[Do not merge] Test for investigating Jenkins test error in PR2917

docker/Dockerfile.onnx-mlir

@@ -47,7 +47,7 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
              -DCMAKE_INSTALL_MESSAGE=NEVER \
              -DONNX_MLIR_ACCELERATORS=${ACCEL} .. \
     && make -j${NPROC} \
-    && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
+#    && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
 # FLOAT16 backend tests only work on ppc64le platform at the moment
     && TEST_NOFLOAT16=${TEST_NOFLOAT16:-$([ "$(uname -m)" = "s390x" ] && echo true || \
                                          ([ "$(uname -m)" = "x86_64" ] &&  echo true || \
@@ -57,16 +57,15 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
                                ([ "$(uname -m)" = "x86_64" ] &&  echo || \
                                ([ "$(uname -m)" = "ppc64le" ] && echo || echo)))} \
     && TEST_ARGS="-mcpu=${TEST_MCPU}" \
-    && make check-docs \
-    && make check-unittest \
-    && make check-multiple-models \
     && make NPROC=${NPROC} \
             CTEST_PARALLEL_LEVEL=${NPROC} \
             TEST_NOFLOAT16=${TEST_NOFLOAT16} \
             TEST_MCPU=${TEST_MCPU} \
             TEST_ARGS="${TEST_ARGS}" \
+	    TEST_CASE_BY_USER="test_min_float16_cpu test_max_float16_cpu" \
+	    TEST_OPTLEVEL=0 \
             -j${NPROC} \
-            check-onnx-backend-numerical \
+            check-onnx-backend-dynamic \
     && make -j${NPROC} install && ldconfig \
 # Clean up
     && cd ${WORK_DIR} \

docker/Dockerfile.onnx-mlir-dev

@@ -45,7 +45,7 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
            ${ONNX_MLIR_ROOT}/utils/cppcheck_scan/cppcheck-scan.py; \
        fi \
     && make -j${NPROC} \
-    && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
+#    && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
 # FLOAT16 backend tests only work on ppc64le platform at the moment
     && TEST_NOFLOAT16=${TEST_NOFLOAT16:-$([ "$(uname -m)" = "s390x" ] && echo true || \
                                          ([ "$(uname -m)" = "x86_64" ] &&  echo true || \
@@ -56,17 +56,16 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
                   ([ "$(uname -m)" = "ppc64le" ] && echo || echo))) \
     && TEST_ARGS="-mcpu=${TEST_MCPU}" \
     && TEST_OPTLEVEL=0 \
-    && make check-docs \
-    && make check-unittest \
-    && make check-multiple-models \
     && make NPROC=${NPROC} \
             CTEST_PARALLEL_LEVEL=${NPROC} \
             TEST_NOFLOAT16=${TEST_NOFLOAT16} \
             TEST_MCPU=${TEST_MCPU} \
             TEST_ARGS="${TEST_ARGS}" \
             TEST_OPTLEVEL=${TEST_OPTLEVEL} \
+	    TEST_CASE_BY_USER="test_min_float16_cpu test_max_float16_cpu" \
+            TEST_OPTLEVEL=0 \
             -j${NPROC} \
-            check-onnx-backend-numerical \
+            check-onnx-backend-dynamic \
     && rm -f Debug/bin/*Test Debug/bin/Perf* Debug/bin/Test* \
 # When building for push event to publish the image, unshallow and
 # rename origin to upstream to make the repo a bit more dev friendly.

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp

@@ -100,4 +100,10 @@ llvm::cl::opt<bool> nnpaEnableSaturation("nnpa-saturation",
                    "Default is false."),
     llvm::cl::init(false), llvm::cl::cat(OnnxMlirCommonOptions));
 
+llvm::cl::opt<bool> nnpaDelayStickifiedConstGen(
+    "nnpa-delay-stickified-const-generation",
+    llvm::cl::desc(
+        "Generate stickified constants when storing constants to file. This "
+        "will be able to reduce peak memory usage. Default is true."),
+    llvm::cl::init(true), llvm::cl::cat(OnnxMlirCommonOptions));
 } // namespace onnx_mlir

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp

@@ -68,6 +68,7 @@ extern llvm::cl::opt<bool> profileZHighIR;
 extern llvm::cl::opt<std::string> nnpaLoadDevicePlacementFile;
 extern llvm::cl::opt<std::string> nnpaSaveDevicePlacementFile;
 extern llvm::cl::opt<bool> nnpaEnableSaturation;
+extern llvm::cl::opt<bool> nnpaDelayStickifiedConstGen;
 
 } // namespace onnx_mlir
 #endif

src/Accelerators/NNPA/Conversion/ZHighToZLow/ZHighToZLow.cpp

@@ -27,6 +27,7 @@
 #include "src/Accelerators/NNPA/Pass/NNPAPasses.hpp"
 #include "src/Accelerators/NNPA/Support/LayoutHelper.hpp"
 #include "src/Accelerators/NNPA/Support/Stickify/Convert.hpp"
+#include "src/Compiler/CompilerOptions.hpp"
 #include "src/Conversion/ONNXToKrnl/ONNXToKrnlCommon.hpp"
 #include "src/Dialect/Krnl/KrnlHelper.hpp"
 #include "src/Support/TypeUtilities.hpp"
@@ -193,25 +194,30 @@ Value insertAllocOrEmitZeroConstant(ArrayRef<IndexExpr> dims,
     ZHighStickifiedConstantOp stickifiedConstant =
         rewriter.create<ZHighStickifiedConstantOp>(loc, resType,
             /*value=*/nullptr,
+            /*stickified=*/nullptr,
+            /*allzero=*/rewriter.getBoolAttr(true),
             /*alignment=*/rewriter.getI64IntegerAttr(4096));
 
-    // Use an dense resource attribute to store stickified data.
-    // Attribute type: tensor<sizeInBytes x i8>
-    int64_t sizeInBytes =
-        affine::getIntOrFloatMemRefSizeInBytes(resType).value();
-    char *rawData = static_cast<char *>(malloc(sizeInBytes));
-    assert(rawData && "failed to allocate memory for stickified data");
-    memset(rawData, 0, sizeInBytes);
-    DenseResourceElementsAttr valueAttr = DenseUI8ResourceElementsAttr::get(
-        RankedTensorType::get({sizeInBytes}, rewriter.getI8Type()),
-        stickifiedConstant.getOperation()
-            ->getDialect()
-            ->getNamespace(), // use the dialect as the blob "hint"
-        HeapAsmResourceBlob::allocateAndCopyWithAlign(
-            llvm::ArrayRef(rawData, sizeInBytes), alignof(char)));
-    stickifiedConstant.setValueAttr(valueAttr);
-    free(rawData);
-
+    // Set all zero in value attribute here, except when stickified constants
+    // are generated at the time constants are stored to file.
+    if (!(storeConstantsToFile && nnpaDelayStickifiedConstGen)) {
+      // Use an dense resource attribute to store stickified data.
+      // Attribute type: tensor<sizeInBytes x i8>
+      int64_t sizeInBytes =
+          affine::getIntOrFloatMemRefSizeInBytes(resType).value();
+      char *rawData = static_cast<char *>(malloc(sizeInBytes));
+      assert(rawData && "failed to allocate memory for stickified data");
+      memset(rawData, 0, sizeInBytes);
+      DenseResourceElementsAttr valueAttr = DenseUI8ResourceElementsAttr::get(
+          RankedTensorType::get({sizeInBytes}, rewriter.getI8Type()),
+          stickifiedConstant.getOperation()
+              ->getDialect()
+              ->getNamespace(), // use the dialect as the blob "hint"
+          HeapAsmResourceBlob::allocateAndCopyWithAlign(
+              llvm::ArrayRef(rawData, sizeInBytes), alignof(char)));
+      stickifiedConstant.setValueAttr(valueAttr);
+      free(rawData);
+    }
     res = stickifiedConstant.getResult();
   } else {
     MultiDialectBuilder<KrnlBuilder, MathBuilder> create(rewriter, loc);
@@ -686,7 +692,7 @@ struct ZHighToZLowUnstickOpLowering : public ConversionPattern {
 };
 
 //===----------------------------------------------------------------------===//
-// Lower ZHigh Stickified Constant to KrnlGlobal
+// Lower ZHigh Stickified Constant to ZLow Stickified Constant
 //===----------------------------------------------------------------------===//
 
 struct ZHighToZLowStickifiedConstantOpLowering : public ConversionPattern {
@@ -699,7 +705,7 @@ struct ZHighToZLowStickifiedConstantOpLowering : public ConversionPattern {
   LogicalResult matchAndRewrite(Operation *op, ArrayRef<Value> operands,
       ConversionPatternRewriter &rewriter) const final {
     Location loc = op->getLoc();
-    ZHighStickifiedConstantOp stickifiedConstOp =
+    ZHighStickifiedConstantOp zhighStickifiedConstOp =
         llvm::dyn_cast<ZHighStickifiedConstantOp>(op);
 
     // Convert ZTensor type to MemRefType.
@@ -713,36 +719,57 @@ struct ZHighToZLowStickifiedConstantOpLowering : public ConversionPattern {
         affine::normalizeMemRefType(mlir::cast<MemRefType>(zMemRefType.value));
     ArrayRef<int64_t> normalizedShape = normalizedType.getShape();
 
-    // Get dense resource attribute.
-    auto blob = mlir::cast<DenseResourceElementsAttr>(
-        stickifiedConstOp.getValue().value())
-                    .getRawHandle()
-                    .getBlob();
-    assert(blob && "Expecting dense resource with a valid blob");
-    ArrayRef<char> data = blob->getData();
-
-    // Validate the stickified tensor.
-    int64_t memRefSizeInBytes = getMemRefEltSizeInBytes(normalizedType);
-    memRefSizeInBytes *= normalizedType.getNumElements();
-    assert((data.size() == static_cast<uint64_t>(memRefSizeInBytes)) &&
-           "The stickified tensor's buffer size and MemRef's size mismatched");
-
-    // Create a KrnlGlobalOp.
-    KrnlGlobalOp constantGlobal =
-        rewriter.create<KrnlGlobalOp>(loc, zMemRefType.value,
-            /*shape=*/
-            rewriter.getI64ArrayAttr(normalizedShape),
-            /*name=*/
-            rewriter.getStringAttr(
-                "constant_stickify_" + std::to_string(constantID)),
-            /*value=*/stickifiedConstOp.getValueAttr(),
-            /*offset=*/nullptr,
-            /*alignment=*/stickifiedConstOp.getAlignmentAttr());
-
+    // When generating stickified constants when storing constants to file,
+    // create ZLowStickifiedConstantOp. Otherwise, create KrnlGlobalOp.
+    Value result;
+    if (storeConstantsToFile && nnpaDelayStickifiedConstGen) {
+      // Create a ZLowStickifiedConstantOp.
+      ZLowStickifiedConstantOp constantOp =
+          rewriter.create<ZLowStickifiedConstantOp>(loc,
+              mlir::cast<MemRefType>(zMemRefType.value),
+              /*shape=*/
+              rewriter.getI64ArrayAttr(normalizedShape),
+              /*value=*/zhighStickifiedConstOp.getValueAttr(),
+              /*name=*/
+              rewriter.getStringAttr(
+                  "constant_stickify_" + std::to_string(constantID)),
+              /*stickified=*/zhighStickifiedConstOp.getStickifiedAttr(),
+              /*allzero=*/zhighStickifiedConstOp.getAllzeroAttr(),
+              /*offset=*/rewriter.getI64IntegerAttr(0),
+              /*alignment=*/zhighStickifiedConstOp.getAlignmentAttr());
+      result = constantOp.getResult();
+    } else {
+      // Get dense resource attribute.
+      auto blob = mlir::cast<DenseResourceElementsAttr>(
+          zhighStickifiedConstOp.getValue().value())
+                      .getRawHandle()
+                      .getBlob();
+      assert(blob && "Expecting dense resource with a valid blob");
+      ArrayRef<char> data = blob->getData();
+
+      // Validate the stickified tensor.
+      int64_t memRefSizeInBytes = getMemRefEltSizeInBytes(normalizedType);
+      memRefSizeInBytes *= normalizedType.getNumElements();
+      assert(
+          (data.size() == static_cast<uint64_t>(memRefSizeInBytes)) &&
+          "The stickified tensor's buffer size and MemRef's size mismatched");
+      // Create a KrnlGlobalOp.
+      KrnlGlobalOp constantOp =
+          rewriter.create<KrnlGlobalOp>(loc, zMemRefType.value,
+              /*shape=*/
+              rewriter.getI64ArrayAttr(normalizedShape),
+              /*name=*/
+              rewriter.getStringAttr(
+                  "constant_stickify_" + std::to_string(constantID)),
+              /*value=*/zhighStickifiedConstOp.getValueAttr(),
+              /*offset=*/nullptr,
+              /*alignment=*/zhighStickifiedConstOp.getAlignmentAttr());
+      result = constantOp.getResult();
+    }
     // Increment constant ID:
     constantID++;
 
-    rewriter.replaceOp(op, constantGlobal.getResult());
+    rewriter.replaceOp(op, result);
     return success();
   }
 };

src/Accelerators/NNPA/Dialect/ZHigh/ZHigh.td

@@ -867,6 +867,8 @@ def ZHighStickifiedConstantOp:ZHigh_Op<"StickifiedConstant", [Pure]> {
     the output tensor's size.
   }];
   let arguments = (ins OptionalAttr<AnyAttr>:$value,
+                       OptionalAttr<BoolAttr>:$stickified,
+                       OptionalAttr<BoolAttr>:$allzero,
                        DefaultValuedAttr<I64Attr, "4096">:$alignment);
   let results = (outs AnyZTensor:$output);
 }

src/Accelerators/NNPA/Dialect/ZLow/CMakeLists.txt

@@ -11,8 +11,16 @@ add_onnx_mlir_library(OMZLowOps
   DEPENDS
   OMZLowIncGen
   OMONNXZLowCombineIncGen
+  OMConstantOpInterface
 
   LINK_LIBS PUBLIC
   MLIRIR
+  MLIRMemRefTransforms
   OMMlirDialects
+  OMLayoutHelper
+  OMStickify
+  OMZHighOps
+
+  ACCEL_INCLUDE_DIRS PRIVATE
+  ${NNPA_INCLUDE_PATH}
   )

src/Accelerators/NNPA/Dialect/ZLow/ZLow.td

@@ -44,6 +44,7 @@ def ZMemRef : MemRefOf<[DLF16]>;
 //===----------------------------------------------------------------------===//
 
 include "mlir/Interfaces/SideEffectInterfaces.td"
+include "src/Interface/ConstantOpInterface.td"
 
 def ZLowAddOp:ZLow_Op<"add", [MemRefsNormalizable,
     DeclareOpInterfaceMethods<MemoryEffectsOpInterface>]> {
@@ -547,4 +548,21 @@ def ZLowConvertF32ToDLF16VectorOp:ZLow_Op<"vec_f32_to_dlf16", [Pure]> {
   ];
 }
 
+def ZLowStickifiedConstantOp:ZLow_Op<"stickifiedConstant", [MemRefsNormalizable,
+    DeclareOpInterfaceMethods<ConstantOpInterface>]> {
+// def ZLowStickifiedConstantOp:ZLow_Op<"stickifiedConstant", [MemRefsNormalizable]> {
+  let summary = "ZLow Stickified Constant operation";
+  let description = [{
+
+  }];
+  let arguments = (ins AnyAttr:$shape,
+                       OptionalAttr<AnyAttr>:$value,
+                       StrAttr:$name,
+                       OptionalAttr<BoolAttr>:$stickified,
+                       OptionalAttr<BoolAttr>:$allzero,
+                       OptionalAttr<I64Attr>:$offset,
+                       DefaultValuedAttr<I64Attr, "4096">:$alignment);
+  let results = (outs ZMemRef:$output);
+}
+
 #endif // ZLOW_OPS