Request: Smart Access Memory information

[bridgman]

Why is BAR showing up larger than my VRAM?

The PCIE standard requires that BAR sizes be a power of two, so with 12GB the BAR size needs to either be rounded up to 16GB or rounded down to 8GB... and rounding down would mean you were not able to access all of your memory.

[Melcar]

5800X
B550 Pro V2
RX 6750 XT

AMD_DEBUG=info glxinfo | grep -E 'vram|name|smart'
name of display: :0
    name = NAVI22
    marketing_name = AMD Radeon RX 6750 XT
    vram_size = 12288 MB
    vram_vis_size = 12272 MB
    vram_type = 9
    has_dedicated_vram = 1
    all_vram_visible = 1
    smart_access_memory = 0
​

dmesg | grep BAR
[    0.522837] pci 0000:08:00.0: BAR 0: assigned to efifb
[    5.319931] [drm] Detected VRAM RAM=12272M, BAR=16384M​

Why is BAR showing up larger than my VRAM?

[leuc]

CPU: AMD Ryzen 9 5900X

MB Name: B350 TOMAHAWK ARCTIC (MS-7A34)
BIOS Version: Version: H.M5
Release Date: 07/18/2022

BIOS version has Re-BAR/4G named option under PCI, disabled by default it seems.

GPU: Radeon RX 480 Gaming X 8GB

AMD_DEBUG=info glxinfo | grep -E 'vram|smart'
vram_size = 8192 MB
vram_vis_size = 8192 MB
vram_type = 5
vram_bit_width = 256
has_dedicated_vram = 1
all_vram_visible = 1
smart_access_memory = 0​

dmesg | grep BAR=
[ 3.321376] [drm] Detected VRAM RAM=8192M, BAR=8192M​

[leuc]

Dell G5 5505 Laptop with Navi 10 RX 5600 and Renoir APU

No BIOS changes needed afaik.

> DRI_PRIME=1 AMD_DEBUG=info glxinfo | grep -E 'vram|smart'
    vram_size = 512 MB
    vram_vis_size = 512 MB
    vram_type = 8
    has_dedicated_vram = 0
    all_vram_visible = 1
    smart_access_memory = 0

    vram_size = 6144 MB
    vram_vis_size = 6128 MB
    vram_type = 9
    has_dedicated_vram = 1
    all_vram_visible = 1
    smart_access_memory = 0​

sudo dmesg | grep BAR=
[ 3.755928] [drm] Detected VRAM RAM=6128M, BAR=8192M
[ 6.833969] [drm] Detected VRAM RAM=512M, BAR=512M​

[middy]

could you run

AMD_DEBUG=info glxinfo | grep smart

and paste the result please?

Memory info:
pte_fragment_size = 2097152
gart_page_size = 4096
gart_size = 16368 MB
vram_size = 16384 MB
vram_vis_size = 16368 MB
vram_type = 9
vram_bit_width = 256
gds_size = 64 kB
gds_gfx_partition_size = 0 kB
max_alloc_size = 13108 MB
min_alloc_size = 256
address32_hi = 4294934528
has_dedicated_vram = 1
all_vram_visible = 1
smart_access_memory = 1
max_tcc_blocks = 16
num_tcc_blocks = 16
tcc_cache_line_size = 128
tcc_rb_non_coherent = 0
pc_lines = 1024
lds_size_per_workgroup = 131072
lds_alloc_granularity = 1024
lds_encode_granularity = 512
max_memory_clock = 1000
ce_ram_size = 65536
l1_cache_size = 16384
l2_cache_size = 4194304

amd 5900x / msi b550 tomahawk (latest bios / agesa) / amd 6900 xt. i am using newly released mesa 22.0 and running kernel 5.16.14 arch linux.

[gand-alf]

My results on a Zen 2 Threadripper 3960X on an Asrock TRX40 Creator board BIOS P1.73 with a RX6800, from what I learned earlier in this thread, rBAR is active:

AMD_DEBUG=info glxinfo | grep vram
vram_size = 16384 MB
vram_vis_size = 16368 MB
vram_type = 9
vram_bit_width = 256
has_dedicated_vram = 1
all_vram_visible = 1

dmesg | grep BAR
[ 0.551386] pci 0000:03:00.0: BAR 0: assigned to efifb
[ 0.564712] pci 0000:46:00.0: VF(n) BAR2 space: [mem 0xcea00000-0xcea6ffff 64bit] (contains BAR2 for 7 VFs)
[ 0.564734] pci 0000:46:00.0: VF(n) BAR4 space: [mem 0xcea80000-0xcea9bfff 64bit] (contains BAR4 for 7 VFs)
[ 3.279622] [drm] Detected VRAM RAM=16368M, BAR=16384M

uname -a
Linux Manjaro 5.15.1-1-MANJARO #1 SMP PREEMPT Sat Nov 6 14:11:10 UTC 2021 x86_64 GNU/Linux

[agd5f]

Ok, apologies if this is asking much.. but

Just so I have this straight - for the RDNA2 series of cards, we want to use amdgpu driver for best open source experience? And by MESA - ELI5 please? long time user, but with this info and google I still am at a loss:

I am getting best of both worlds? Adequate performance and no proprietary driver right? And Mesa being the directx equivelant of linux?

Is there a thread for "best optimization options" with amdgpu + latest stable mesa out there that has AMD gpu users back that you could point me / others of interest to?

I'm not exactly sure what you are asking. amdgpu is the kernel driver for AMD GPUs from the last 5-6 years or so. mesa contains the user mode drivers for OpenGL, video encode/decode, etc. These user mode drivers talk to the amdgpu kernel driver. All of them are open source and directly supported by AMD. AMD also provides packaged drivers which contain the same amdgpu and mesa drivers as what is upstream. They are just packaged for convenience. AMD also provides some closed source add on user mode driver packages which use the same kernel driver to support specific workstation applications. Unless you are using one of those workstation applications your best bet is to just use the upstream drivers. The upstream drivers are the preferred drivers for gaming and other workloads. As for optimizations, you shouldn't really need to change anything. The default configuration should be fine the the vast majority of users.

[Entzilla]

The kernel driver will always resize the BAR if possible. Mesa has additional optimizations.

Ok, apologies if this is asking much.. but

Just so I have this straight - for the RDNA2 series of cards, we want to use amdgpu driver for best open source experience? And by MESA - ELI5 please? long time user, but with this info and google I still am at a loss:

OpenGL: renderer: AMD Radeon RX 6800 XT (SIENNA_CICHLID DRM 3.42.0
5.14.2-1-MANJARO LLVM 12.0.1)
v: 4.6 Mesa 21.2.1 direct render: Yes

And:

Graphics:
Device-1: AMD Navi 21 [Radeon RX 6800/6800 XT / 6900 XT]
vendor: Sapphire Limited driver: amdgpu v: kernel bus-ID: 0f:00.0

I am getting best of both worlds? Adequate performance and no proprietary driver right? And Mesa being the directx equivelant of linux?

Is there a thread for "best optimization options" with amdgpu + latest stable mesa out there that has AMD gpu users back that you could point me / others of interest to?

[White Wolf]

Thanks all for help! So it looks like my SAM works properly.

[agd5f]

The kernel driver will always resize the BAR if possible. Mesa has additional optimizations.