[AMD] release/3.2.x AMD perf cherry picks (triton-lang#5191)

Cherry pick list: - triton-lang#4925 - triton-lang#5053 - triton-lang#5019 - triton-lang#5002 - triton-lang#4935 - required additional cherry picks triton-lang#4991 and triton-lang#4951 - triton-lang#4998 - triton-lang#4925 - triton-lang#5281 - triton-lang#5308 - All previous LLVM hash PRs before triton-lang#5308 --------- Co-authored-by: Ilya V <[email protected]> Co-authored-by: Lei Zhang <[email protected]> Co-authored-by: Lixun Zhang <[email protected]> Co-authored-by: Keren Zhou <[email protected]> Co-authored-by: Alexander Efimov <[email protected]> Co-authored-by: Kyle Wang <[email protected]> Co-authored-by: Jungwook Park <[email protected]> Co-authored-by: peterbell10 <[email protected]> Co-authored-by: Hongtao Yu <[email protected]> (cherry picked from commit 2d8093c)
jataylo · Dec 18, 2024 · 93f31aa · 93f31aa
1 parent e74f027
commit 93f31aa
Show file tree

Hide file tree

Showing 42 changed files with 1,556 additions and 971 deletions.
diff --git a/cmake/llvm-hash.txt b/cmake/llvm-hash.txt
@@ -1 +1 @@
-b5cc222d7429fe6f18c787f633d5262fac2e676f
+1f20eee6dc367bd202895e3eedb03974a628ef16
diff --git a/include/triton/Analysis/Utility.h b/include/triton/Analysis/Utility.h
@@ -212,11 +212,7 @@ bool cvtNeedsSharedMemory(RankedTensorType srcTy, RankedTensorType dstTy);
 
 bool atomicNeedsSharedMemory(Value result);
 
-bool isBlockedToDotShortcut(RankedTensorType &srcTy, RankedTensorType &dstT);
-
-bool isMfmaToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy);
-
-bool isMmaToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy);
+bool isBlockedToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy);
 
 // Return true if the src and dst layout match.
 bool matchMmaV3AndDotOperandLayout(RankedTensorType srcTy,

diff --git a/include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h b/include/triton/Conversion/TritonGPUToLLVM/ElementwiseOpToLLVMBase.h
@@ -18,14 +18,6 @@ namespace gpu {
 SmallVector<Value> reorderValues(const SmallVector<Value> &values, Type inType,
                                  Type ouType);
 
-SmallVector<Value> unpackI32(const SmallVector<Value> &inValues, Type srcTy,
-                             ConversionPatternRewriter &rewriter, Location loc,
-                             const LLVMTypeConverter *typeConverter);
-
-SmallVector<Value> packI32(const SmallVector<Value> &inValues, Type srcTy,
-                           ConversionPatternRewriter &rewriter, Location loc,
-                           const LLVMTypeConverter *typeConverter);
-
 Type getElementType(Value value);
 
 class MultipleOperandsRange
@@ -187,8 +179,8 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
     for (auto operand : adaptor.getOperands()) {
       auto argTy = op->getOperand(0).getType();
       auto subOperands = unpackLLElements(loc, operand, rewriter);
-      subOperands = unpackI32(subOperands, argTy, rewriter, loc,
-                              this->getTypeConverter());
+      subOperands = unpackI32s(subOperands, argTy, rewriter, loc,
+                               this->getTypeConverter());
       allOperands.resize(subOperands.size());
       for (auto v : llvm::enumerate(subOperands))
         allOperands[v.index()].push_back(v.value());
@@ -215,7 +207,7 @@ class ElementwiseOpConversionBase : public ConvertOpToLLVMPattern<SourceOp> {
     }
     resultVals = maybeDeduplicate(op, resultVals);
     resultVals =
-        packI32(resultVals, resultTy, rewriter, loc, this->getTypeConverter());
+        packI32s(resultVals, resultTy, rewriter, loc, this->getTypeConverter());
     Value view = packLLElements(loc, this->getTypeConverter(), resultVals,
                                 rewriter, resultTy);
     rewriter.replaceOp(op, view);

diff --git a/include/triton/Conversion/TritonGPUToLLVM/Utility.h b/include/triton/Conversion/TritonGPUToLLVM/Utility.h
@@ -1388,6 +1388,67 @@ inline Value getStructFromSharedMemoryObject(Location loc,
   return llvmStruct;
 }
 
+// For some reasons, LLVM's NVPTX backend inserts unnecessary (?) integer
+// instructions to pack & unpack sub-word integers.  A workaround is to
+// store the results of tensors with dot operand encodings in i32 to
+// facilitate instructions such as `ldmatrix`.
+//
+// TODO: Confirm if the problem is still there.
+inline bool requiresI32Conversion(Type type) {
+  auto tensorTy = dyn_cast<RankedTensorType>(type);
+  if (!tensorTy)
+    return false;
+  auto dotOpEnc = dyn_cast<DotOperandEncodingAttr>(tensorTy.getEncoding());
+  if (!dotOpEnc)
+    return false;
+  auto parent = dyn_cast<NvidiaMmaEncodingAttr>(dotOpEnc.getParent());
+  if (!(parent && parent.getVersionMajor() < 3))
+    return false;
+  return true;
+}
+
+inline SmallVector<Value> packI32s(const SmallVector<Value> &inValues,
+                                   Type type, RewriterBase &rewriter,
+                                   Location loc,
+                                   const LLVMTypeConverter *typeConverter) {
+  if (!requiresI32Conversion(type))
+    return inValues;
+  Type eltTy =
+      typeConverter->convertType(cast<RankedTensorType>(type).getElementType());
+
+  SmallVector<Value> outValues;
+  int vecWidth = 32 / eltTy.getIntOrFloatBitWidth();
+  auto vecTy = vec_ty(eltTy, vecWidth);
+  for (int i = 0; i < inValues.size(); i += vecWidth) {
+    Value vec = undef(vecTy);
+    for (int j = 0; j < vecWidth; j++) {
+      vec = insert_element(vec, inValues[i + j], i32_val(j));
+    }
+    outValues.push_back(bitcast(vec, i32_ty));
+  }
+  return outValues;
+}
+
+inline SmallVector<Value> unpackI32s(const SmallVector<Value> &inValues,
+                                     Type type, RewriterBase &rewriter,
+                                     Location loc,
+                                     const LLVMTypeConverter *typeConverter) {
+  if (!requiresI32Conversion(type))
+    return inValues;
+  Type eltTy =
+      typeConverter->convertType(cast<RankedTensorType>(type).getElementType());
+
+  SmallVector<Value> outValues;
+  for (auto v : inValues) {
+    auto vecTy = vec_ty(eltTy, 32 / eltTy.getIntOrFloatBitWidth());
+    auto vec = bitcast(v, vecTy);
+    for (int i = 0; i < 32 / eltTy.getIntOrFloatBitWidth(); i++) {
+      outValues.push_back(extract_element(vec, i32_val(i)));
+    }
+  }
+  return outValues;
+}
+
 inline SmallVector<Value> unpackLLElements(Location loc, Value llvmStruct,
                                            RewriterBase &rewriter) {
   assert(bool(llvmStruct) && "can not unpack null values");

diff --git a/include/triton/Dialect/Triton/IR/TritonOps.td b/include/triton/Dialect/Triton/IR/TritonOps.td
@@ -727,6 +727,10 @@ def TT_ReduceOp: TT_Op<"reduce",
       llvm::SmallVector<RankedTensorType> getInputTypes();
       llvm::SmallVector<Type> getElementTypes();
       unsigned getNumOperands();
+
+      // Returns the CombineOp iff this ReduceOp's region contains only
+      // one CombineOp other than the return, or nullptr if not applicable.
+      ::mlir::Operation *getSingleCombiner();
     }];
 }
 

diff --git a/include/triton/Tools/LinearLayout.h b/include/triton/Tools/LinearLayout.h
@@ -679,6 +679,13 @@ class LinearLayout {
   // (i.e. every input bit affects the output).
   llvm::MapVector<StringAttr, int32_t> getFreeVariableMasks() const;
 
+  // Increase an input dimension without affecting the output dimension.  The
+  // added free variables are mapped to 0, ensuring that the new input
+  // dimensions correspond directly to the existing output space.  The function
+  // errors out if `newInDimSize` is less than the current size or the new size
+  // is not a power of 2.
+  LinearLayout resize(StringAttr inDim, int32_t newInDimSize) const;
+
   std::string toString() const;
 
   friend bool operator==(LinearLayout lhs, LinearLayout rhs);

diff --git a/lib/Analysis/Allocation.cpp b/lib/Analysis/Allocation.cpp
@@ -113,7 +113,7 @@ ScratchConfig getScratchConfigForCvt(RankedTensorType srcTy,
   Attribute srcLayout = srcTy.getEncoding();
   Attribute dstLayout = dstTy.getEncoding();
 
-  assert(!isMfmaToDotShortcut(srcTy, dstTy));
+  assert(cvtNeedsSharedMemory(srcTy, dstTy));
 
   // FIXME This is NOT entirely correct
   // This should be getElemOrder, but we don't have such a method

diff --git a/lib/Analysis/Utility.cpp b/lib/Analysis/Utility.cpp
@@ -536,7 +536,7 @@ bool supportMMA(Value value, int version) {
          (elemTy.isInteger(8) && version >= 2);
 }
 
-bool isBlockedToDotShortcut(RankedTensorType &srcTy, RankedTensorType &dstTy) {
+bool isBlockedToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy) {
   auto blockedLayout = dyn_cast<BlockedEncodingAttr>(srcTy.getEncoding());
   auto dotOperandLayout = dyn_cast<DotOperandEncodingAttr>(dstTy.getEncoding());
   if (blockedLayout == nullptr || dotOperandLayout == nullptr)
@@ -605,22 +605,6 @@ bool isBlockedToDotShortcut(RankedTensorType &srcTy, RankedTensorType &dstTy) {
   return matrixDimsCompatible && bDimCompatible;
 }
 
-bool isMfmaToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy) {
-  auto mfmaLayout = dyn_cast<AMDMfmaEncodingAttr>(srcTy.getEncoding());
-  auto dotOperandLayout = dyn_cast<DotOperandEncodingAttr>(dstTy.getEncoding());
-  if (mfmaLayout == nullptr || dotOperandLayout == nullptr)
-    return false;
-  // TODO: Remove the restriction on the warpsPerCTA once chain dot testing is
-  // improved. In addition, we can enable this shortcut for regular MFMA
-  // layout when opIdx == 1.
-  return mfmaLayout.getWarpsPerCTA()[1] == 1 &&
-         dotOperandLayout.getOpIdx() == 0 && mfmaLayout.getIsTransposed() &&
-         dotOperandLayout.getKWidth() == getContigPerThread(mfmaLayout)[1] &&
-         dotOperandLayout.getParent() == mfmaLayout &&
-         (mfmaLayout.getMDim() == 32 || mfmaLayout.getMDim() == 16) &&
-         (srcTy.getElementType().isF16() || srcTy.getElementType().isBF16());
-}
-
 // For MMAV3 dotOperand layout matches mma operand for f16 and bf16 cases.
 bool matchMmaV3AndDotOperandLayout(RankedTensorType srcTy,
                                    RankedTensorType dstTy) {
@@ -655,8 +639,46 @@ std::optional<LinearLayout> minimalCvtLayout(RankedTensorType srcTy,
       toLinearLayout(dstTy.getShape(), dstTy.getEncoding());
   if (!(srcLayout.has_value() && dstLayout.has_value()))
     return std::nullopt;
+  StringAttr kRegister = StringAttr::get(ctx, "register");
+  StringAttr kLane = StringAttr::get(ctx, "lane");
+  StringAttr kWarp = StringAttr::get(ctx, "warp");
+  StringAttr kBlock = StringAttr::get(ctx, "block");
+  auto numSrcRegs = srcLayout->getInDimSize(kRegister);
+  auto numDstRegs = dstLayout->getInDimSize(kRegister);
+  // The `invertAndCompose` function will generate a layout that is injective
+  // by assigning new output dimensions to free variables.  For instance,
+  // consider a scenario where `srcLayout` has a free variable in the lane
+  // dimension, while `dstLayout` has two free variables in the lane
+  // dimension and also a larger number of registers.
+  // The injective form of `srcLayout` will add only a single additional row
+  // to the transformation matrix, whereas the injective form of `dstLayout`
+  // will add two additional rows.  This discrepancy causes misleading results
+  // because the matrices end up with a different number of rows.
+  //
+  // Take `dstLayout ⋅ srcLayout^-1` as an example:
+  //
+  //   - `injective(dstLayout)`: [n, m] → [n + 2, m]
+  //   - `injective(srcLayout)`: [n, m] → [n + 1, m]
+  //   - `injective(srcLayout)^-1`: [n + 1, m] → [m, n + 1]
+  //   - `injective(dstLayout) ⋅ injective(srcLayout)^-1`: [n + 2, m] ⋅ [m, n +
+  //   1] → [n + 2, n + 1]
+  //
+  // Here, the `(n + 1)`-th row added by `dstLayout` represents the free
+  // variable in registers, and the `(n + 2)`-th row represents the free
+  // variable in lanes.  However, the `(n + 1)`-th row added by `srcLayout`
+  // represents the free variable in lanes.  As a result, the `(n + 1)`-th row
+  // in two layouts do not correspond to the same free variable.
+  //
+  // To address this issue, we pad the free variables in `srcLayout` and
+  // `dstLayout` to ensure they have the same number of registers.  This
+  // guarantees that the resulting matrices have the same number of rows,
+  // ensuring consistency in the composition process.
+  auto numRegs = std::max(numSrcRegs, numDstRegs);
+  auto srcLayoutWithFreeRegs = srcLayout->resize(kRegister, numRegs);
+  auto dstLayoutWithFreeRegs = dstLayout->resize(kRegister, numRegs);
   // comp describes the layout function to create dst from src.
-  LinearLayout comp = dstLayout->invertAndCompose(*srcLayout);
+  LinearLayout comp =
+      dstLayoutWithFreeRegs.invertAndCompose(srcLayoutWithFreeRegs);
   // We try to quotient by the largest subspace first
   auto dims = SmallVector<StringRef>{"block", "warp", "lane", "register"};
   for (auto dim : dims) {
@@ -693,15 +715,14 @@ bool cvtNeedsWarpShuffle(RankedTensorType srcTy, RankedTensorType dstTy) {
 }
 
 bool cvtNeedsSharedMemory(RankedTensorType srcTy, RankedTensorType dstTy) {
-  // TODO(jlebar): Remove these special cases (`isMmaToDotShortcut`,
-  // `isBlockedToDotShortcut` and `isMfmaToDotShortcut`) once they're fully
-  // subsumed by the linear-layout checks.
+  // TODO(jlebar): Remove these special cases (`isBlockedToDotShortcut` and
+  // `isMfmaToDotShortcut`) once they're fully subsumed by the linear-layout
+  // checks.
   // TODO(Keren): We didn't check `cvtNeedsWarpShuffle` here because it's not
   // supported yet in Triton's backend.
   return !cvtReordersRegisters(srcTy, dstTy) &&
          !isBlockedToDotShortcut(srcTy, dstTy) &&
-         !isMmaToDotShortcut(srcTy, dstTy) &&
-         !isMfmaToDotShortcut(srcTy, dstTy);
+         !matchMmaV3AndDotOperandLayout(srcTy, dstTy);
 }
 
 bool atomicNeedsSharedMemory(Value value) {
@@ -711,20 +732,6 @@ bool atomicNeedsSharedMemory(Value value) {
   return true;
 }
 
-bool isMmaToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy) {
-  if (matchMmaV3AndDotOperandLayout(srcTy, dstTy))
-    return true;
-  // dot_op<opIdx=0, parent=#mma> = #mma
-  // when #mma = MmaEncoding<version=2, warpsPerCTA=[..., 1]>
-  auto mmaLayout = dyn_cast<NvidiaMmaEncodingAttr>(srcTy.getEncoding());
-  auto dotOperandLayout = dyn_cast<DotOperandEncodingAttr>(dstTy.getEncoding());
-  return mmaLayout && dotOperandLayout && mmaLayout.getVersionMajor() == 2 &&
-         mmaLayout.getWarpsPerCTA()[1] == 1 &&
-         dotOperandLayout.getOpIdx() == 0 &&
-         dotOperandLayout.getParent() == mmaLayout &&
-         !srcTy.getElementType().isF32();
-}
-
 namespace {
 
 /// A data structure similar to SetVector but maintains
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		b5cc222d7429fe6f18c787f633d5262fac2e676f
		1f20eee6dc367bd202895e3eedb03974a628ef16