[Feature]: hipcc should try to detect --offload-arch on Windows

[ScottTodd]

Suggestion Description

This code uses rocm_agent_enumerator on Linux:

llvm-project/amd/hipcc/src/hipBin_amd.h

Lines 778 to 796 in 4e3aee9

	// No AMDGPU target specified at commandline. So look for HCC_AMDGPU_TARGET
	if (default_amdgpu_target == 1) {
	if (!var.hccAmdGpuTargetEnv_.empty()) {
	targetsStr = var.hccAmdGpuTargetEnv_;
	} else if (os != windows) {
	// Else try using rocm_agent_enumerator
	string ROCM_AGENT_ENUM;
	ROCM_AGENT_ENUM = roccmPath + "/bin/rocm_agent_enumerator";
	targetsStr = ROCM_AGENT_ENUM +" -t GPU";
	SystemCmdOut sysOut = hipBinUtilPtr_->exec(targetsStr.c_str());
	regex toReplace("\n+");
	targetsStr = hipBinUtilPtr_->replaceRegex(sysOut.out, toReplace, ",");
	}
	default_amdgpu_target = 0;
	}
	// Parse the targets collected in targetStr
	// and set corresponding compiler options.
	vector<string> targets = hipcc::utils::splitStr(targetsStr, ',');
	string GPU_ARCH_OPT = " --offload-arch=";

Windows currently has hipInfo.exe (which is similar to rocminfo). Windows and Linux have amdgpu-arch. It seems like hipcc could try using amdgpu-arch? I'm also wondering if we would want to provide rocm_agent_enumerator on Windows too, and why we have both that and amdgpu-arch.

Context: ROCm/TheRock#1571. We had a test running on Linux and Windows that omitted --offload-arch. This was passing on Linux but silently failing on Windows, until ROCm/rocm-systems@e79eaaa turned the silent failure into a crash. That PR changes the test to explicitly pass --offload-arch on both platforms using amdgpu-arch to look up the gfx target.

Operating System

Windows

GPU

N/A

ROCm Component

hipcc