a fix for .b0 ptx and some other code improvements

ctx.cpp

@@ -1483,6 +1483,10 @@ FunctionEmitContext::ProgramIndexVectorPTX(bool is32bits) {
     llvm::Value *__warpszm1 = BinaryOperator(llvm::Instruction::Add, __warpsz, LLVMInt32(-1), "__warpszm1");
     llvm::Value *laneIdx = BinaryOperator(llvm::Instruction::And, __tid_x, __warpszm1, "__laneidx");
     llvm::Value *index = InsertInst(llvm::UndefValue::get(LLVMTypes::Int32VectorType), laneIdx, 0, "__laneIdxV");
+#if 0
+    if (!is32bits)
+      index = ZExtInst(index, LLVMTypes::Int64VectandType);
+#endif
     return index;
 }
 

examples_ptx/ptxcc/ptx.ll

@@ -29,6 +29,7 @@ TAB [\t]*
 ".visible"      { return TOKEN_VISIBLE; }
 ".global"       { return TOKEN_GLOBAL; }
 ".param"        { return TOKEN_PARAM; }
+".b0"           { LLSETTOKEN( TOKEN_B32);}   /* fix for buggy llvm-ptx generator */
 ".b8"           { LLSETTOKEN( TOKEN_B8);}
 ".b16"          { LLSETTOKEN( TOKEN_B16);}
 ".b32"          { LLSETTOKEN( TOKEN_B32);}

examples_ptx/ptxcc/ptxcc

@@ -1,6 +1,6 @@
 #!/bin/sh
 
-PTXSRC=$1
+PTXSRC=$1__tmp_ptx.ptx
 PTXCU=$1___tmp_ptx.cu
 PTXSH=$1___tmp_ptx.sh
 
@@ -9,6 +9,7 @@ NVCCPARM=${@:2}
 DEPTX=dePTX
 NVCC=nvcc
 
+$(cat $1 | sed 's/\.b0/\.b32/g' > $PTXSRC) &&
 $DEPTX < $PTXSRC > $PTXCU &&
 $NVCC -arch=sm_35 -G -dc $NVCCPARM -dryrun $PTXCU 2>&1 | \
   sed 's/\#\$//g'| \

module.cpp

@@ -444,15 +444,36 @@ Module::AddGlobalVariable(const std::string &name, const Type *type, Expr *initE
         return;
     }
 
-#if 1
     if (g->target->getISA() == Target::NVPTX && 
+#if 0
+        !type->IsConstType()  &&
+#endif
+#if 1
         at != NULL &&
+#endif
         type->IsVaryingType())
     {
-        Error(pos, "Global \"varying\" variables are not yet supported in \"nvptx\" target.");
+      Error(pos, "Global \"varying\" variables are not yet supported in \"nvptx\" target.");
-        return;
+      return;
-    }
+#if 0
+        int nel = 32;  /* warp-size */
+        if (type->IsArrayType())
+        {
+          const ArrayType *at = CastType<ArrayType>(type);
+          /* we must scale # elements by 4, because a thread-block will run 4 warps
+           * or 128 threads.
+           * ***note-to-me***:please define these value (128threads/4warps)
+           * in nvptx-target definition
+           * instead of compile-time constants 
+           */
+          nel *= at->GetElementCount();
+          assert (!type->IsSOAType());
+          type = new ArrayType(at->GetElementType()->GetAsUniformType(), nel);
+        }
+        else
+          type = new ArrayType(type->GetAsUniformType(), nel);
 #endif
+    }
 
 
     llvm::Type *llvmType = type->LLVMType(g->ctx);

tests/operators2.ispc

@@ -27,33 +27,24 @@ struct S operator/(struct S rr, struct S rv) {
     return c;
 }
 
-#ifdef __NVPTX__
-uniform struct S _a[programCount];
-uniform struct S _b[programCount];
-uniform struct S _d[programCount];
-#define global_a _a[programIndex]
-#define global_b _b[programIndex]
-#define global_d _d[programIndex]
-#else
-struct S global_a;
-struct S global_b;
-struct S d;
-#endif
 
 export void f_f(uniform float RET[], uniform float aFOO[]) {
+    struct S a;
+    struct S b;
+    struct S d;
     int T = programIndex;
-    global_a.a = aFOO[programIndex];
+    a.a = aFOO[programIndex];
-    global_b.a = -aFOO[programIndex];
+    b.a = -aFOO[programIndex];
     if (programIndex == 3)
         off = 1;
     else
         off = 0;
     if (T % 2)
-        global_d = global_a + global_b;
+        d = a + b;
     else
-        global_d = global_a / global_b;
+        d = a / b;
 
-    RET[programIndex] = global_d.a;
+    RET[programIndex] = d.a;
 }
 
 export void result(uniform float RET[4]) {