[PATCH v2 4/4] drm/amdgpu: Use xgmi APIs for init and bandwidth

[Kim, Jonathan]

[Public]

> -----Original Message-----
> From: Lazar, Lijo <Lijo.Lazar at amd.com>
> Sent: Monday, February 10, 2025 2:01 AM
> To: amd-gfx at lists.freedesktop.org
> Cc: Zhang, Hawking <Hawking.Zhang at amd.com>; Deucher, Alexander
> <Alexander.Deucher at amd.com>; Kim, Jonathan <Jonathan.Kim at amd.com>
> Subject: [PATCH v2 4/4] drm/amdgpu: Use xgmi APIs for init and bandwidth
>
> Initialize xgmi related static information during early_init. Use xgmi
> API to get max bandwidth details.
>
> Signed-off-by: Lijo Lazar <lijo.lazar at amd.com>
> ---
> v2:
>       Move XGMI info init to early init phase (Jon)
>
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 6 ++++--
>  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 3 +++
>  2 files changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> index 2c1b38c5cfc6..e6bc70fb1eb6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> @@ -576,6 +576,7 @@ int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct
> amdgpu_device *dst,
>                                           bool is_min)
>  {
>       struct amdgpu_device *adev = dst, *peer_adev;
> +     uint32_t link_max_bandwidth;
>       int num_links;
>
>       if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2))
> @@ -593,8 +594,9 @@ int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct
> amdgpu_device *dst,
>               num_links = 0;
>       }
>
> -     /* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for
> bandwidth. */
> -     return (num_links * 16 * 25000)/BITS_PER_BYTE;
> +     link_max_bandwidth = amdgpu_xgmi_get_max_bandwidth(adev);
> +
> +     return (num_links * link_max_bandwidth)/BITS_PER_BYTE;
>  }

I have a WIP patch that deprecates the amdgpu_amdkfd_get_xgmi_* calls and consolidates them into amdgpu_xgmi.c.
This should hopefully reduce the middleman and make life easier for both parts of the driver.
It'll take a day or two to fully test, but I'll send you a compile tested version offline in the meantime to give you a sense of what I'll be doing.

Jon

>
>  int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool
> is_min)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 315ab2d6717a..09aac4542ed7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -2749,6 +2749,9 @@ static int amdgpu_device_ip_early_init(struct
> amdgpu_device *adev)
>       if (!total)
>               return -ENODEV;
>
> +     if (adev->gmc.xgmi.supported)
> +             amdgpu_xgmi_early_init(adev);
> +
>       ip_block = amdgpu_device_ip_get_ip_block(adev,
> AMD_IP_BLOCK_TYPE_GFX);
>       if (ip_block->status.valid != false)
>               amdgpu_amdkfd_device_probe(adev);
> --
> 2.25.1