1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
//! Apple GPU monitoring (M1/M2/M3/M4 integrated GPUs)
//!
//! This module provides GPU monitoring for Apple Silicon using powermetrics
//! and Metal Performance Shaders framework.
use super::*;
use crate::error::Error;
use crate::silicon::apple::AppleGpuCounters;
/// Apple GPU implementation
pub struct AppleGpu {
index: u32,
name: String,
/// GPU core count, where `system_profiler` reported one.
///
/// `None` rather than a default. This was `u32` seeded with `8` and parsed
/// with `unwrap_or(8)`, and it reached `GpuStaticInfo::shader_cores` as
/// `Some(8)` -- a sentinel wrapped in the very type that exists to say
/// whether anything was read. `shader_cores` is `Option<u32>` precisely so
/// that an unread count can be `None`, and the reader filled it in anyway.
cores: Option<u32>,
}
impl AppleGpu {
/// Create new Apple GPU instance
///
/// Formerly also derived a `max_frequency` and `max_power` from a core-count
/// bucket (1400–1550 MHz, 20–130 W). Neither was read from the hardware, and both
/// escaped as though they were: `max_frequency` was reported as the GPU's
/// `graphics_max` clock, and `max_power` was the denominator of a power
/// `usage_percent`. Apple exposes neither figure, so both are now `None`.
pub fn new(index: u32, name: String, cores: Option<u32>) -> Self {
Self { index, name, cores }
}
/// Detect Apple GPUs
pub fn detect_gpus() -> Result<Vec<AppleGpu>, Error> {
// Use system_profiler to get GPU info
let output = std::process::Command::new("system_profiler")
.args(["-detailLevel", "basic", "SPDisplaysDataType"])
.env("LC_ALL", "C")
.output()
.map_err(|e| Error::CommandExecutionFailed(format!("system_profiler: {}", e)))?;
if !output.status.success() {
return Err(Error::CommandExecutionFailed(
"system_profiler could not enumerate displays".into(),
));
}
Ok(Self::parse_inventory(&String::from_utf8_lossy(
&output.stdout,
)))
}
fn parse_inventory(text: &str) -> Vec<Self> {
let mut gpus: Vec<Self> = Vec::new();
let mut current = None;
for line in text.lines() {
let Some((key, value)) = line.trim().split_once(':') else {
continue;
};
match key {
"Chipset Model" => {
current = None;
if value.trim().starts_with("Apple ") {
current = Some(gpus.len());
gpus.push(Self::new(
gpus.len() as u32,
format!("{} GPU", value.trim()),
None,
));
}
}
"Total Number of Cores" => {
if let Some(index) = current {
gpus[index].cores = value.trim().parse::<u32>().ok().filter(|n| *n > 0);
}
}
_ => {}
}
}
gpus
}
/// This GPU's entry among IOKit's accelerators: the one whose `model` is
/// this GPU's chipset, or the only one when this is the only Apple GPU.
fn counters_in<'a>(&self, gpus: &'a [AppleGpuCounters]) -> Option<&'a AppleGpuCounters> {
let chipset = self.name.strip_suffix(" GPU").unwrap_or(&self.name);
gpus.iter()
.find(|gpu| gpu.model.as_deref() == Some(chipset))
.or_else(|| (self.index == 0 && gpus.len() == 1).then(|| &gpus[0]))
}
/// The live figures IOKit's counters carry.
fn dynamic_from(counters: &AppleGpuCounters) -> GpuDynamicInfo {
let mut vendor_specific = Vec::new();
if let Some(tiler) = counters.tiler_utilization {
vendor_specific.push(("tiler".to_string(), tiler));
}
GpuDynamicInfo {
utilization: counters.device_utilization,
// The GPU has no memory of its own; it uses the SoC's unified
// memory, and this is how much of it. There is no GPU total to
// divide by -- the share the GPU may take is the OS's to decide --
// so neither free nor a percentage is reported.
memory: GpuMemory {
used: counters.in_use_system_memory,
..GpuMemory::unreported()
},
clocks: GpuClocks {
graphics: None, // Needs `powermetrics`, which needs root
graphics_max: None, // Apple does not publish a GPU clock ceiling
memory: None, // Unified memory, no separate memory clock
memory_max: None,
sm: None,
video: None,
},
power: GpuPower {
draw: None, // Needs `powermetrics`, which needs root
limit: None,
default_limit: None,
// A real draw divided by a guessed ceiling is not a measurement.
// This used to divide by a `max_power` picked from a core-count
// bucket, so the percentage looked derived from a measurement
// while its denominator was invented.
usage_percent: None,
},
thermal: GpuThermal {
temperature: None,
// Apple publishes no GPU thermal limits. These were 100/110 with
// a "conservative estimate" comment, which reached callers as
// read thresholds.
max_temperature: None,
critical_temperature: None,
fan_speed: None,
fan_rpm: None,
},
pcie: PcieLinkInfo {
current_gen: None, // Integrated GPU, no PCIe
current_width: None,
max_gen: None,
max_width: None,
current_speed: None,
max_speed: None,
tx_throughput: None,
rx_throughput: None,
},
engines: GpuEngines {
graphics: counters.renderer_utilization,
compute: None,
encoder: None,
decoder: None,
copy: None,
vendor_specific,
},
processes: Vec::new(),
}
}
}
#[cfg(feature = "apple")]
impl Gpu for AppleGpu {
fn vendor(&self) -> GpuVendor {
GpuVendor::Apple
}
fn index(&self) -> usize {
self.index as usize
}
fn name(&self) -> Result<String, crate::Error> {
Ok(self.name.clone())
}
fn static_info(&self) -> Result<GpuStaticInfo, crate::Error> {
Ok(GpuStaticInfo {
index: self.index as usize,
vendor: GpuVendor::Apple,
name: self.name.clone(),
pci_bus_id: None,
uuid: None,
vbios_version: None,
driver_version: None,
compute_capability: None,
shader_cores: self.cores,
l2_cache: None,
num_engines: None,
integrated: Some(true), // Apple Silicon GPUs are always integrated
})
}
fn dynamic_info(&self) -> Result<GpuDynamicInfo, crate::Error> {
// IOKit's counters for this GPU, which need no root. This read only
// `powermetrics`, which does, so without root an Apple GPU had no live
// figure at all.
let gpus = crate::silicon::apple::read_gpu_counters()
.map_err(crate::Error::CommandExecutionFailed)?;
let counters = self.counters_in(&gpus).ok_or_else(|| {
crate::Error::DeviceNotFound(format!("IOKit lists no IOAccelerator for {}", self.name))
})?;
Ok(Self::dynamic_from(counters))
}
fn processes(&self) -> Result<Vec<GpuProcess>, crate::Error> {
// Process tracking is not available via powermetrics, so no list was
// read. This used to return an empty list, which reads as "none".
Err(crate::Error::NotSupported(
concat!(
"GPU process lists are not read on macOS; ",
"powermetrics reports no per-process GPU use"
)
.to_string(),
))
}
fn kill_process(&self, _pid: u32) -> Result<(), crate::Error> {
Err(crate::Error::Unsupported(
"Process killing not supported on Apple GPUs".to_string(),
))
}
fn set_power_limit(&mut self, _limit_mw: u32) -> Result<(), crate::Error> {
Err(crate::Error::Unsupported(
"Power limit control not supported on Apple GPUs".to_string(),
))
}
}
#[cfg(test)]
mod inventory_tests {
use super::AppleGpu;
use crate::silicon::apple::AppleGpuCounters;
/// IOKit's counters become the GPU's live figures, matched to the GPU by
/// chipset, and what they do not carry stays unread.
#[test]
fn iokit_counters_are_this_gpus_live_figures() {
let gpu = AppleGpu::new(0, "Apple M1 GPU".into(), Some(8));
let other = AppleGpuCounters {
model: Some("Apple Paravirtual device".into()),
..Default::default()
};
let mine = AppleGpuCounters {
model: Some("Apple M1".into()),
device_utilization: Some(12),
renderer_utilization: Some(11),
tiler_utilization: Some(7),
in_use_system_memory: Some(423_624_704),
..Default::default()
};
let gpus = [other, mine.clone()];
assert_eq!(gpu.counters_in(&gpus), Some(&mine));
let dynamic = AppleGpu::dynamic_from(&mine);
assert_eq!(dynamic.utilization, Some(12));
assert_eq!(dynamic.engines.graphics, Some(11));
assert_eq!(
dynamic.engines.vendor_specific,
vec![("tiler".to_string(), 7)]
);
assert_eq!(dynamic.memory.used, Some(423_624_704));
assert_eq!(
dynamic.memory.total, None,
"unified memory has no GPU total"
);
assert_eq!(dynamic.memory.utilization, None);
assert_eq!(dynamic.clocks.graphics, None);
assert_eq!(dynamic.power.draw, None);
// Two GPUs neither of which is this one: no match, not a guess.
let strangers = [AppleGpuCounters::default(), AppleGpuCounters::default()];
assert_eq!(gpu.counters_in(&strangers), None);
// The only accelerator, for the only Apple GPU, is the same device.
let lone = [AppleGpuCounters::default()];
assert!(gpu.counters_in(&lone).is_some());
}
#[test]
fn apple_inventory_uses_enumerated_devices_and_keeps_unknown_core_counts() {
let gpus = AppleGpu::parse_inventory("Chipset Model: Apple M1\nTotal Number of Cores: 8\nChipset Model: Intel Iris\nTotal Number of Cores: 24\nChipset Model: Apple M4\n");
assert_eq!(gpus.len(), 2);
assert_eq!(gpus[0].name, "Apple M1 GPU");
assert_eq!(gpus[0].cores, Some(8));
assert_eq!(gpus[1].cores, None);
assert!(AppleGpu::parse_inventory("").is_empty());
assert!(AppleGpu::parse_inventory("Chipset Model: Intel Iris\n").is_empty());
}
}