[NNPA] AffineLoad and AffineStore from/to zTensor without zlow.stick and zlow.unstick

src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp

@@ -139,9 +139,14 @@ void addPassesNNPA(mlir::OwningOpRef<mlir::ModuleOp> &module,
       else {
         // Partially lower Krnl ops to Affine dialect.
         addKrnlToAffinePasses(pm);
+        // Replace zlow.stick, zlow.unstick by inserting dlf16 conversion
+        // directly into affine.for loops. This must be done before
+        // normalize-memrefs so that access indices are automatically generated.
+        pm.addPass(zlow::createZLowInsertDLF16ConversionPass());
+        pm.addPass(mlir::createCanonicalizerPass());
         // Normalize MemRefs.
         normalizeMemRefsPasses(pm);
-        // Some Knrl ops, e.g. KrnlMemset, potentially exist and will be lowered
+        // Some Krnl ops, e.g. KrnlMemset, potentially exist and will be lowered
         // to Affine when its operands are normalized.
         addKrnlToAffinePasses(pm);
         // Optimizations at ZLow.

src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVMCommon.cpp

@@ -29,9 +29,12 @@ namespace zlow {
 
 ApiRegistry RegisterAllApis(MLIRContext *context) {
   auto voidTy = LLVM::LLVMVoidType::get(context);
-  auto opaquePtrTy = LLVM::LLVMPointerType::get(IntegerType::get(context, 8));
+  auto int8Ty = IntegerType::get(context, 8);
+  auto int16Ty = IntegerType::get(context, 16);
   auto int32Ty = IntegerType::get(context, 32);
   auto int64Ty = IntegerType::get(context, 64);
+  auto float32Ty = FloatType::getF32(context);
+  auto opaquePtrTy = LLVM::LLVMPointerType::get(int8Ty);
 
   // Declare API type as an enum value, its string name and an LLVM Type
   // specifying its signature.
@@ -74,9 +77,13 @@ ApiRegistry RegisterAllApis(MLIRContext *context) {
     ApiSpec(API::ZDNN_MATMUL_BCAST_OP, "zdnn_matmul_bcast_op", int32Ty, {opaquePtrTy, opaquePtrTy, opaquePtrTy, int64Ty, opaquePtrTy}, false),
     ApiSpec(API::ZDNN_CONV2D, "zdnn_conv2d", int32Ty, {opaquePtrTy, opaquePtrTy, opaquePtrTy, int64Ty, int64Ty, int64Ty, int64Ty, opaquePtrTy, opaquePtrTy}, false),
     ApiSpec(API::ZDNN_AVGPOOL2D, "zdnn_avgpool2d", int32Ty, {opaquePtrTy, int64Ty, int64Ty, int64Ty, int64Ty, int64Ty, opaquePtrTy}, false),
-    ApiSpec(API:: ZDNN_MAXPOOL2D, "zdnn_maxpool2d", int32Ty, {opaquePtrTy, int64Ty, int64Ty, int64Ty, int64Ty, int64Ty, opaquePtrTy}, false),
-    ApiSpec(API:: ZDNN_MEANREDUCE2D, "zdnn_meanreduce2d", int32Ty, {opaquePtrTy, opaquePtrTy}, false),
-    ApiSpec(API:: ZDNN_BATCHNORM, "zdnn_batchnorm", int32Ty, {opaquePtrTy, opaquePtrTy, opaquePtrTy, opaquePtrTy}, false),
+    ApiSpec(API::ZDNN_MAXPOOL2D, "zdnn_maxpool2d", int32Ty, {opaquePtrTy, int64Ty, int64Ty, int64Ty, int64Ty, int64Ty, opaquePtrTy}, false),
+    ApiSpec(API::ZDNN_MEANREDUCE2D, "zdnn_meanreduce2d", int32Ty, {opaquePtrTy, opaquePtrTy}, false),
+    ApiSpec(API::ZDNN_BATCHNORM, "zdnn_batchnorm", int32Ty, {opaquePtrTy, opaquePtrTy, opaquePtrTy, opaquePtrTy}, false),
+    // Scalar operations
+    // Note: these APIs are for testing purpose only.
+    ApiSpec(API::DLF16_TO_F32, "cnvt_1_dlf16_to_fp32", float32Ty, {int16Ty}, false),
+    ApiSpec(API::F32_TO_DLF16, "cnvt_1_fp32_to_dlf16", int16Ty, {float32Ty}, false),
   };
   // clang-format on
 

src/Accelerators/NNPA/Conversion/ZLowToLLVM/ZLowToLLVMCommon.hpp

@@ -56,7 +56,10 @@ enum class API {
   ZDNN_AVGPOOL2D,
   ZDNN_MAXPOOL2D,
   ZDNN_MEANREDUCE2D,
-  ZDNN_BATCHNORM
+  ZDNN_BATCHNORM,
+  // Scalar operations.
+  DLF16_TO_F32,
+  F32_TO_DLF16,
 };
 
 // Obtain a zDNN API for an elementwise ZLow operation.