AMDGPU: Fix supported CL features

[jvesely]

Summary

Missing parentheses
Inconsistent whitespace
add has_fp64 macro

[yaxunl]

LGTM.

[jvesely]

add tests
EG support the same feature set as NI

[yaxunl]

test/Misc/amdgcn.languageOptsOpenCL.cl

5 ↗

(On Diff #57692)

#pragma OPENCL EXTENSION cl_xxx : enable

[jvesely]

remove subgroups support.
better tests.

note: I had to add OPENCL_VERSION defines manually, these can be removed once clang correctly reports the version

[jvesely]

even better tests
use existing OPENCL_C_VERSION
use -cl-std=CL for explicit OpenCL1.0

[arsenm]

lib/Basic/Targets.cpp

2096

What is the current problem with subgroups? There's no reason we can't support this

[jvesely]

I'll update this patch based on the outcome of D20447

lib/Basic/Targets.cpp

2096

no problem (other than not implementing any of the required builtins, or exposing cl2.0), since having 1 subgroup per threadgroup is perfectly OK (dunno why this is even an extension).
The bigger question is whether we want to export all potentially available extensions, or only those that are actually implemented.

[jvesely]

report only exported extensions.

int64 atomics, fp16, and 3d image writes are actually not currently exported, but should be more or less done for SI+

[arsenm]

test/Misc/r600.languageOptsOpenCL.cl

4 ↗

(On Diff #58850)

We should probably add builtin device macros for these

[jvesely]

test/Misc/r600.languageOptsOpenCL.cl

4 ↗

(On Diff #58850)

When we ran into similar problem while adding fma. We opted for feature macro(HAS_FMAF) instead of device macro. I suppose I can add HAS_FP64, but it won't have many uses since everyone checks for cl_khr_fp64.

[jvesely]

add has_fp64 macro

[arsenm]

lib/Basic/Targets.cpp

2024–2025

I don't think we need this. I want device macros for other tuning and intrinsic availability reasons. Right now there are builtins that only work on some subtargets but no way to test for that

[jvesely]

lib/Basic/Targets.cpp

2024–2025

Why not have macro per feature that determines intrinsic/optimization availability?
it was preferred last year [0]
at least on r600 it works nicer than separating EG/EG+FP64/NI_but_eg_isa/NI+FP64+CM_ISA
and the feature selection is done in two places (llvm+clang) instead of every piece of compiled code.

[0]http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20150420/127643.html

[Anastasia]

test/Misc/amdgcn.languageOptsOpenCL.cl

1 ↗

(On Diff #59080)

Negative testing?

[jvesely]

tests all extensions against expected outcome (add negative tests)
enable cl_khr_icd (works ok with mesa)

[jvesely]

ping

[arsenm]

LGTM

lib/Basic/Targets.cpp

GPU = llvm::StringSwitch<GPUKind>(Name)
       .Case("rs880",    GK_R600)
       .Case("rv670",    GK_R600_DOUBLE_OPS)
       .Case("rv710",    GK_R700)
       .Case("rv730",    GK_R700)
       .Case("rv740",    GK_R700_DOUBLE_OPS)
       .Case("rv770",    GK_R700_DOUBLE_OPS)
       .Case("palm",     GK_EVERGREEN)
       .Case("cedar",    GK_EVERGREEN)
       .Case("sumo",     GK_EVERGREEN)
       .Case("sumo2",    GK_EVERGREEN)
       .Case("redwood",  GK_EVERGREEN)
       .Case("juniper",  GK_EVERGREEN)
       .Case("hemlock",  GK_EVERGREEN_DOUBLE_OPS)
       .Case("cypress",  GK_EVERGREEN_DOUBLE_OPS)
       .Case("barts",    GK_NORTHERN_ISLANDS)
       .Case("turks",    GK_NORTHERN_ISLANDS)
       .Case("caicos",   GK_NORTHERN_ISLANDS)
       .Case("cayman",   GK_CAYMAN)
case GK_VOLCANIC_ISLANDS:
       hasFMAF = true;
       hasLDEXPF = true;
       break;
     }
 
     return true;
   }
 
-   void setSupportedOpenCLOpts() override {
-     auto &Opts = getSupportedOpenCLOpts();
-     Opts.cl_clang_storage_class_specifiers = 1;
-     Opts.cl_khr_gl_sharing = 1;
-     Opts.cl_khr_gl_event = 1;
-     Opts.cl_khr_d3d10_sharing = 1;
-     Opts.cl_khr_subgroups = 1;
 
-     if (hasFP64)
-       Opts.cl_khr_fp64 = 1;
-     if (GPU >= GK_NORTHERN_ISLANDS) {
-       Opts.cl_khr_byte_addressable_store = 1;
-       Opts.cl_khr_global_int32_base_atomics = 1;
-       Opts.cl_khr_global_int32_extended_atomics = 1;
-       Opts.cl_khr_local_int32_base_atomics = 1;
-       Opts.cl_khr_local_int32_extended_atomics = 1;
-     }
-     if (GPU >= GK_SOUTHERN_ISLANDS)
-       Opts.cl_khr_fp16 = 1;
-       Opts.cl_khr_int64_base_atomics = 1;
-       Opts.cl_khr_int64_extended_atomics = 1;
-       Opts.cl_khr_3d_image_writes = 1;
-       Opts.cl_khr_gl_msaa_sharing = 1;
+    if (GPU >= GK_SOUTHERN_ISLANDS) {
     }
+  }
 };
 
 const Builtin::Info AMDGPUTargetInfo::BuiltinInfo[] = {
 #define BUILTIN(ID, TYPE, ATTRS)                \
   { #ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr },
 #include "clang/Basic/BuiltinsAMDGPU.def"
 };
 const char * const AMDGPUTargetInfo::GCCRegNames[] = {