cpuinfo reports incorrect L2 cache size on ARM (Neoverse V1/V2) #369
Description
🐛 Describe the bug
cpuinfo reports incorrect L2 cache size on ARM (Neoverse V1/V2)
Downstream users relying on cpuinfo via PyTorch get under-reported L2 sizes on these ARM platforms (could affect tuning/allocations).
#include <cstddef>
#include <cstdio>
#include <cpuinfo.h>
size_t l2_bytes() {
if (!cpuinfo_initialize()) return 0;
const cpuinfo_processor* p = cpuinfo_get_current_processor();
if (!p || !p->cache.l2) return 0;
return p->cache.l2->size; // bytes
}
int main() { std::printf("%zu\n", l2_bytes()); }
g++ -O2 -std=c++17 l2c_test.cpp -I"$PYTORCH_ROOT/third_party/cpuinfo/include" -L"$PYTORCH_ROOT/build/lib" -lcpuinfo -lpthread -o l2c_test
- Observed:
- Graviton3 (Neoverse V1, expected 1 MiB per core): program prints 262144 bytes (256 KB) for the current processor.
- Graviton4 (Neoverse V2, expected 2 MiB per core): program prints 1048576 bytes (1 MiB).
- Expected:
- Graviton3: ~1 MiB per core.
- Graviton4: ~2 MiB per core.