Jungfraujoch/common/CUDAWrapper.cu

// SPDX-FileCopyrightText: 2024 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only

#include <fstream>
#include <thread>
#include <atomic>
#include <mutex>
#include <vector>

#include <cuda.h>

#include "CUDAWrapper.h"
#include "ThreadAffinity.h"
#include "JFJochException.h"

inline void cuda_err(cudaError_t val) {
    if (val != cudaSuccess)
        throw JFJochException(JFJochExceptionCategory::GPUCUDAError, cudaGetErrorString(val));
}

int32_t get_gpu_count() {
    int device_count;
    cudaError_t val = cudaGetDeviceCount(&device_count);
    switch (val) {
        case cudaSuccess:
            return device_count;
        case cudaErrorNoDevice:
        case cudaErrorInsufficientDriver:
            return 0;
        default:
            throw JFJochException(JFJochExceptionCategory::GPUCUDAError, cudaGetErrorString(val));
    }

}

std::vector<std::string> get_gpu_names() {
    std::vector<std::string> names;
    const int32_t count = get_gpu_count();
    names.reserve(count);
    for (int32_t i = 0; i < count; i++) {
        cudaDeviceProp prop{};
        // A device that cannot be queried still exists and still gets work, so it is listed - just
        // without a name. Losing the whole list over one unreadable device would be worse.
        if (cudaGetDeviceProperties(&prop, i) == cudaSuccess)
            names.emplace_back(prop.name);
        else
            names.emplace_back("unknown GPU");
    }
    return names;
}

void set_gpu(int32_t dev_id) {
    auto dev_count = get_gpu_count();

    // Ignore if no GPU present
    if (dev_count > 0) {
        if ((dev_id < 0) || (dev_id >= dev_count))
            throw JFJochException(JFJochExceptionCategory::InputParameterInvalid, "Device ID cannot be negative");

        cuda_err(cudaSetDevice(dev_id));
    }
}

namespace {
    std::atomic<bool> gpu_numa_binding{false};

    // The NUMA node of each device, read once.
    int GpuNumaNode(int32_t dev_id) {
        static std::mutex m;
        static std::vector<int> node;
        std::lock_guard<std::mutex> lock(m);
        if (node.empty()) {
            const int32_t count = get_gpu_count();
            node.assign(count, -1);
            for (int32_t i = 0; i < count; i++) {
                char bus_id[32] = {};
                if (cudaDeviceGetPCIBusId(bus_id, sizeof(bus_id), i) == cudaSuccess)
                    node[i] = NumaNodeOfPciDevice(bus_id);
            }
        }
        return dev_id < static_cast<int32_t>(node.size()) ? node[dev_id] : -1;
    }
}

void enable_gpu_numa_binding() {
    gpu_numa_binding = true;
}

void pin_gpu(int32_t dev_id) {
    if (get_gpu_count() == 0)
        return;
    set_gpu(dev_id);
    if (gpu_numa_binding)
        PinThreadToNumaNode(GpuNumaNode(dev_id));
}

ScopedGpuPin::ScopedGpuPin(int32_t dev_id) {
    if (get_gpu_count() == 0)
        return;
    cuda_err(cudaGetDevice(&prev_device));
    const int node = gpu_numa_binding ? GpuNumaNode(dev_id) : -1;
    if (node >= 0)
        prev_cpus = GetThreadCpus();
    set_gpu(dev_id);
    PinThreadToNumaNode(node);
}

ScopedGpuPin::~ScopedGpuPin() {
    SetThreadCpus(prev_cpus);
    if (prev_device >= 0)
        cudaSetDevice(prev_device);   // not set_gpu: a destructor must not throw
}

void pin_gpu() {
    static std::atomic<uint32_t> counter{0};
    auto dev_count = get_gpu_count();
    if (dev_count > 0)
        pin_gpu(static_cast<int32_t>(counter.fetch_add(1) % dev_count));
}

void set_gpu_blocking_sync() {
    // Through the driver's primary-context flags, which a context created later takes: the runtime's
    // cudaSetDevice creates the device's context, serially, and on four cards that was most of a
    // second before the run had started. The function is taken from the driver at run time, so the
    // program does not link against it and still starts on a machine without one.
    void *fn = nullptr;
    cudaDriverEntryPointQueryResult found{};
    if (get_gpu_count() == 0
        || cudaGetDriverEntryPointByVersion("cuDevicePrimaryCtxSetFlags", &fn, 12000, cudaEnableDefault,
                                            &found) != cudaSuccess
        || found != cudaDriverEntryPointSuccess) {
        cudaGetLastError();
        return;
    }
    const auto set_flags = reinterpret_cast<CUresult (*)(CUdevice, unsigned int)>(fn);
    for (int32_t i = 0; i < get_gpu_count(); i++)
        set_flags(i, CU_CTX_SCHED_BLOCKING_SYNC);
}

void create_gpu_contexts(int32_t first_device) {
    std::vector<std::thread> threads;
    for (int32_t i = first_device; i < get_gpu_count(); i++)
        threads.emplace_back([i] {
            // A device that fails here fails again, with its error, where it is used.
            if (cudaSetDevice(i) != cudaSuccess || cudaFree(nullptr) != cudaSuccess)
                cudaGetLastError();
        });
    for (auto &t : threads)
        t.join();
}

void release_gpu_contexts() {
    std::vector<std::thread> threads;
    for (int32_t i = 0; i < get_gpu_count(); i++)
        threads.emplace_back([i] {
            if (cudaSetDevice(i) != cudaSuccess || cudaDeviceReset() != cudaSuccess)
                cudaGetLastError();
        });
    for (auto &t : threads)
        t.join();
}

void cuda_clear_error() {
    cudaGetLastError();
}

void cuda_throw_if_context_lost() {
    // cudaFree(nullptr) frees nothing, but a lost context fails it. The last error cannot say the same:
    // once cudaGetLastError() has returned a sticky error, it and cudaPeekAtLastError() report success.
    const cudaError_t err = cudaFree(nullptr);
    if (err != cudaSuccess)
        throw JFJochException(JFJochExceptionCategory::GPUCUDAError,
                              std::string("CUDA device unusable after an unrecoverable error: ")
                              + cudaGetErrorString(err));
}