[Issue]: Windows HSA_OVERRIDE_GFX_VERSION doesn't work

[umarinkovic]

Problem Description

On Linux, export HSA_OVERRIDE_GFX_VERSION={MAJOR}.{MINOR}.{PATCH} tricks ROCm into thinking the available card is of the architecture specified in the env variable, e.g. export HSA_OVERRIDE_GFX_VERSION=10.3.0 to set the current ISA to gfx1030. This is useful for people with older/unsupported cards that are of the same architecture family as a supported card.

On Windows this doesn't work.

I'm just curious as to why the hack doesn't work on Windows and is this circumventable somehow?

Operating System

Windows 10.0.26100

CPU

AMD Ryzen 9 7900X 12-Core Processor

GPU

AMD Radeon RX 6700XT

ROCm Version

7

ROCm Component

rocBLAS

Steps to Reproduce

set HSA_OVERRIDE_GFX_VERSION=X.Y.Z
run any rocBLAS test suite and it will report the actual GPU you have and not the overridden one

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[schung-amd]

Hi @umarinkovic, to clarify,

run any rocBLAS test suite and it will report the actual GPU you have and not the overridden one

Is the reported GPU what you're seeing in the output of rocblas-test, something like "Device ID 0 : AMD Radeon RX 9070 XT gfx1201"? If so, that does not change with HSA_OVERRIDE_GFX_VERSION on either Linux or Windows.

[umarinkovic]

Hi @umarinkovic, to clarify,

run any rocBLAS test suite and it will report the actual GPU you have and not the overridden one

Is the reported GPU what you're seeing in the output of rocblas-test, something like "Device ID 0 : AMD Radeon RX 9070 XT gfx1201"? If so, that does not change with HSA_OVERRIDE_GFX_VERSION on either Linux or Windows.

This is the output after running rocblas_rtest.py -t psdb and overriding the gfx version on my AMD Radeon RX 6700XT (gfx1031) on Linux:

***
*** Running: quick_pre_checkin
***
rocblas-test --gtest_output=xml --gtest_color=yes --gtest_filter=*quick*:*pre_checkin*-*known_bug*
rocBLAS info: client (OPENMP) reduced omp_set_num_threads to 110
rocBLAS warning: LD_LIBRARY_PATH override may use incompatible rocblas
rocBLAS info: Using reference library 'OpenBLAS::OpenBLAS'
rocBLAS version: 5.1.0.f005caf7c3
rocBLAS-commit-hash: a3cfc707edd382d33afaa8fcf6fc5053049869cc
Tensile-commit-hash:
hipBLASLt version: 1.1.0 commit-hash: f005caf7c3
Query device success: there are 1 devices
-------------------------------------------------------------------------------
Device ID 0 : AMD Radeon RX 6700 XT gfx1030
with 12.9 GB memory, max. SCLK 2855 MHz, max. MCLK 1000 MHz, memoryBusWidth 24 Bytes, compute capability 10.3
maxGridDimX 2147483647, sharedMemPerBlock 65.5 KB, maxThreadsPerBlock 1024, warpSize 32
-------------------------------------------------------------------------------

[schung-amd]

Interesting, I'll try to repro that. I don't see that happening on my end but my environment likely differs in some spots. Thanks for the report!