[SYCL][Driver] Update support for HIP/AMDGCN targets in DeviceConfigFile (#15414)

This change adds new supported targets, gfx941 and gfx942. It also fixes
the supported subgroup size config value for AMD RDNA GPUs, due to the
fact that the ROCm driver does not allow support for wave64 mode in HIP,
for gfx10 and gx11 family of GPUs, which are based on the RDNA
architecture. This is a result of wavefront/cross-lane functions that
only work with wave32.

Author: GeorgeWeb

clang/test/Driver/sycl-device-traits-macros-amdgcn.cpp

@@ -50,6 +50,10 @@
 // RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-DEVICE-TRIPLE
 // RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amd_gpu_gfx940 \
 // RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
+// RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amd_gpu_gfx941 \
+// RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
+// RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amd_gpu_gfx942 \
+// RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-DEVICE-TRIPLE
 // RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amd_gpu_gfx1010 \
 // RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
@@ -156,6 +160,12 @@
 // RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx940 \
 // RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-OFFLOAD-ARCH
+// RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx941 \
+// RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-OFFLOAD-ARCH
+// RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx942 \
+// RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-OFFLOAD-ARCH
 // RUN: %clangxx -fsycl -nogpulib -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=gfx1010 \
 // RUN:   -fsycl-libspirv-path=%S/Inputs/SYCL/libspirv.bc -### %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=CHECK-SYCL-AMDGCN-AMD-AMDHSA-OFFLOAD-ARCH

llvm/include/llvm/SYCLLowerIR/DeviceConfigFile.td

@@ -275,7 +275,11 @@ class HipTargetInfo<string targetName, list<Aspect> aspectList, list<int> subGro
 defvar HipSubgroupSizesGCN2 = [16];     // gfx7
 defvar HipSubgroupSizesGCN3 = [16];     // gfx8, GCN 3rd gen and 4th gen have the same subgroup sizes
 defvar HipSubgroupSizesGCN5 = [64];     // gfx900-gfx906 GCN5.0 (known as "Vega"), gfx90c GCN5.1 (known as "Vega 7nm")
-defvar HipSubgroupSizesRDNA = [32, 64]; // gfxX10-gfx11 (encapsulates RDNA1..3), natively 32 (64-waves mode available)
+// According to the "Accelerator and GPU hardware specifications table" docs,
+// (see: https://rocm.docs.amd.com/en/latest/reference/gpu-arch-specs.html)
+// the ROCm driver selects wave32 mode for the gfx10 and gfx11 family of GPUs.
+// Also, see relevant ROCm issue: https://github.com/ROCm/hipamd/issues/59
+defvar HipSubgroupSizesRDNA = [32];     // gfxX10-gfx11 (encapsulates RDNA1..3), (wave64 mode available but not used).
 defvar HipSubgroupSizesCDNA = [64];     // gfx908, gfx90a (encapsulates CDNA1..2)
 
 defvar HipMinAspects = [AspectGpu, AspectFp64, AspectOnline_compiler, AspectOnline_linker, AspectQueue_profiling,
@@ -290,9 +294,18 @@ def : HipTargetInfo<"amd_gpu_gfx908", !listconcat(HipMinAspects, AllUSMAspects,
 def : HipTargetInfo<"amd_gpu_gfx90a", !listconcat(HipMinAspects, AllUSMAspects,
     [AspectExt_intel_device_info_uuid, AspectExt_oneapi_graph, AspectExt_oneapi_limited_graph, AspectExt_oneapi_native_assert]),
     HipSubgroupSizesCDNA>;
+// TODO: Need to verify whether device-side asserts (oneapi_native_assert) are
+// now working for the new CDNA3 gfx940, gfx941, gfx942 GPUs and fixed for the
+// other supported, gfx1030 and gfx1100, RDNA3 GPUs.
 def : HipTargetInfo<"amd_gpu_gfx940", !listconcat(HipMinAspects, AllUSMAspects,
     [AspectExt_intel_device_info_uuid, AspectExt_oneapi_graph, AspectExt_oneapi_limited_graph]),
     HipSubgroupSizesCDNA>;
+def : HipTargetInfo<"amd_gpu_gfx941", !listconcat(HipMinAspects, AllUSMAspects,
+    [AspectExt_intel_device_info_uuid, AspectExt_oneapi_graph, AspectExt_oneapi_limited_graph]),
+    HipSubgroupSizesCDNA>;
+def : HipTargetInfo<"amd_gpu_gfx942", !listconcat(HipMinAspects, AllUSMAspects,
+    [AspectExt_intel_device_info_uuid, AspectExt_oneapi_graph, AspectExt_oneapi_limited_graph]),
+    HipSubgroupSizesCDNA>;
 def : HipTargetInfo<"amd_gpu_gfx1030", !listconcat(HipMinAspects, AllUSMAspects,
     [AspectExt_intel_device_info_uuid, AspectExt_oneapi_graph, AspectExt_oneapi_limited_graph]),
     HipSubgroupSizesRDNA>;
@@ -332,7 +345,5 @@ def : HipTargetInfo<"amd_gpu_gfx1103", !listconcat(HipMinAspects, AllUSMAspects)
 def : HipTargetInfo<"amd_gpu_gfx1150", !listconcat(HipMinAspects, AllUSMAspects), HipSubgroupSizesRDNA>;
 def : HipTargetInfo<"amd_gpu_gfx1151", !listconcat(HipMinAspects, AllUSMAspects), HipSubgroupSizesRDNA>;
 // TBA
-def : HipTargetInfo<"amd_gpu_gfx941", [], []>;  // CDNA 3
-def : HipTargetInfo<"amd_gpu_gfx942", [], []>;  // CDNA 3
 def : HipTargetInfo<"amd_gpu_gfx1200", [], []>; // RDNA 4
 def : HipTargetInfo<"amd_gpu_gfx1201", [], []>; // RDNA 4