[Openmp-commits] [openmp] [OpenMP][AMDGPU] Adapt dynamic callstack sizes to HIP behavior (PR #74080)

[Joseph Huber via Openmp-commits]

================
@@ -1872,6 +1873,25 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     else
       return Plugin::error("Unexpected AMDGPU wavefront %d", WavefrontSize);
 
+    // To determine the correct scratch memory size per thread, we need to check
+    // the device architecure generation. According to AOT_OFFLOADARCHS we may
+    // assume that AMDGPU offload archs are prefixed with "gfx" and suffixed
+    // with a two char arch specialization. In-between is the 1-2 char
+    // generation number we want to extract.
+    StringRef Arch(ComputeUnitKind);
+    unsigned GfxGen = 0u;
+    if (!llvm::to_integer(Arch.slice(sizeof("gfx") - 1, Arch.size() - 2),
+                          GfxGen))
+      return Plugin::error("Invalid GFX architecture string");
+
+    // See: 'getMaxWaveScratchSize' in 'llvm/lib/Target/AMDGPU/GCNSubtarget.h'.
+    // But we need to divide by WavefrontSize.
+    // For generations pre-gfx11: use 13-bit field in units of 256-dword,
+    // otherwise: 15-bit field in units of 64-dword.
+    MaxThreadScratchSize = (GfxGen < 11)
+                               ? ((256 * 4) / WavefrontSize) * ((1 << 13) - 1)
+                               : ((64 * 4) / WavefrontSize) * ((1 << 15) - 1);
+
----------------
jhuber6 wrote:

That function is in a utility class, but could easily be made static or put somewhere more common. It uses an internal `generation` enum, but we could probably just convert to that here.

https://github.com/llvm/llvm-project/pull/74080