response to comments

Signed-off-by: Alexandre Eichenberger <alexe@us.ibm.com>

src/Conversion/ONNXToKrnl/Math/Reduction.cpp

@@ -416,7 +416,6 @@ bool emitFullSIMDReductionFor(ConversionPatternRewriter &rewriter, Location loc,
   int64_t totVL =
       computeSuitableUnrollFactor(inputType, collapsedInnermostLoops, mix,
           canOverCompute, simdLoopStaticTripCount, simdOnly);
-#if 1 // new approach that enables parallelism
   // Test if loop trip count is long enough for a parallel execution.
   if (enableParallel) {
     int64_t parId;
@@ -454,6 +453,7 @@ bool emitFullSIMDReductionFor(ConversionPatternRewriter &rewriter, Location loc,
     // where each (possibly virtual) thread is responsible for one chunk.
     // Second round computes the final reduction done by one thread.
 
+    // TODO: this should not be hardwired but gotten from an option.
     int64_t tNum = 8;
 
     // Round 1.
@@ -506,81 +506,6 @@ bool emitFullSIMDReductionFor(ConversionPatternRewriter &rewriter, Location loc,
         output2, tmp1, tmp2, alloc1, alloc2, divisorForMean);
   }
 
-#else // old approach
-
-  // Compute type of small temporary reduction vector.
-  MemRefType outputType = MemRefType::get({}, elementType);
-  MemRefType redType = MemRefType::get({totVL}, elementType);
-  VectorType vecType = VectorType::get({totVL}, elementType);
-
-  SmallVector<Value, 2> inputs, tmps, outputs, initVals;
-  SmallVector<DimsExpr, 2> inputAFs, tmpAFs, outputAFs;
-  DimsExpr emptyAF;
-  DimsExpr zeroAF(1, zero);
-
-  // Initialize data for input (same for both reduction).
-  inputs.emplace_back(flatInput);
-  inputAFs.emplace_back(zeroAF);
-
-  // Init data for 1st reduction: tmp (redAlloc), output (alloc), and init.
-  tmps.emplace_back(create.mem.alignedAlloc(redType));
-  tmpAFs.emplace_back(zeroAF);
-  /*output*/ alloc1 = create.mem.alloc(outputType);
-  outputs.emplace_back(alloc1);
-  outputAFs.emplace_back(emptyAF);
-  initVals.emplace_back(getIdentityValue<ONNXReductionOp1>(
-      rewriter, create.getLoc(), elementType));
-  // Init data for 2nd reduction.
-  alloc2 = nullptr;
-  if (hasTwoRed) {
-    tmps.emplace_back(create.mem.alignedAlloc(redType));
-    tmpAFs.emplace_back(zeroAF);
-    /*output*/ alloc2 = create.mem.alloc(outputType);
-    outputs.emplace_back(alloc2);
-    outputAFs.emplace_back(emptyAF);
-    initVals.emplace_back(getIdentityValue<ONNXReductionOp2>(
-        rewriter, create.getLoc(), elementType));
-  }
-  create.krnl.simdReduceIE(
-      lb, ub, totVL, simdOnly, inputs, inputAFs, tmps, tmpAFs, outputs,
-      outputAFs, initVals,
-      /* reduction function */
-      [&](const KrnlBuilder &kb, ArrayRef<Value> inputVals,
-          ArrayRef<Value> tmpVals, llvm::SmallVectorImpl<Value> &resultVals,
-          int64_t VL) {
-        Type type = (VL > 1) ? vecType : elementType;
-        // First reduction, enqueue result.
-        Value accumulatedVec1 = emitScalarOpFor<ONNXReductionOp1>(
-            rewriter, create.getLoc(), op, type, {tmpVals[0], inputVals[0]});
-        resultVals.emplace_back(accumulatedVec1);
-        if (hasTwoRed) {
-          // Has a second reduction, also enqueue result.
-          Value accumulatedVec2 = emitScalarOpFor<ONNXReductionOp2>(
-              rewriter, create.getLoc(), op, type, {tmpVals[1], inputVals[0]});
-          resultVals.emplace_back(accumulatedVec2);
-        }
-      },
-      /* post reduction function*/
-      [&](const KrnlBuilder &kb, ArrayRef<Value> tmpVals,
-          llvm::SmallVectorImpl<Value> &scalarOutputs, int64_t VL) {
-        // Perform horizontal reductions.
-        Value res1 = create.vec.reduction(
-            getCombiningKind<ONNXReductionOp1>(), tmpVals[0]);
-        scalarOutputs.emplace_back(res1);
-        if (hasTwoRed) {
-          Value res2 = create.vec.reduction(
-              getCombiningKind<ONNXReductionOp2>(), tmpVals[1]);
-          scalarOutputs.emplace_back(res2);
-        }
-        // Handle means if any.
-        if (divideByMean<ONNXReductionOp1>())
-          scalarOutputs[0] = create.math.div(scalarOutputs[0], divisorForMean);
-        if (hasTwoRed && divideByMean<ONNXReductionOp2>())
-          scalarOutputs[1] = create.math.div(scalarOutputs[1], divisorForMean);
-      });
-
-#endif
-
   if (hasTwoRed)
     onnxToKrnlSimdReport(op, /*successful*/ true, totVL,
         simdLoopStaticTripCount, "fused reduction to a scalar");