[dcompute] add support for OpenCL image I/O

.cirrus.yml

@@ -183,7 +183,9 @@ task:
     rm llvm.tar.xz
   clone_submodules_early_script: |
     cd $CIRRUS_WORKING_DIR
-    git submodule update --init --depth $CIRRUS_CLONE_DEPTH
+    # remove --depth from following command because of
+    # https://github.com/ldc-developers/ldc/pull/3835#issuecomment-928517319
+    git submodule update --init # --depth $CIRRUS_CLONE_DEPTH
   remove_some_tests_script: |
     cd $CIRRUS_WORKING_DIR/tests/d2/dmd-testsuite
     # The -lowmem tests don't work with an ltsmaster host compiler

dmd/id.d

@@ -554,6 +554,7 @@ immutable Msgtable[] msgtable =
     { "dcPointer", "Pointer" },
     { "dcReflect", "__dcompute_reflect" },
     { "RTInfoImpl" },
+    { "opencl" },
 
     // IN_LLVM
     { "io" },

dmd/id.h

@@ -77,6 +77,7 @@ struct Id
     static Identifier *LDC_extern_weak;
     static Identifier *LDC_profile_instr;
     static Identifier *dcReflect;
+    static Identifier *opencl;
     static Identifier *criticalenter;
     static Identifier *criticalexit;
     static Identifier *attributes;

gen/dcompute/druntime.cpp

@@ -18,7 +18,7 @@
 #include "dmd/module.h"
 #include "dmd/template.h"
 
-bool isFromLDC_DCompute(Dsymbol *sym) {
+bool isFromLDC_Mod(Dsymbol *sym, Identifier* id) {
   auto mod = sym->getModule();
   if (!mod)
     return false;
@@ -31,7 +31,14 @@ bool isFromLDC_DCompute(Dsymbol *sym) {
   if (moduleDecl->packages.ptr[0] != Id::ldc)
     return false;
 
-  return moduleDecl->id == Id::dcompute;
+  return moduleDecl->id == id;
+}
+
+bool isFromLDC_DCompute(Dsymbol *sym) {
+  return isFromLDC_Mod(sym,Id::dcompute);
+}
+bool isFromLDC_OpenCL(Dsymbol *sym) {
+  return isFromLDC_Mod(sym,Id::opencl);
 }
 
 llvm::Optional<DcomputePointer> toDcomputePointer(StructDeclaration *sd) {

gen/dcompute/druntime.h

@@ -23,6 +23,7 @@ class Dsymbol;
 class Type;
 
 bool isFromLDC_DCompute(Dsymbol *sym);
+bool isFromLDC_OpenCL(Dsymbol *sym);
 
 struct DcomputePointer {
   int addrspace;

gen/dcompute/targetOCL.cpp

@@ -12,6 +12,7 @@
 #include "dmd/id.h"
 #include "dmd/identifier.h"
 #include "dmd/template.h"
+#include "dmd/module.h"
 #include "gen/abi-spirv.h"
 #include "gen/dcompute/target.h"
 #include "gen/dcompute/druntime.h"
@@ -38,6 +39,7 @@
 
 namespace {
 class TargetOCL : public DComputeTarget {
+  bool usedImage;
 public:
   TargetOCL(llvm::LLVMContext &c, int oclversion)
       : DComputeTarget(c, oclversion, OpenCL, "ocl", "spv", createSPIRVABI(),
@@ -63,25 +65,35 @@ class TargetOCL : public DComputeTarget {
     // opencl.used.optional.core.features
     // opencl.compiler.options
     // opencl.enable.FP_CONTRACT (-ffast-math)
-    llvm::Metadata *SPIRVerElts[] = {
+    {
+      llvm::Metadata *SPIRVerElts[] = {
         llvm::ConstantAsMetadata::get(
-            llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), 1)),
+          llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), 1)),
         llvm::ConstantAsMetadata::get(
-            llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), 2))};
-    llvm::NamedMDNode *SPIRVerMD =
+          llvm::ConstantInt::get(llvm::Type::getInt32Ty(ctx), 2))};
+      llvm::NamedMDNode *SPIRVerMD =
         _ir->module.getOrInsertNamedMetadata("opencl.spir.version");
-    SPIRVerMD->addOperand(llvm::MDNode::get(ctx, SPIRVerElts));
-
+      SPIRVerMD->addOperand(llvm::MDNode::get(ctx, SPIRVerElts));
+    }
     // Add OpenCL version
-    llvm::Metadata *OCLVerElts[] = {
+    {
+      llvm::Metadata *OCLVerElts[] = {
         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
-            llvm::Type::getInt32Ty(ctx), tversion / 100)),
+          llvm::Type::getInt32Ty(ctx), tversion / 100)),
         llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
-            llvm::Type::getInt32Ty(ctx), (tversion % 100) / 10))};
-    llvm::NamedMDNode *OCLVerMD =
+          llvm::Type::getInt32Ty(ctx), (tversion % 100) / 10))};
+      llvm::NamedMDNode *OCLVerMD =
         _ir->module.getOrInsertNamedMetadata("opencl.ocl.version");
-
-    OCLVerMD->addOperand(llvm::MDNode::get(ctx, OCLVerElts));
+      OCLVerMD->addOperand(llvm::MDNode::get(ctx, OCLVerElts));
+    }
+    if (usedImage) {
+      llvm::NamedMDNode *OCLUsedCoreFeatures =
+        _ir->module.getOrInsertNamedMetadata("opencl.used.optional.core.features");
+      llvm::Metadata *OCLUsedCoreFeaturesElts[] = {
+        llvm::MDString::get(ctx,"cl_images")
+      };
+      OCLUsedCoreFeatures->addOperand(llvm::MDNode::get(ctx, OCLUsedCoreFeaturesElts));
+    }
   }
   enum KernArgMD {
       KernArgMD_addr_space,
@@ -99,40 +111,25 @@ class TargetOCL : public DComputeTarget {
 
     unsigned i = 0;
     // TODO: Handle Function attibutes
-    llvm::SmallVector<llvm::Metadata *, 8> kernelMDArgs;
-    kernelMDArgs.push_back(llvm::ConstantAsMetadata::get(llf));
-    // MDNode for the kernel argument address space qualifiers.
 
     std::array<llvm::SmallVector<llvm::Metadata *, 8>,count_KernArgMD> paramArgs;
-    std::array<const char*,count_KernArgMD> args = {{
+    std::array<llvm::StringRef,count_KernArgMD> args = {{
       "kernel_arg_addr_space",
       "kernel_arg_access_qual",
       "kernel_arg_type",
       "kernel_arg_type_qual",
       "kernel_arg_base_type",
       "kernel_arg_name"
     }};
-
-    for (auto md : args) {
-      paramArgs[i].push_back(llvm::MDString::get(ctx, md));
-      i++;
-    }
 
     VarDeclarations *vs = fd->parameters;
     for (i = 0; i < vs->length; i++) {
       VarDeclaration *v = (*vs)[i];
       decodeTypes(paramArgs, v);
     }
-
-    for (auto &md : paramArgs)
-      kernelMDArgs.push_back(llvm::MDNode::get(ctx, md));
-    ///-------------------------------
-    /// TODO: Handle Function attibutes
-    ///-------------------------------
-    llvm::MDNode *kernelMDNode = llvm::MDNode::get(ctx, kernelMDArgs);
-    llvm::NamedMDNode *OpenCLKernelMetadata =
-        _ir->module.getOrInsertNamedMetadata("opencl.kernels");
-    OpenCLKernelMetadata->addOperand(kernelMDNode);
+    for (i = 0; i < count_KernArgMD; i++) {
+      llf->setMetadata(args[i],llvm::MDNode::get(ctx,paramArgs[i]));
+    }
   }
 
   std::string mod2str(MOD mod) {
@@ -162,7 +159,7 @@ class TargetOCL : public DComputeTarget {
     return ss.str();
   }
 
-  void decodeTypes(std::array<llvm::SmallVector<llvm::Metadata *, 8>,count_KernArgMD> attrs,
+  void decodeTypes(std::array<llvm::SmallVector<llvm::Metadata *, 8>,count_KernArgMD>& attrs,
                    VarDeclaration *v)
   {
     llvm::Optional<DcomputePointer> ptr;
@@ -174,8 +171,29 @@ class TargetOCL : public DComputeTarget {
     if (v->type->ty == TY::Tstruct &&
         (ptr = toDcomputePointer(static_cast<TypeStruct *>(v->type)->sym))) {
       addrspace = ptr->addrspace;
-      tyName = basicTypeToString(ptr->type) + "*";
-      baseTyName = tyName;
+      tyName = basicTypeToString(ptr->type);
+
+      auto ts = ptr->type->isTypeStruct();
+      if (ts && isFromLDC_OpenCL(ts->sym)) {
+        // TODO: Pipes
+        auto name = std::string(ts->toChars());
+        if (!usedImage && name.rfind("ldc.opencl.image",0) == 0)
+          usedImage = true;
+        // parse access qualifiers from ldc.opencl.*_XX_t types like ldc.opencl.image1d_ro_t
+        // 4 == length of "XX_t"
+        name = name.substr(name.length()-4, 2);
+        if (name == "ro")
+          accessQual = "read_only";
+        else if (name == "wo")
+          accessQual = "write_only";
+        else if (name == "rw")
+          accessQual = "read_write";
+        // ldc.opencl.{sampler_t, reserve_id_t} do not get an access qualifier
+      } else {
+        // things like e.g. GlobalPointer!opencl.image1d_[ro,wo,rw]_t/sampler_t must not have the additional *
+        // but all others e.g. GlobalPointer!int must
+        tyName += "*";
+      }
       // there is no volatile or restrict (yet) in D
       typeQuals = mod2str(ptr->type->mod);
       // TODO: Images and Pipes They are global pointers to opaques

ir/irtypestruct.cpp

@@ -56,6 +56,12 @@ IrTypeStruct *IrTypeStruct::get(StructDeclaration *sd) {
 
   // if it's a forward declaration, all bets are off, stick with the opaque
   if (sd->sizeok != Sizeok::done) {
+    // but rewrite the name for special OpenCL types
+    if (isFromLDC_OpenCL(sd)) {
+      const int prefixlen = 4; // == sizeof("ldc.")
+      LLStructType *st = static_cast<LLStructType *>(t->type);
+      st->setName(st->getName().substr(prefixlen));
+    }
     return t;
   }
 

tests/codegen/dcompute_cl_images.d

@@ -0,0 +1,43 @@
+// REQUIRES: target_SPIRV
+// RUN: %ldc -c -mdcompute-targets=ocl-220 -m64 -I%S/inputs -mdcompute-file-prefix=%t -output-ll -output-o %s && FileCheck %s < %t_ocl220_64.ll
+@compute(CompileFor.deviceOnly) module dcompute_cl_images;
+import ldc.dcompute;
+import ldc.opencl;
+
+// CHECK-DAG: %opencl.image1d_ro_t = type opaque
+// CHECK-DAG: %opencl.image1d_wo_t = type opaque
+// CHECK-DAG: %"ldc.dcompute.Pointer!(AddrSpace.Global, image1d_ro_t).Pointer" = type { %opencl.image1d_ro_t addrspace(1)* }
+// CHECK-DAG: %"ldc.dcompute.Pointer!(AddrSpace.Global, image1d_wo_t).Pointer" = type { %opencl.image1d_wo_t addrspace(1)* }
+// CHECK-DAG: %"ldc.dcompute.Pointer!(AddrSpace.Shared, sampler_t).Pointer" = type { %opencl.sampler_t addrspace(2)* }
+// CHECK-DAG: %opencl.sampler_t = type opaque
+
+pragma(mangle,"__translate_sampler_initializer")
+    Sampler makeSampler(int);
+
+pragma(mangle,"_Z11read_imagef11ocl_image1d_ro11ocl_sampleri")
+    __vector(float[4]) read(GlobalPointer!image1d_ro_t, Sampler, int);
+
+pragma(mangle,"_Z12write_imagef11ocl_image1d_woiDv4_f")
+    void write(GlobalPointer!image1d_wo_t,int,__vector(float[4]));
+
+@kernel void img(GlobalPointer!image1d_ro_t src, GlobalPointer!image1d_wo_t dst)
+{
+// CHECK: %{{[0-9+]}} = call spir_func %opencl.sampler_t addrspace(2)* @__translate_sampler_initializer(i32 0) {{.*}}
+// CHECK: %{{[0-9+]}} = call spir_func <4 x float> @_Z11read_imagef11ocl_image1d_ro11ocl_sampleri(%opencl.image1d_ro_t addrspace(1)* %.DComputePointerRewrite_arg, %opencl.sampler_t addrspace(2)* %.DComputePointerRewrite_arg1, i32 0) {{.*}}
+    auto x = src.read(makeSampler(0), 0);
+// CHECK: call spir_func void @_Z12write_imagef11ocl_image1d_woiDv4_f(%opencl.image1d_wo_t addrspace(1)* %.DComputePointerRewrite_arg2, i32 0, <4 x float> %{{[0-9+]}})
+    dst.write(0,x);
+}
+
+@kernel void img2(GlobalPointer!image1d_ro_t src, Sampler samp)
+{
+// CHECK: %{{[0-9+]}} = call spir_func <4 x float> @_Z11read_imagef11ocl_image1d_ro11ocl_sampleri(%opencl.image1d_ro_t addrspace(1)* %.DComputePointerRewrite_arg, %opencl.sampler_t addrspace(2)* %.DComputePointerRewrite_arg1, i32 0) {{.*}}
+    auto x = src.read(samp, 0);
+}
+
+// metadata
+// CHECK: !{{[0-9+]}} = !{!"read_only", !"write_only"}
+// CHECK: !{{[0-9+]}} = !{!"image1d_ro_t", !"image1d_wo_t"}
+
+// CHECK: !{{[0-9+]}} = !{!"read_only", !"none"}
+// CHECK: !{{[0-9+]}} = !{!"image1d_ro_t", !"sampler_t"}