device/cuda/adapters.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-License-Identifier: BSD-3-Clause | ||
| 2 | #include "chrono_batch_render/device/cuda/device.h" | ||
| 3 | |||
| 4 | #include <algorithm> | ||
| 5 | |||
| 6 | #include <fmt/format.h> | ||
| 7 | |||
| 8 | #include "chrono_batch_render/device/compute/compute.h" | ||
| 9 | |||
| 10 | namespace chrono::batch_render::device::cuda { | ||
| 11 | namespace { | ||
| 12 | |||
| 13 | ✗ | int Attribute(const Api& cuda, CUdevice device, CUdevice_attribute attribute) { | |
| 14 | ✗ | int value = 0; | |
| 15 | ✗ | cuda.Check(cuda.cuDeviceGetAttribute(&value, attribute, device), "cuDeviceGetAttribute"); | |
| 16 | ✗ | return value; | |
| 17 | } | ||
| 18 | |||
| 19 | ✗ | AdapterInfo Describe(const Api& cuda, int ordinal) { | |
| 20 | ✗ | CUdevice device = 0; | |
| 21 | ✗ | cuda.Check(cuda.cuDeviceGet(&device, ordinal), "cuDeviceGet"); | |
| 22 | |||
| 23 | ✗ | char name[256] = {}; | |
| 24 | ✗ | cuda.Check(cuda.cuDeviceGetName(name, int(sizeof(name)), device), "cuDeviceGetName"); | |
| 25 | ✗ | CUuuid uuid{}; | |
| 26 | ✗ | cuda.Check(cuda.cuDeviceGetUuid(&uuid, device), "cuDeviceGetUuid"); | |
| 27 | ✗ | size_t memory = 0; | |
| 28 | ✗ | cuda.Check(cuda.cuDeviceTotalMem(&memory, device), "cuDeviceTotalMem"); | |
| 29 | |||
| 30 | ✗ | compute::ComputeDevice reported = { | |
| 31 | .backend = Backend::Cuda, | ||
| 32 | ✗ | .ordinal = uint32_t(ordinal), | |
| 33 | ✗ | .name = name, | |
| 34 | .memory = memory, | ||
| 35 | ✗ | .integrated = Attribute(cuda, device, CU_DEVICE_ATTRIBUTE_INTEGRATED) != 0, | |
| 36 | ✗ | .warpSize = uint32_t(Attribute(cuda, device, CU_DEVICE_ATTRIBUTE_WARP_SIZE)), | |
| 37 | ✗ | .maxThreads = uint32_t(Attribute(cuda, device, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK)), | |
| 38 | ✗ | .textures = true}; | |
| 39 | ✗ | std::copy(std::begin(uuid.bytes), std::end(uuid.bytes), reported.uuid.begin()); | |
| 40 | ✗ | AdapterInfo adapter = compute::DescribeCompute(reported); | |
| 41 | |||
| 42 | ✗ | const Version capability{Attribute(cuda, device, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR), | |
| 43 | ✗ | Attribute(cuda, device, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR)}; | |
| 44 | ✗ | adapter.architecture = fmt::format("sm_{}{}", capability.major, capability.minor); | |
| 45 | ✗ | adapter.apiVersion = cuda.driverVersion; | |
| 46 | ✗ | adapter.driverVersion = cuda.driverVersion; | |
| 47 | |||
| 48 | ✗ | if (capability < kMinimumComputeCapability) { | |
| 49 | ✗ | adapter.unsupported = | |
| 50 | ✗ | fmt::format("compute capability is {}; Chrono Batch Render needs {} or later", | |
| 51 | ✗ | capability.ToString(), kMinimumComputeCapability.ToString()); | |
| 52 | } | ||
| 53 | ✗ | return adapter; | |
| 54 | ✗ | } | |
| 55 | |||
| 56 | } | ||
| 57 | |||
| 58 | 28 | std::vector<AdapterInfo> ListAdapters() { | |
| 59 | 28 | const Api& cuda = GetApi(); | |
| 60 | ✗ | int count = 0; | |
| 61 | ✗ | cuda.Check(cuda.cuDeviceGetCount(&count), "cuDeviceGetCount"); | |
| 62 | ✗ | std::vector<AdapterInfo> adapters; | |
| 63 | ✗ | adapters.reserve(count); | |
| 64 | ✗ | for (int ordinal = 0; ordinal < count; ++ordinal) { | |
| 65 | ✗ | adapters.push_back(Describe(cuda, ordinal)); | |
| 66 | } | ||
| 67 | ✗ | return adapters; | |
| 68 | ✗ | } | |
| 69 | |||
| 70 | } | ||
| 71 |