Is DG1 supported with Linux xe KMD?

[ky-bd]

I see that there are xe KMD support and DG1 support, but I can't get them working together. clinfo -l with the following debug settings shows:

export NEOReadDebugKeys=1
export PrintDriverDiagnostics=5
export PrintDebugSettings=1
export PrintDebugMessages=1
export PrintXeLogs=1
export PrintBOCreateDestroyResult=1

# clinfo -l
Non-default value of debug variable: PrintDriverDiagnostics = 5
Non-default value of debug variable: PrintDebugSettings = 1
Non-default value of debug variable: PrintDebugMessages = 1
Non-default value of debug variable: PrintXeLogs = 1
Non-default value of debug variable: PrintBOCreateDestroyResult = 1
Shared System USM NOT allowed: KMD does not support
EXT_SET_PAT support is: disabled
INFO: System Info query failed!
WARNING: Failed to query memory info
WARNING: Failed to query engine info
WARNING: Topology query failed!
FATAL: Cannot query EU total parameter!
Platform #0: Intel(R) OpenCL
 `-- Device #0: AMD Ryzen 9 5900X 12-Core Processor

strace clinfo -l output (the ioctl operations):

...
openat(AT_FDCWD, "/dev/dri/by-path", O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_DIRECTORY) = 5
fstat(5, {st_mode=S_IFDIR|0755, st_size=80, ...}) = 0
getdents64(5, 0x571b5acd2090 /* 4 entries */, 32768) = 144
getdents64(5, 0x571b5acd2090 /* 0 entries */, 32768) = 0
close(5)                                = 0
openat(AT_FDCWD, "/dev/dri/by-path/pci-0000:2f:00.0-render", O_RDWR|O_CLOEXEC) = 5
ioctl(5, DRM_IOCTL_VERSION, 0x7fff929d60a0) = 0
ioctl(5, DRM_IOCTL_VERSION, 0x7fff929d6080) = 0
ioctl(5, DRM_IOCTL_XE_DEVICE_QUERY, 0x7fff929d6140) = 0
ioctl(5, DRM_IOCTL_XE_DEVICE_QUERY, 0x7fff929d6140) = 0
write(2, "Shared System USM NOT allowed: K"..., 52Shared System USM NOT allowed: KMD does not support
) = 52
ioctl(5, DRM_IOCTL_VERSION, 0x7fff929d5cf0) = 0
ioctl(5, DRM_IOCTL_I915_GEM_CREATE_EXT, 0x7fff929d5bf0) = -1 EINVAL
write(1, "EXT_SET_PAT support is: disabled"..., 33EXT_SET_PAT support is: disabled
) = 33
ioctl(5, DRM_IOCTL_I915_REG_READ, 0x7fff929d5bf0) = -1 EINVAL
ioctl(5, DRM_IOCTL_I915_REG_READ, 0x7fff929d5bf0) = -1 EINVAL
readlink("/proc/self/exe", "/usr/bin/clinfo", 511) = 15
ioctl(5, DRM_IOCTL_I915_QUERY, 0x7fff929d5ca0) = -1 EINVAL
write(1, "INFO: System Info query failed!\n", 32INFO: System Info query failed!
) = 32
ioctl(5, DRM_IOCTL_I915_QUERY, 0x7fff929d5b80) = -1 EINVAL
write(2, "WARNING: Failed to query memory "..., 37WARNING: Failed to query memory info
) = 37
ioctl(5, DRM_IOCTL_I915_QUERY, 0x7fff929d5a20) = -1 EINVAL
write(2, "WARNING: Failed to query engine "..., 37WARNING: Failed to query engine info
) = 37
ioctl(5, DRM_IOCTL_I915_QUERY, 0x7fff929d5b30) = -1 EINVAL
write(2, "WARNING: Topology query failed!\n", 32WARNING: Topology query failed!
) = 32
ioctl(5, DRM_IOCTL_I915_GETPARAM, 0x7fff929d5bf0) = -1 EINVAL
write(2, "FATAL: Cannot query EU total par"..., 40FATAL: Cannot query EU total parameter!
) = 40
close(5)                                = 0
munmap(0x720be1000000, 35286328)        = 0
munmap(0x720be3643000, 864456)          = 0
close(4)                                = 0
...

By looking into the code, I found that the runtime queries the device with IoctlHelperXe if xe KMD is found:

compute-runtime/shared/source/os_interface/linux/drm_neo.cpp

Lines 2016 to 2023 in 2459623

	bool Drm::queryDeviceIdAndRevision() {
	auto drmVersion = Drm::getDrmVersion(getFileDescriptor());
	if ("xe" == drmVersion) {
	this->setPerContextVMRequired(false);
	return IoctlHelperXe::queryDeviceIdAndRevision(*this);
	}
	return IoctlHelperI915::queryDeviceIdAndRevision(*this);
	}

... but sets up a product-specific ioctl helper for DG1:

compute-runtime/shared/source/os_interface/linux/drm_neo.cpp

Lines 1258 to 1273 in 2459623

	void Drm::setupIoctlHelper(const PRODUCT_FAMILY productFamily) {
	if (!this->ioctlHelper) {
	auto drmVersion = Drm::getDrmVersion(getFileDescriptor());
	auto productSpecificIoctlHelperCreator = ioctlHelperFactory[productFamily];
	if (productSpecificIoctlHelperCreator && !debugManager.flags.IgnoreProductSpecificIoctlHelper.get()) {
	this->ioctlHelper = productSpecificIoctlHelperCreator.value()(*this);
	} else if ("xe" == drmVersion) {
	this->ioctlHelper = IoctlHelperXe::create(*this);
	} else {
	std::string prelimVersion = "";
	getPrelimVersion(prelimVersion);
	this->ioctlHelper = IoctlHelper::getI915Helper(productFamily, prelimVersion, *this);
	}
	this->ioctlHelper->initialize();
	}
	}

... which is an IoctlHelperImpl<IGFX_DG1>:

compute-runtime/shared/source/os_interface/linux/local/dg1/enable_ioctl_helper_dg1.cpp

Lines 12 to 18 in 2459623

	struct EnableProductIoctlHelperDg1 {
	EnableProductIoctlHelperDg1() {
	ioctlHelperFactory[IGFX_DG1] = IoctlHelperImpl<IGFX_DG1>::get;
	}
	};
	static EnableProductIoctlHelperDg1 enableIoctlHelperDg1;

... and it is inherited from the IoctlHelperUpstream then IoctlHelperI915, instead of the IoctlHelperXe.

compute-runtime/shared/source/os_interface/linux/ioctl_helper.h

Lines 384 to 385 in 2459623

	template <PRODUCT_FAMILY gfxProduct>
	class IoctlHelperImpl : public IoctlHelperUpstream {

compute-runtime/shared/source/os_interface/linux/ioctl_helper.h

Line 321 in 2459623

class IoctlHelperUpstream : public IoctlHelperI915 {

So the device is queried with i915 ioctl commands and causes the clinfo failed to probe the device.

The clinfo works fine with i915 KMD:

# clinfo -l
Non-default value of debug variable: PrintDriverDiagnostics = 5
Non-default value of debug variable: PrintDebugSettings = 1
Non-default value of debug variable: PrintDebugMessages = 1
Non-default value of debug variable: PrintXeLogs = 1
Non-default value of debug variable: PrintBOCreateDestroyResult = 1
EXT_SET_PAT support is: disabled
INFO: System Info query failed!
WARNING: Failed to request OCL Turbo Boost
Created new BO with GEM_USERPTR, handle: BO-1
NEO_CACHE_PERSISTENT is enabled. Cache is located in: /root/.cache/neo_compiler_cache

Performing GEM_CREATE_EXT with { size: 4096, memory class: 0, memory instance: 0 }
GEM_CREATE_EXT with EXT_MEMORY_REGIONS has returned: 0 BO-2 with size: 4096
Performing GEM_CREATE_EXT with { size: 4096, memory class: 0, memory instance: 0 }
GEM_CREATE_EXT with EXT_MEMORY_REGIONS has returned: 0 BO-3 with size: 4096
Performing GEM_CREATE_EXT with { size: 4096, memory class: 0, memory instance: 0 }
GEM_CREATE_EXT with EXT_MEMORY_REGIONS has returned: 0 BO-4 with size: 4096
Performing GEM_CREATE_EXT with { size: 4096, memory class: 0, memory instance: 0 }
GEM_CREATE_EXT with EXT_MEMORY_REGIONS has returned: 0 BO-5 with size: 4096
Performing GEM_CREATE_EXT with { size: 4096, memory class: 0, memory instance: 0 }
GEM_CREATE_EXT with EXT_MEMORY_REGIONS has returned: 0 BO-6 with size: 4096
computeUnitsUsedForScratch: 768
hwInfo: {96, 672}: (16, 1, 6)
Platform #0: Intel(R) OpenCL Graphics
 `-- Device #0: Intel(R) Iris(R) Xe MAX Graphics
Platform #1: Intel(R) OpenCL
 `-- Device #0: AMD Ryzen 9 5900X 12-Core Processor
Calling gem close on handle: BO-2
Calling gem close on handle: BO-3
Calling gem close on handle: BO-4
Calling gem close on handle: BO-5
Calling gem close on handle: BO-6
Calling gem close on handle: BO-1

Environment

• CPU: AMD Ryzen 9 5900X
• GPU: Intel Corporation DG1 [Iris Xe MAX Graphics] [8086:4905] (rev 01)
• MB: MSI MAG X570 TOMAHAWK WIFI (MS-7C84) (UEFI boot, resizable BAR and above 4G decoding enabled, CSM disabled)
• OS: Ubuntu 24.04.4 LTS
• Kernel: 6.17.0-14-generic
• Kernel cmdline: xe.force_probe=4905 acpi_enforce_resources=lax pci=nommconf pcie_acs_override=downstream pcie_aspm=off
• Driver: xe (i915 blacklisted in modprobe.d)
• OpenCL loader: Khronos OpenCL ICD Loader 3.0.7
• OpenCL runtime (ICD): intel-opencl-icd 26.05.37020.3-1~24.04~ppa2
• Level Zero: libze-intel-gpu1 26.05.37020.3-1~24.04~ppa2, libze1 1.27.0-1~24.04~ppa1
• GMM: libigdgmm12 22.9.0-1~24.04~ppa1
• Intel GSC: 0.9.5-1~24.04~ppa2
• linux-firmware: 20240318.git3b128b60-0ubuntu2.25

[HmnSn]

Simply swapping the order here enables support for DG1 on xe, but I’m not sure what side effects this might cause.

diff --git a/shared/source/os_interface/linux/drm_neo.cpp b/shared/source/os_interface/linux/drm_neo.cpp
index 2ec412c967..56d5f4c516 100644
--- a/shared/source/os_interface/linux/drm_neo.cpp
+++ b/shared/source/os_interface/linux/drm_neo.cpp
@@ -1259,10 +1259,10 @@ void Drm::setupIoctlHelper(const PRODUCT_FAMILY productFamily) {
     if (!this->ioctlHelper) {
         auto drmVersion = Drm::getDrmVersion(getFileDescriptor());
         auto productSpecificIoctlHelperCreator = ioctlHelperFactory[productFamily];
-        if (productSpecificIoctlHelperCreator && !debugManager.flags.IgnoreProductSpecificIoctlHelper.get()) {
-            this->ioctlHelper = productSpecificIoctlHelperCreator.value()(*this);
-        } else if ("xe" == drmVersion) {
+        if ("xe" == drmVersion) {
             this->ioctlHelper = IoctlHelperXe::create(*this);
+        } else if (productSpecificIoctlHelperCreator && !debugManager.flags.IgnoreProductSpecificIoctlHelper.get()) {
+            this->ioctlHelper = productSpecificIoctlHelperCreator.value()(*this);
         } else {
             std::string prelimVersion = "";
             getPrelimVersion(prelimVersion);

[kgibala]

Hi @ky-bd,

Thank you for the detailed report and for taking the time to investigate the code paths involved.

To clarify the current support status: DG1 is not officially supported by the xe kernel driver. The official and fully supported kernel driver for DG1 is i915. When you use xe.force_probe=4905 in your kernel command line, you are explicitly bypassing the xe driver's built-in guard that would otherwise warn:

[ 3.299922] xe 0000:03:00.0: Your graphics device 4905 is not officially supported
by xe driver in this kernel version. To force Xe probe, use xe.force_probe='4905'
and i915.force_probe='!4905' module parameters or CONFIG_DRM_XE_FORCE_PROBE='4905'
and CONFIG_DRM_I915_FORCE_PROBE='!4905' configuration options.

By forcing the probe, you are opting into unsupported territory.

As an experiment, you can enable the xeKMD module on DG1 using the experimental debug flag IgnoreProductSpecificIoctlHelper:

NEOReadDebugKeys=1 IgnoreProductSpecificIoctlHelper=1

NEOReadDebugKeys=1 IgnoreProductSpecificIoctlHelper=1 PrintXeLogs=1 clinfo -l
IoctlHelperXe::IoctlHelperXe
IoctlHelperXe::initialize
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID 0x14905
REV_ID 0x1
DEVICE_ID 0x4905
DRM_XE_QUERY_CONFIG_FLAGS 0x3
DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM ON
DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT 0x1000
DRM_XE_QUERY_CONFIG_VA_BITS 0x30
DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY 0x1
DRM_XE_QUERY_CONFIG_FLAG_HAS_LOW_LATENCY ON
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getDrmParamValue 0x2a
QueryHwconfigTable => IoctlHelperXe::ioctl 0x10
-> IoctlHelperXe::ioctl Query id=0x2a f=0x0 len=0
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getDrmParamValue 0x1a MemoryClassSystem
-> IoctlHelperXe::getDrmParamValue 0x19 MemoryClassDevice
-> IoctlHelperXe::getIoctlRequestValue 0x10
-> IoctlHelperXe::getIoctlRequestValue 0x10
numberHwEngines=5
-> IoctlHelperXe::getDrmParamValue 0x11 EngineClassVideo
-> IoctlHelperXe::getDrmParamValue 0x12 EngineClassVideoEnhance
class: rcs, instance: 0, gt_id: 0, tile: 0
...
Platform #0: Intel(R) OpenCL Graphics
`-- Device #0: Intel(R) Iris(R) Xe MAX Graphics

We will dispatch this topic further internally to explore the options for enabling this as a default flow. We will post updates once they become available.

Thank you for your cooperation.