Re: [RFC] amdkfd: Add GPU family name property to KFD topology

[Mario Limonciello]

On 2/11/2026 4:25 PM, Alex Deucher wrote:
On Wed, Feb 11, 2026 at 5:14 PM Mario Limonciello
<mario.limoncie...@amd.com> wrote:

Add a new 'gpu_family' property to the KFD topology sysfs interface that
exposes the GPU family name in ROCm TheRock format (e.g., gfx110X, gfx90X).

The property is algorithmically derived from gfx_target_version, which
encodes the version as MMMNNRR (major.minor.revision). The family name
follows the pattern "gfx{major}{minor}X", making it easy for users and
tools to identify which ROCm compiler target to use.

The property only appears for GFX9 and newer GPUs, as older generations
predate TheRock targets. Unknown or unsupported GPUs will not display
the property.

Example usage:
   $ cat /sys/class/kfd/kfd/topology/nodes/1/properties | grep gpu_family
   gpu_family gfx115X

I would call this isa_version or isa_family.  gfx_target_version is
the ISA version of a chip (i.e., the compiler target).  It's not the
same as the GC IP version or the chip family.  That should help
differentiate it and avoid confusion.

Alex


Thanks, I guess of those options I would say based on how things are 
named in TheRock isa_family feels more logical.

But now I'm not sure if directionally this property will really be 
useful or not.  I noticed that the apt packages and wheels aren't 
actually split
up this way anymore.


Signed-off-by: Mario Limonciello <mario.limoncie...@amd.com>
---
The idea is that this can help a user figure out which wheels they need
for their hardware.

Cc: stella.laure...@amd.com
Cc: slobodan.jo...@amd.com
Cc: filip.janko...@amd.com
  drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 37 +++++++++++++++++++++++
  drivers/gpu/drm/amd/amdkfd/kfd_topology.h |  1 +
  2 files changed, 38 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c 
b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index 08c63b4cc9a5f..ab029d87eb9ea 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -471,6 +471,9 @@ static ssize_t node_show(struct kobject *kobj, struct 
attribute *attr,
                               dev->node_props.max_slots_scratch_cu);
         sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
                               dev->node_props.gfx_target_version);
+       if (dev->node_props.gpu_family[0])
+               sysfs_show_gen_prop(buffer, offs, "gpu_family %s\n",
+                                   dev->node_props.gpu_family);
         sysfs_show_32bit_prop(buffer, offs, "vendor_id",
                               dev->node_props.vendor_id);
         sysfs_show_32bit_prop(buffer, offs, "device_id",
@@ -2048,6 +2051,37 @@ static void kfd_topology_set_capabilities(struct 
kfd_topology_device *dev)
         kfd_topology_set_dbg_firmware_support(dev);
  }

+/**
+ * kfd_get_gpu_family_name - Get GPU family name from gfx_target_version
+ * @gfx_target_version: Numeric GPU target version
+ * @family_name: Output buffer for family name
+ * @size: Size of output buffer
+ *
+ * Converts gfx_target_version to TheRock family format (e.g., "gfx110X").
+ * The gfx_target_version encodes version as MMMNNRR where:
+ *   MMM = major version
+ *   NN  = minor version
+ *   RR  = revision
+ * Family format is "gfx{major}{minor}X", e.g., 110501 (11.5.01) -> "gfx115X"
+ * If the version is not recognized, family_name is set to empty string.
+ */
+static void kfd_get_gpu_family_name(uint32_t gfx_target_version,
+                                   char *family_name, size_t size)
+{
+       int major, minor;
+
+       /* Only support TheRock GPU families (GFX9 and newer) */
+       if (gfx_target_version < 90000) {
+               family_name[0] = '\0';
+               return;
+       }
+
+       major = gfx_target_version / 10000;
+       minor = (gfx_target_version / 100) % 100;
+
+       snprintf(family_name, size, "gfx%d%dX", major, minor);
+}
+
  int kfd_topology_add_device(struct kfd_node *gpu)
  {
         uint32_t gpu_id;
@@ -2105,6 +2139,9 @@ int kfd_topology_add_device(struct kfd_node *gpu)

         dev->node_props.gfx_target_version =
                                 gpu->kfd->device_info.gfx_target_version;
+       kfd_get_gpu_family_name(dev->node_props.gfx_target_version,
+                               dev->node_props.gpu_family,
+                               sizeof(dev->node_props.gpu_family));
         dev->node_props.vendor_id = gpu->adev->pdev->vendor;
         dev->node_props.device_id = gpu->adev->pdev->device;
         dev->node_props.capability |=
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h 
b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
index 3de8ec0043bb4..8e52dd59b53dd 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.h
@@ -81,6 +81,7 @@ struct kfd_node_properties {
         uint32_t eop_buffer_size;
         uint32_t debug_memory_size;
         char name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
+       char gpu_family[8];  /* GPU family name in TheRock format (e.g., 
gfx110X) */
  };

  struct kfd_mem_properties {
--
2.53.0