
+        transa,
+        transb,
+        m,
+        n,
+        k,
+        alpha,
+        a,
+        a_type,
+        lda,
+        b,
+        b_type,
+        ldb,
+        beta,
+        c,
+        c_type,
+        ldc,
+        c,
+        c_type,
+        ldc,
+        compute_type,
+        algo,
+        0,
+        0,
+    )
+}

zluda_blas/src/lib.rs

@@ -47,10 +47,16 @@ cuda_macros::cublas_function_declarations!(
     implemented
         <= [
             cublasCreate_v2,
+            cublasDestroy_v2,
+            cublasGemmEx,
+            cublasGetMathMode,
+            cublasLtCreate,
+            cublasLtDestroy,
             cublasSetMathMode,
+            cublasSetStream_v2,
+            cublasSetWorkspace_v2,
             cublasSgemmStridedBatched,
             cublasSgemm_v2,
-            cublasDestroy_v2
         ],
     implemented_and_always_succeeds
         <= [

zluda_common/src/lib.rs

@@ -1,4 +1,4 @@
-use cuda_types::{cublas::*, cuda::*};
+use cuda_types::{cublas::*, cuda::*, nvml::*};
 use hip_runtime_sys::*;
 use rocblas_sys::*;
 use std::{
@@ -22,6 +22,16 @@ impl CudaErrorType for cublasError_t {
     const NOT_SUPPORTED: Self = Self::NOT_SUPPORTED;
 }
 
+impl CudaErrorType for rocblas_error {
+    const INVALID_VALUE: Self = Self::invalid_value;
+    const NOT_SUPPORTED: Self = Self::not_implemented;
+}
+
+impl CudaErrorType for nvmlError_t {
+    const INVALID_VALUE: Self = Self::INVALID_ARGUMENT;
+    const NOT_SUPPORTED: Self = Self::NOT_SUPPORTED;
+}
+
 /// Used to try to convert CUDA API values into our internal representation.
 ///
 /// Similar to [`TryFrom`], but we can implement this for primitive types. We also provide conversions from pointers to references.
@@ -142,7 +152,11 @@ from_cuda_nop!(
     CUcontext,
     cublasHandle_t,
     cublasStatus_t,
-    CUlaunchConfig
+    CUlaunchConfig,
+    cublasMath_t,
+    nvmlDevice_t,
+    nvmlFieldValue_t,
+    nvmlGpuFabricInfo_t
 );
 from_cuda_transmute!(
     CUuuid => hipUUID,
@@ -215,6 +229,76 @@ impl<'a, E: CudaErrorType> FromCuda<'a, cublasMath_t, E> for rocblas_math_mode {
     }
 }
 
+impl<'a, E: CudaErrorType> FromCuda<'a, rocblas_math_mode, E> for cublasMath_t {
+    fn from_cuda(mode: &'a rocblas_math_mode) -> Result<Self, E> {
+        Ok(match *mode {
+            rocblas_math_mode_::rocblas_default_math => cublasMath_t::CUBLAS_DEFAULT_MATH,
+            rocblas_math_mode::rocblas_xf32_xdl_math_op => cublasMath_t::CUBLAS_TF32_TENSOR_OP_MATH,
+            _ => return Err(E::NOT_SUPPORTED),
+        })
+    }
+}
+
+impl<'a, E: CudaErrorType> FromCuda<'a, cuda_types::cublas::cudaDataType, E> for rocblas_datatype {
+    fn from_cuda(mode: &'a cuda_types::cublas::cudaDataType) -> Result<Self, E> {
+        Ok(match *mode {
+            cudaDataType_t::CUDA_R_16F => rocblas_datatype::rocblas_datatype_f16_r,
+            cudaDataType_t::CUDA_R_32F => rocblas_datatype::rocblas_datatype_f32_r,
+            cudaDataType_t::CUDA_R_64F => rocblas_datatype::rocblas_datatype_f64_r,
+            cudaDataType_t::CUDA_C_16F => rocblas_datatype::rocblas_datatype_f16_c,
+            cudaDataType_t::CUDA_C_32F => rocblas_datatype::rocblas_datatype_f32_c,
+            cudaDataType_t::CUDA_C_64F => rocblas_datatype::rocblas_datatype_f64_c,
+            cudaDataType_t::CUDA_R_8I => rocblas_datatype::rocblas_datatype_i8_r,
+            cudaDataType_t::CUDA_R_8U => rocblas_datatype::rocblas_datatype_u8_r,
+            cudaDataType_t::CUDA_R_32I => rocblas_datatype::rocblas_datatype_i32_r,
+            cudaDataType_t::CUDA_R_32U => rocblas_datatype::rocblas_datatype_u32_r,
+            cudaDataType_t::CUDA_C_8I => rocblas_datatype::rocblas_datatype_i8_c,
+            cudaDataType_t::CUDA_C_8U => rocblas_datatype::rocblas_datatype_u8_c,
+            cudaDataType_t::CUDA_C_32I => rocblas_datatype::rocblas_datatype_i32_c,
+            cudaDataType_t::CUDA_C_32U => rocblas_datatype::rocblas_datatype_u32_c,
+            cudaDataType_t::CUDA_R_16BF => rocblas_datatype::rocblas_datatype_bf16_r,
+            cudaDataType_t::CUDA_C_16BF => rocblas_datatype::rocblas_datatype_bf16_c,
+            cudaDataType_t::CUDA_R_8F_UE4M3 => rocblas_datatype::rocblas_datatype_f8_r,
+            cudaDataType_t::CUDA_R_8F_E5M2 => rocblas_datatype::rocblas_datatype_bf8_r,
+            _ => return Err(E::NOT_SUPPORTED),
+        })
+    }
+}
+
+impl<'a, E: CudaErrorType> FromCuda<'a, cuda_types::cublas::cublasComputeType_t, E>
+    for rocblas_computetype
+{
+    fn from_cuda(mode: &'a cuda_types::cublas::cublasComputeType_t) -> Result<Self, E> {
+        Ok(match *mode {
+            cublasComputeType_t::CUBLAS_COMPUTE_32F => {
+                rocblas_computetype::rocblas_compute_type_f32
+            }
+            _ => return Err(E::NOT_SUPPORTED),
+        })
+    }
+}
+
+impl<'a, E: CudaErrorType> FromCuda<'a, cuda_types::cublas::cublasComputeType_t, E>
+    for rocblas_datatype
+{
+    fn from_cuda(mode: &'a cuda_types::cublas::cublasComputeType_t) -> Result<Self, E> {
+        Ok(match *mode {
+            cublasComputeType_t::CUBLAS_COMPUTE_16F => rocblas_datatype::rocblas_datatype_f16_r,
+            cublasComputeType_t::CUBLAS_COMPUTE_32F => rocblas_datatype::rocblas_datatype_f32_r,
+            cublasComputeType_t::CUBLAS_COMPUTE_64F => rocblas_datatype::rocblas_datatype_f64_r,
+            _ => return Err(E::NOT_SUPPORTED),
+        })
+    }
+}
+
+impl<'a, E: CudaErrorType> FromCuda<'a, cuda_types::cublas::cublasGemmAlgo_t, E>
+    for rocblas_gemm_algo
+{
+    fn from_cuda(_: &'a cuda_types::cublas::cublasGemmAlgo_t) -> Result<Self, E> {
+        Ok(rocblas_gemm_algo::rocblas_gemm_algo_standard)
+    }
+}
+
 /// Represents an object that can be sent across the API boundary.
 ///
 /// Some CUDA calls operate on an opaque handle. For example, `cuModuleLoadData` will load a

zluda_ml/Cargo.toml

@@ -11,6 +11,10 @@ crate-type = ["cdylib"]
 [dependencies]
 cuda_macros = { path = "../cuda_macros" }
 cuda_types = { path = "../cuda_types" }
+zluda_common = { path = "../zluda_common" }
+
+[target.'cfg(unix)'.dependencies]
+rocm_smi-sys = { path = "../ext/rocm_smi-sys" }
 
 [package.metadata.zluda]
 linux_symlinks = [

zluda_ml/src/impl_common.rs

@@ -1,6 +1,8 @@
 use cuda_types::nvml::*;
 use std::{ffi::CStr, ptr};
 
+const VERSION: &'static CStr = c"550.77";
+
 #[cfg(debug_assertions)]
 pub(crate) fn unimplemented() -> nvmlReturn_t {
     unimplemented!()
@@ -11,20 +13,10 @@ pub(crate) fn unimplemented() -> nvmlReturn_t {
     nvmlReturn_t::ERROR_NOT_SUPPORTED
 }
 
-pub(crate) fn error_string(_result: cuda_types::nvml::nvmlReturn_t) -> *const ::core::ffi::c_char {
-    c"".as_ptr()
-}
-
-pub(crate) fn init_v2() -> cuda_types::nvml::nvmlReturn_t {
-    nvmlReturn_t::SUCCESS
-}
-
-const VERSION: &'static CStr = c"550.77";
-
 pub(crate) fn system_get_driver_version(
     result: *mut ::core::ffi::c_char,
     length: ::core::ffi::c_uint,
-) -> cuda_types::nvml::nvmlReturn_t {
+) -> nvmlReturn_t {
     if result == ptr::null_mut() {
         return nvmlReturn_t::ERROR_INVALID_ARGUMENT;
     }
@@ -37,3 +29,7 @@ pub(crate) fn system_get_driver_version(
     }
     nvmlReturn_t::SUCCESS
 }
+
+pub(crate) fn error_string(_result: nvmlReturn_t) -> *const ::core::ffi::c_char {
+    c"".as_ptr()
+}

zluda_ml/src/impl_unix.rs

@@ -0,0 +1,77 @@
+use cuda_types::nvml::*;
+use rocm_smi_sys::*;
+use std::mem;
+use zluda_common::{from_cuda_object, ZludaObject};
+
+pub(crate) use crate::impl_common::error_string;
+pub(crate) use crate::impl_common::system_get_driver_version;
+
+pub(crate) struct Device {
+    _index: u32,
+}
+
+impl ZludaObject for Device {
+    const COOKIE: usize = 0x79443851e7cee0d9;
+
+    type Error = nvmlError_t;
+    type CudaHandle = nvmlDevice_t;
+
+    fn drop_checked(&mut self) -> nvmlReturn_t {
+        Ok(())
+    }
+}
+
+from_cuda_object!(Device);
+
+pub(crate) unsafe fn init() -> rsmi_status_t {
+    rsmi_init(0)
+}
+
+pub(crate) unsafe fn init_v2() -> rsmi_status_t {
+    rsmi_init(0)
+}
+
+pub(crate) unsafe fn init_with_flags(_flags: ::core::ffi::c_uint) -> rsmi_status_t {
+    rsmi_init(0)
+}
+
+pub(crate) unsafe fn shutdown() -> rsmi_status_t {
+    rsmi_shut_down()
+}
+
+pub(crate) unsafe fn device_get_count_v2(device_count: &mut ::core::ffi::c_uint) -> rsmi_status_t {
+    rsmi_num_monitor_devices(device_count)
+}
+
+pub(crate) unsafe fn device_get_field_values(
+    _device: &Device,
+    values_count: ::core::ffi::c_int,
+    values: &mut cuda_types::nvml::nvmlFieldValue_t,
+) -> nvmlReturn_t {
+    for field in std::slice::from_raw_parts_mut(values, values_count as usize) {
+        get_field_value(field)?;
+    }
+    Ok(())
+}
+
+unsafe fn get_field_value(field: &mut nvmlFieldValue_st) -> Result<(), nvmlError_t> {
+    *field = mem::zeroed();
+    field.nvmlReturn = nvmlReturn_t::ERROR_NOT_SUPPORTED;
+    Ok(())
+}
+
+pub(crate) unsafe fn device_get_gpu_fabric_info(
+    _device: &Device,
+    gpu_fabric_info: &mut cuda_types::nvml::nvmlGpuFabricInfo_t,
+) -> nvmlReturn_t {
+    *gpu_fabric_info = mem::zeroed();
+    Ok(())
+}
+
+pub(crate) fn device_get_handle_by_index_v2(
+    index: ::core::ffi::c_uint,
+    device: &mut cuda_types::nvml::nvmlDevice_t,
+) -> nvmlReturn_t {
+    *device = Device { _index: index }.wrap();
+    nvmlReturn_t::SUCCESS
+}

zluda_ml/src/impl_win.rs

@@ -0,0 +1,50 @@
+use cuda_types::nvml::*;
+
+pub(crate) use crate::impl_common::error_string;
+pub(crate) use crate::impl_common::system_get_driver_version;
+
+pub(crate) unsafe fn init() -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn init_v2() -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn init_with_flags(_flags: ::core::ffi::c_uint) -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn shutdown() -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn device_get_count_v2(_device_count: &mut ::core::ffi::c_uint) -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn device_get_field_values(
+    _device: cuda_types::nvml::nvmlDevice_t,
+    _values_count: ::core::ffi::c_int,
+    _values: &mut cuda_types::nvml::nvmlFieldValue_t,
+) -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+unsafe fn get_field_value(_field: &mut nvmlFieldValue_st) -> Result<(), nvmlError_t> {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) unsafe fn device_get_gpu_fabric_info(
+    _device: cuda_types::nvml::nvmlDevice_t,
+    _gpu_fabric_info: &mut cuda_types::nvml::nvmlGpuFabricInfo_t,
+) -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}
+
+pub(crate) fn device_get_handle_by_index_v2(
+    _index: ::core::ffi::c_uint,
+    _device: &mut cuda_types::nvml::nvmlDevice_t,
+) -> nvmlReturn_t {
+    crate::impl_common::unimplemented()
+}

zluda_ml/src/lib.rs

@@ -1,4 +1,7 @@
+#[cfg_attr(windows, path = "impl_win.rs")]
+#[cfg_attr(unix, path = "impl_unix.rs")]
 mod r#impl;
+mod impl_common;
 
 macro_rules! unimplemented_fn {
     ($($abi:literal fn $fn_name:ident( $($arg_id:ident : $arg_type:ty),* ) -> $ret_type:ty;)*) => {
@@ -6,19 +9,32 @@ macro_rules! unimplemented_fn {
             #[no_mangle]
             #[allow(improper_ctypes_definitions)]
             pub extern $abi fn $fn_name ( $( $arg_id : $arg_type),* ) -> $ret_type {
-                r#impl::unimplemented()
+                impl_common::unimplemented()
             }
         )*
     };
 }
 
 macro_rules! implemented_fn {
+    ($($abi:literal fn $fn_name:ident( $($arg_id:ident : $arg_type:ty),* ) -> $ret_type:ty;)*) => {
+        $(
+            #[no_mangle]
+            #[allow(improper_ctypes_definitions)]
+            pub unsafe extern $abi fn $fn_name ( $( $arg_id : $arg_type),* ) -> $ret_type {
+                cuda_macros::nvml_normalize_fn!( crate::r#impl::$fn_name ) ( $( zluda_common::FromCuda::<_, cuda_types::nvml::nvmlError_t>::from_cuda(&$arg_id )?),*)?;
+                Ok(())
+            }
+        )*
+    };
+}
+
+macro_rules! implemented_unnormalized {
     ($($abi:literal fn $fn_name:ident( $($arg_id:ident : $arg_type:ty),* ) -> $ret_type:ty;)*) => {
         $(
             #[no_mangle]
             #[allow(improper_ctypes_definitions)]
             pub extern $abi fn $fn_name ( $( $arg_id : $arg_type),* ) -> $ret_type {
-                cuda_macros::nvml_normalize_fn!( crate::r#impl::$fn_name ) ( $( $arg_id ),* )
+                cuda_macros::nvml_normalize_fn!( crate::r#impl::$fn_name ) ( $( $arg_id),*)
             }
         )*
     };
@@ -26,5 +42,17 @@ macro_rules! implemented_fn {
 
 cuda_macros::nvml_function_declarations!(
     unimplemented_fn,
-    implemented_fn <= [nvmlErrorString, nvmlInit_v2, nvmlSystemGetDriverVersion]
+    implemented_fn
+        <= [
+            nvmlDeviceGetCount_v2,
+            nvmlDeviceGetFieldValues,
+            nvmlDeviceGetGpuFabricInfo,
+            nvmlDeviceGetHandleByIndex_v2,
+            nvmlInit,
+            nvmlInitWithFlags,
+            nvmlInit_v2,
+            nvmlShutdown,
+            nvmlSystemGetDriverVersion,
+        ],
+    implemented_unnormalized <= [nvmlErrorString,]
 );