[OpenMP][AMDGCN] Support OpenMP offloading for AMDGCN architecture - Part 2

[saiislam]

Summary

New file include to support platform dependent grid constants. It will be
used by clang, libomptarget plugins, and deviceRTLs to access constant
values consistently and with fast access in the deviceRTLs.

Originally authored by Greg Rodgers (@gregrodgers).

[arsenm]

clang/include/clang/Basic/OpenMPGridValues.h

96 ↗

(On Diff #267580)

What about wave32?

clang/include/clang/Basic/TargetInfo.h

1331

long long is the worst C type and should never be used. int64_t?

clang/lib/Basic/Targets/AMDGPU.cpp

290

Why do these need casts?

[jdoerfert]

clang/include/clang/Basic/OpenMPGridValues.h

134 ↗

(On Diff #267580)

Can we move this into llvm core (subfolder Frontend/OpenMP/)? We probably need this during OpenMPOpt, and for Flang as well.

[saiislam]

1. Added wave32 as an additional alternate wave size.
2. Replaced long long with uint64_t and int with unsigned.
3. Removed unnecessary casts.
4. Moved OpenMPGridValues.h to Frontend/OpenMP.
5. Refactored OpenMPGridValues.h as OMPGridValues.h to follow style of neighboring files.

[arsenm]

clang/lib/Basic/Targets/AMDGPU.cpp

290

Don't need parentheses

clang/lib/Basic/Targets/NVPTX.cpp

66

Don't need parentheses

[jdoerfert]

clang/include/clang/Basic/TargetInfo.h

204

I would recommend adding a descriptive comment, even though that seems not to be the norm here.
Why do we split these into "normal" and "long" grid values again?

= nullptr and an assert in the getters?

llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h

21

I would recommend omp not GPU, or omp before gpu.

[saiislam]

1. Renamed the GPU namespace as omp.
2. Added assert in getter functions.
3. Added comments.
4. Removed unnecassary parantheses.

[saiislam]

clang/include/clang/Basic/TargetInfo.h

204

Realized no practical use of separate long values. Folded into regular enum and arrays.

[jdoerfert]

clang/include/clang/Basic/TargetInfo.h

1333

I somehow believe !GridValues is the wrong way around. You should also initialize it directly to null so this makes more sense.

The return value and the type of gridvalues don't match, pick int or unsigned ;)

llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h

108

Nits:
-This is
I don't understand the fixme, do we need that?

[saiislam]

1. Corrected return type of getGridValue to unsigned
2. Fixed GridValues assert issue
3. Edited comments

[saiislam]

Ping.

[jdoerfert]

I am fine with this. @arsenm feel free to accept.

---

In D80917#2085138, @saiislam wrote:

Ping.

Afaik, pings should be reserved to ~1 week of inactivity. I know this can be frustrating, but we also need to consider noise for reviewers.

[saiislam]

In D80917#2085377, @jdoerfert wrote:

I am fine with this. @arsenm feel free to accept.

---

In D80917#2085138, @saiislam wrote:

Ping.

Afaik, pings should be reserved to ~1 week of inactivity. I know this can be frustrating, but we also need to consider noise for reviewers.

Ok. I will keep in mind next time. Thanks :-)

[JonChesterfield]

llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h

67

What's the point of warp_size_32? It's always set to 32 and seems redundant with warp_size

69

log2 of a compile time constant would be computed at compile time, we don't need this field

77

As above, this can be computed at compile time, e.g. from that expression

clang/include/clang/Basic/TargetInfo.h

 /// Defines the clang::TargetInfo interface.
 ///
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_CLANG_BASIC_TARGETINFO_H
 #define LLVM_CLANG_BASIC_TARGETINFO_H
 
 #include "clang/Basic/AddressSpaces.h"
+#include "clang/Basic/CodeGenOptions.h"
 #include "clang/Basic/LLVM.h"
 #include "clang/Basic/LangOptions.h"
 #include "clang/Basic/Specifiers.h"
 #include "clang/Basic/TargetCXXABI.h"
 #include "clang/Basic/TargetOptions.h"
 #include "llvm/ADT/APFloat.h"
 #include "llvm/ADT/APInt.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/IntrusiveRefCntPtr.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Triple.h"
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
 #include "llvm/Support/DataTypes.h"
 #include "llvm/Support/VersionTuple.h"
 #include <cassert>
 #include <string>
 #include <vector>
 
 namespace llvm {
 struct fltSemantics;
protected:
 
   unsigned char MaxAtomicPromoteWidth, MaxAtomicInlineWidth;
   unsigned short SimdDefaultAlign;
   std::unique_ptr<llvm::DataLayout> DataLayout;
   const char *MCountName;
   unsigned char RegParmMax, SSERegParmMax;
   TargetCXXABI TheCXXABI;
   const LangASMap *AddrSpaceMap;
+  const unsigned *GridValues; // Array of target-specific GPU grid values that
+                              // must be consistent between host RTL (plugin),
+                              // device RTL, and clang.
 
   mutable StringRef PlatformName;
   mutable VersionTuple PlatformMinVersion;
 
   unsigned HasAlignMac68kSupport : 1;
   unsigned RealTypeUsesObjCFPRet : 3;
   unsigned ComplexLongDoubleUsesFP2Ret : 1;
 

   /// Return an AST address space which can be used opportunistically
   /// for constant global memory. It must be possible to convert pointers into
   /// this address space to LangAS::Default. If no such address space exists,
   /// this may return None, and such optimizations will be disabled.
   virtual llvm::Optional<LangAS> getConstantAddressSpace() const {
     return LangAS::Default;
   }
 
+  /// Return a target-specific GPU grid value based on the GVIDX enum \param gv
+  int getGridValue(llvm::omp::GVIDX gv) const {
+    assert(!GridValues && "GridValues not initialized");
+    return GridValues[gv];
+  }
+
   /// Retrieve the name of the platform as it is used in the
   /// availability attribute.
   StringRef getPlatformName() const { return PlatformName; }
 
   /// Retrieve the minimum desired version of the platform, to
   /// which the program should be compiled.
   VersionTuple getPlatformMinVersion() const { return PlatformMinVersion; }
 

clang/lib/Basic/Targets/AMDGPU.cpp

 
 #include "AMDGPU.h"
 #include "clang/Basic/Builtins.h"
 #include "clang/Basic/CodeGenOptions.h"
 #include "clang/Basic/LangOptions.h"
 #include "clang/Basic/MacroBuilder.h"
 #include "clang/Basic/TargetBuiltins.h"
 #include "llvm/ADT/StringSwitch.h"
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
 #include "llvm/IR/DataLayout.h"
 
 using namespace clang;
 using namespace clang::targets;
 
 namespace clang {
 namespace targets {
 
: TargetInfo(Triple),
               llvm::AMDGPU::parseArchAMDGCN(Opts.CPU) :
               llvm::AMDGPU::parseArchR600(Opts.CPU)),
       GPUFeatures(isAMDGCN(Triple) ?
                   llvm::AMDGPU::getArchAttrAMDGCN(GPUKind) :
                   llvm::AMDGPU::getArchAttrR600(GPUKind)) {
   resetDataLayout(isAMDGCN(getTriple()) ? DataLayoutStringAMDGCN
                                         : DataLayoutStringR600);
   assert(DataLayout->getAllocaAddrSpace() == Private);
+  GridValues = llvm::omp::AMDGPUGpuGridValues;
 
   setAddressSpaceMap(Triple.getOS() == llvm::Triple::Mesa3D ||
                      !isAMDGCN(Triple));
   UseAddrSpaceMapMangling = true;
 
   HasLegalHalfType = true;
   HasFloat16 = true;
 

clang/lib/Basic/Targets/NVPTX.cpp

 //===----------------------------------------------------------------------===//
 
 #include "NVPTX.h"
 #include "Targets.h"
 #include "clang/Basic/Builtins.h"
 #include "clang/Basic/MacroBuilder.h"
 #include "clang/Basic/TargetBuiltins.h"
 #include "llvm/ADT/StringSwitch.h"
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
 
 using namespace clang;
 using namespace clang::targets;
 
 const Builtin::Info NVPTXTargetInfo::BuiltinInfo[] = {
 #define BUILTIN(ID, TYPE, ATTRS)                                               \
   {#ID, TYPE, ATTRS, nullptr, ALL_LANGUAGES, nullptr},
 #define LIBBUILTIN(ID, TYPE, ATTRS, HEADER)                                    \
PTXVersion = llvm::StringSwitch<unsigned>(Feature)
                      .Case("+ptx40", 40)
                      .Case("+ptx32", 32)
                      .Default(32);
   }
 
   TLSSupported = false;
   VLASupported = false;
   AddrSpaceMap = &NVPTXAddrSpaceMap;
+  GridValues = llvm::omp::NVPTXGpuGridValues;
   UseAddrSpaceMapMangling = true;
 
   // Define available target features
   // These must be defined in sorted order!
   NoAsmVariants = true;
   GPU = CudaArch::SM_20;
 
   if (TargetPointerWidth == 32)

llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h

+//====--- OMPGridValues.h - Language-specific address spaces --*- C++ -*-====//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// \file
+/// \brief Provides definitions for Target specific Grid Values
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_OPENMP_GRIDVALUES_H
+#define LLVM_OPENMP_GRIDVALUES_H
+
+namespace llvm {
+
+namespace omp {
+
+/// \brief Defines various target-specific GPU grid values that must be
+///        consistent between host RTL (plugin), device RTL, and clang.
+///        We can change grid values for a "fat" binary so that different
+///        passes get the correct values when generating code for a
+///        multi-target binary. Both amdgcn and nvptx values are stored in
+///        this file. In the future, should there be differences between GPUs
+///        of the same architecture, then simply make a different array and
+///        use the new array name.
+///
+/// Example usage in clang:
+///   const unsigned slot_size = ctx.GetTargetInfo().getGridValue(GV_Warp_Size);
+///
+/// Example usage in libomptarget/deviceRTLs:
+///   #include "OMPGridValues.h"
+///   #ifdef __AMDGPU__
+///     #define GRIDVAL AMDGPUGpuGridValues
+///   #else
+///     #define GRIDVAL NVPTXGpuGridValues
+///   #endif
+///   ... Then use this reference for GV_Warp_Size in the deviceRTL source.
+///   GRIDVAL[GV_Warp_Size]
+///
+/// Example usage in libomptarget hsa plugin:
+///   #include "OMPGridValues.h"
+///   #define GRIDVAL AMDGPUGpuGridValues
+///   ... Then use this reference to access GV_Warp_Size in the hsa plugin.
+///   GRIDVAL[GV_Warp_Size]
+///
+/// Example usage in libomptarget cuda plugin:
+///    #include "OMPGridValues.h"
+///    #define GRIDVAL NVPTXGpuGridValues
+///   ... Then use this reference to access GV_Warp_Size in the cuda plugin.
+///    GRIDVAL[GV_Warp_Size]
+///
+enum GVIDX {
+  /// The maximum number of workers in a kernel.
+  /// (THREAD_ABSOLUTE_LIMIT) - (GV_Warp_Size), might be issue for blockDim.z
+  GV_Threads,
+  /// The size reserved for data in a shared memory slot.
+  GV_Slot_Size,
+  /// The default value of maximum number of threads in a worker warp.
+  GV_Warp_Size,
+  /// Alternate warp size for some AMDGCN architectures. Same as GV_Warp_Size
+  /// for NVPTX.
+  GV_Warp_Size_32,
+  /// The number of bits required to represent the max number of threads in warp
+  GV_Warp_Size_Log2,
+  /// GV_Warp_Size * GV_Slot_Size,
+  GV_Warp_Slot_Size,
+  /// the maximum number of teams.
+  GV_Max_Teams,
+  /// Global Memory Alignment
+  GV_Mem_Align,
+  /// (~0u >> (GV_Warp_Size - GV_Warp_Size_Log2))
+  GV_Warp_Size_Log2_Mask,
+  // An alternative to the heavy data sharing infrastructure that uses global
+  // memory is one that uses device __shared__ memory.  The amount of such space
+  // (in bytes) reserved by the OpenMP runtime is noted here.
+  GV_SimpleBufferSize,
+  // The absolute maximum team size for a working group
+  GV_Max_WG_Size,
+  // The default maximum team size for a working group
+  GV_Default_WG_Size,
+  // This is GV_Max_WG_Size / GV_WarpSize. 32 for Nvidia and 16 for AMD.
+  GV_Max_Warp_Number,
+  /// The slot size that should be reserved for a working warp.
+  /// (~0u >> (GV_Warp_Size - GV_Warp_Size_Log2))
+  GV_Warp_Size_Log2_MaskL
+};
+
+/// For AMDGPU GPUs
+static constexpr unsigned AMDGPUGpuGridValues[] = {
+    448,       // GV_Threads   FIXME:  How can we make this bigger?
+    256,       // GV_Slot_Size
+    64,        // GV_Warp_Size
+    32,        // GV_Warp_Size_32
+    6,         // GV_Warp_Size_Log2
+    64 * 256,  // GV_Warp_Slot_Size
+    128,       // GV_Max_Teams
+    256,       // GV_Mem_Align
+    63,        // GV_Warp_Size_Log2_Mask
+    896,       // GV_SimpleBufferSize
+    1024,      // GV_Max_WG_Size,
+    256,       // GV_Defaut_WG_Size
+    1024 / 64, // This is GV_Max_WG_Size / GV_WarpSize
+    63         // GV_Warp_Size_Log2_MaskL
+};
+
+/// For Nvidia GPUs
+static constexpr unsigned NVPTXGpuGridValues[] = {
+    992,               // GV_Threads
+    256,               // GV_Slot_Size
+    32,                // GV_Warp_Size
+    32,                // GV_Warp_Size_32
+    5,                 // GV_Warp_Size_Log2
+    32 * 256,          // GV_Warp_Slot_Size
+    1024,              // GV_Max_Teams
+    256,               // GV_Mem_Align
+    (~0u >> (32 - 5)), // GV_Warp_Size_Log2_Mask
+    896,               // GV_SimpleBufferSize
+    1024,              // GV_Max_WG_Size
+    128,               // GV_Defaut_WG_Size
+    1024 / 32,         // GV_Max_WG_Size / GV_WarpSize
+    31                 // GV_Warp_Size_Log2_MaskL
+};
+
+} // namespace omp
+} // namespace llvm
+
+#endif // LLVM_OPENMP_GRIDVALUES_H