develancer / amd-opencl-test-case

Simple test case for a common AMD OpenCL compiler error

AMD OpenCL error test case

Reproducible on RX 5700 XT with ROCm (4.0.1) drivers under Ubuntu 20.04.1 with 5.4.0-65 kernel. And many other, I guess.

The code itself is a highly modified version of a part of code from johncantrell's bip39-solver-gpu, which, in turn, is an OpenCL port of secp256k1.

How to compile?

If you have ROCm or OpenCL libraries installed in a custom directory (e.g. /opt/rocm) you'll have to modify the Makefile to reflect it. Simply replace

CFLAGS=-std=gnu99 -Wall -Wextra -Wno-deprecated-declarations -O3

with

CFLAGS=-std=gnu99 -Wall -Wextra -Wno-deprecated-declarations -O3 -I/opt/rocm/opencl/include
LDFLAGS=-L/opt/rocm/opencl/lib

Then, run

make

How to run?

USAGE: ./weird platform device [ compiler_flags ]

so, for example, if GPU card is device #1 on platform #1:

./weird 1 1

The correct computation result is

bbde464b6355ee6de6deba5ae860f8a66524937eee81dde224a0214efd795d09

but the 5700 XT produces

77262ca4b90e3fcb55f32ba92841024688802f53e75b16196c399de799377ba7

instead.

Running the kernel with no optimizations does not help, it still produces incorrect results (different, though):

./weird 1 1 -cl-opt-disable
b155da6459a3e7f864a7c1217e83f35dadb2f74d2bd1afdb22901fbec1a10f53

Why?

It seems that AMD OpenCL compiler does not correctly handle overflow when multiplying 32-bit integers cast to 64-bit. It results in the upper part of the result being truncated. There may be some other issues as well.

The previous version of this bug report is on the branch previous.