fix(amdgpu): force sysfs hwmon only for Vega20/MI50 metrics

[johncoffee715]

Bug

MangoHud 0.8.4 displays incorrect GPU metrics (zeroed or exorbitant values) on AMD Radeon Pro VII / MI50 (Vega 20, PCI ID 0x66A1).

Root cause

The SMU gpu_metrics binary blob (/sys/bus/pci/devices/0000:05:00.0/gpu_metrics) returns incorrect values on Vega 20 / gfx906. The gpu_metrics file exists and is readable, so gpu_metrics_is_valid is set to true, but the data inside (temperatures, clocks, power, fan speed) is unreliable.

The code in amdgpu.cpp then overrides the correct sysfs hwmon values with the incorrect gpu_metrics values:

if (gpu_metrics_is_valid) {
    metrics.temp = amdgpu_common_metrics.gpu_temp_c;  // overrides correct hwmon temp
    // ...
}

Fix

Force gpu_metrics_is_valid = false so MangoHud reads exclusively from sysfs hwmon — the same data source amdgpu_top uses successfully on this hardware.

Verification

• amdgpu_top 0.11.5 reads all metrics correctly via DRM ioctl + sysfs hwmon on this GPU
• With this patch, MangoHud reads from the exact same sysfs hwmon path (/sys/bus/pci/devices/0000:05:00.0/hwmon/hwmon0/)
• All metrics remain available via sysfs: gpu_busy_percent, freq1/2_input, temp1/2/3_input, power1_input/cap, fan1_input, in0_input, mem_info_vram_total/used
• Only throttle status is lost (gpu_metrics-only field, minor)

Test environment

• GPU: AMD Radeon Pro VII (Vega 20, PCI ID 0x66A1, VBIOS 113-D1640700-100)
• OS: CachyOS Linux (Arch-based), kernel 6.x
• Driver: amdgpu (open)
• Vulkan: RADV VEGA20
• amdgpu_top 0.11.5 confirms all sysfs hwmon sensors return correct values

Alternative approach

Instead of hardcoding false, a config option (amdgpu_sysfs_only) or a device ID allowlist could be added. I chose the simple approach because Vega 20 is the only known affected ASIC, and the sysfs hwmon path is strictly more reliable on this hardware.

[johncoffee715]

Update: Added RAPL debug logging + powercap udev rule

Two additional commits pushed:

1. RAPL debug logging (cpu.cpp)

Added SPDLOG_TRACE to get_cpu_power_rapl() so the actual CPU power value is logged:

RAPL: power = 81 W (delta=4777271 uJ, dt=58520 us)

This helps users verify CPU power is being read correctly.

2. Contrib: RAPL powercap udev rule

On Intel CPUs, /sys/class/powercap/intel-rapl:0/energy_uj defaults to 0400 root:root, preventing MangoHud from reading CPU power consumption. The included udev rule (contrib/99-powercap-rapl.rules) makes it readable by the video group.

Install:

sudo contrib/install-rapl-fix.sh

Verification

Tested with MANGOHUD_LOG_LEVEL=trace vkcube:

[trace] RAPL: power = 81 W (delta=4777271 uJ, dt=58520 us)
[trace] RAPL: power = 78 W (delta=39406820 uJ, dt=500434 us)
[trace] RAPL: power = 73 W
...

CPU power now displays correctly in the overlay.