experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[ROCm] Populating memory bandwidth information in the DeviceDescription

Browse Source 
Currently when running on the ROCm platform, the `deviceMemoryBandwidth` information displayed is incorrect.
for e.g
```
2020-12-04 14:13:18.503661: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1738] Found device 0 with properties:
pciBusID: 0000:1c:00.0 name: Vega 10 [Radeon Instinct MI25]     ROCm AMD GPU ISA: gfx900
coreClock: 1.5GHz coreCount: 64 deviceMemorySize: 15.98GiB deviceMemoryBandwidth: -1B/s
```

This commit fixes that by quering the GPU for that information, and populating it correctly in TF
```
2020-12-04 19:41:12.864439: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1745] Found device 0 with properties:
pciBusID: 0000:63:00.0 name: Vega 10 [Radeon Instinct MI25]     ROCm AMD GPU ISA: gfx900
coreClock: 1.5GHz coreCount: 64 deviceMemorySize: 15.98GiB deviceMemoryBandwidth: 450.61GiB/s
```
This commit is contained in:
Deven Desai 2020-04-08 22:44:41 +00:00
parent ef9f660b77
commit e1962fce81
1 changed files with 5 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/stream_executor/rocm/rocm_gpu_executor.cc
View File

@ -856,6 +856,11 @@ GpuExecutor::CreateDeviceDescription(int device_ordinal) {
float clock_rate_ghz = static_cast<float>(prop.clockRate) / 1e6;
builder.set_clock_rate_ghz(clock_rate_ghz);
// mem_bandwidth = 2 * mem_bus_width_in_bytes * mem_clock_rate_in_hz
int64 memory_bandwidth = 2 * (int64(prop.memoryBusWidth) / 8) *
(int64(prop.memoryClockRate) * 1000);
builder.set_memory_bandwidth(memory_bandwidth);
}
{