[test_locality]: Fails to run correctly on PVC GPU.

[etiotto]

The end-to-end core test test_locality fails to run correctly on a PVC GPU. We need to investigate the cause of the functional problem(s).

[alexbaden]

The problem is an unresolved function:

L0 module build log: error : unresolved external symbol _Z12get_global_sizej at offset 676 in instructions segment #0 (aka kernel : kernel_0d1d2de)
error : unresolved external symbol _Z12get_global_sizej at offset 692 in instructions segment #0 (aka kernel : kernel_0d1d2de)

The OpenCL function get_global_size cannot be found when lowering LLVM to SPIRV - specifically we fail during zeCreateModule (though the failure is not detected until we actually attempt to lauch the kernel, later).

Coming out of the MLIR pipeline in Triton we have several external, mangled (opencl?) functions. If I dump the IR from the IGC shader compilation pipeline, those mangled functions gradually disappear - except for get_global_size which is still in place.

IGC Before Unification Passes:

; Function Attrs: nounwind
declare spir_func i64 @_Z12get_local_idj(i32) #0

; Function Attrs: nounwind
declare spir_func i64 @_Z12get_group_idj(i32) #0

; Function Attrs: nounwind
declare spir_func i64 @_Z12get_global_sizej(i32) #0

; Function Attrs: nounwind
declare spir_func float @_Z21sub_group_shuffle_xorfj(float, i32) #0

; Function Attrs: nounwind
declare spir_func void @_Z7barrierj(i32) #0

; Function Attrs: nounwind
define spir_kernel void @kernel_0d1d2de(float addrspace(1)* nocapture readonly %0, float addrspace(1)* nocapture %1, i32 %2, float addrspace(3)* nocapture %3) #0 !dbg !265 {
  %5 = call spir_func i64 @_Z12get_local_idj(i32 0) #0, !dbg !268
  %6 = trunc i64 %5 to i32, !dbg !268
  %7 = and i32 %6, 31, !dbg !268
  %8 = lshr i32 %6, 3, !dbg !268
  %9 = and i32 %8, 15, !dbg !268
  %10 = or i32 %9, 16, !dbg !268
  %11 = shl i32 %6, 2, !dbg !268
  %12 = and i32 %11, 28, !dbg !268
  %13 = call spir_func i64 @_Z12get_group_idj(i32 0) #0, !dbg !269
  %14 = trunc i64 %13 to i32, !dbg !269
  %15 = call spir_func i64 @_Z12get_group_idj(i32 1) #0, !dbg !270
  %16 = trunc i64 %15 to i32, !dbg !270
  %17 = call spir_func i64 @_Z12get_global_sizej(i32 1) #0, !dbg !271
  %18 = trunc i64 %17 to i32, !dbg !271
  %19 = shl i32 %14, 5, !dbg !272
  %20 = or i32 %19, %9, !dbg !273
  %21 = or i32 %10, %19, !dbg !273
  %22 = or i32 %19, %7, !dbg !273
  %23 = add i32 %2, 31, !dbg !274
  %24 = sdiv i32 %23, 32, !dbg !278
  %25 = mul i32 %20, %2, !dbg !279
  %26 = mul i32 %21, %2, !dbg !279
  %27 = sext i32 %25 to i64, !dbg !280
  %28 = getelementptr float, float addrspace(1)* %0, i64 %27, !dbg !280
  %29 = sext i32 %26 to i64, !dbg !280
  %30 = getelementptr float, float addrspace(1)* %0, i64 %29, !dbg !280
  %31 = icmp sgt i32 %24, %16, !dbg !281
  br i1 %31, label %.lr.ph, label %._crit_edge, !dbg !281

IGC After Unification Pass(es)

; Function Attrs: nounwind
declare spir_func i64 @_Z12get_global_sizej(i32) #0

; Function Attrs: convergent nounwind
define spir_kernel void @kernel_0d1d2de(float addrspace(1)* nocapture readonly %0, float addrspace(1)* nocapture %1, i32 %2, float addrspace(3)* nocapture %3, <8 x i32> %r0, <8 x i32> %payloadHeader, i16 %localIdX, i16 %localIdY, i16 %localIdZ, i8* %privateBase, i32 %bufferOffset, i32 %bufferOffset1) #1 !dbg !327 {
  call void @llvm.genx.GenISA.CatchAllDebugLine(), !dbg !331
  %scalar = extractelement <8 x i32> %r0, i32 0
  %scalar22 = extractelement <8 x i32> %r0, i32 1
  %scalar23 = extractelement <8 x i32> %r0, i32 2
  %scalar24 = extractelement <8 x i32> %r0, i32 3
  %scalar25 = extractelement <8 x i32> %r0, i32 4
  %scalar26 = extractelement <8 x i32> %r0, i32 5
  %scalar27 = extractelement <8 x i32> %r0, i32 6
  %scalar28 = extractelement <8 x i32> %r0, i32 7
  %5 = zext i16 %localIdX to i32, !dbg !332
  %6 = and i32 %5, 31, !dbg !332
  %7 = lshr i32 %5, 3, !dbg !332
  %8 = and i32 %7, 15, !dbg !332
  %9 = or i32 %8, 16, !dbg !332
  %10 = shl nuw nsw i32 %5, 2, !dbg !332
  %11 = and i32 %10, 28, !dbg !332
  %12 = call spir_func i64 @_Z12get_global_sizej(i32 1) #7, !dbg !333
  %13 = trunc i64 %12 to i32, !dbg !333
  %14 = shl i32 %scalar22, 5, !dbg !334
  %15 = or i32 %14, %8, !dbg !335
  %16 = or i32 %9, %14, !dbg !335
  %17 = or i32 %14, %6, !dbg !335
  %18 = add i32 %2, 31, !dbg !336
  %19 = sdiv i32 %18, 32, !dbg !340
  %20 = mul i32 %15, %2, !dbg !341
  %21 = mul i32 %16, %2, !dbg !341
  %22 = sext i32 %20 to i64, !dbg !342
  %23 = getelementptr float, float addrspace(1)* %0, i64 %22, !dbg !342
  %24 = sext i32 %21 to i64, !dbg !342
  %25 = getelementptr float, float addrspace(1)* %0, i64 %24, !dbg !342
  %26 = icmp sgt i32 %19, %scalar27, !dbg !343
  br i1 %26, label %.lr.ph, label %._crit_edge, !dbg !343

[etiotto]

Closing because PR #247 is merged in.