[AMD] Improve shared layout for Wmma's operands (#7319)

Swizzling is always disabled for Wmma's B operand, it should be disabled
only when k dimension is not contiguous.

Both vectorSize, perPhase and maxPhase are now determined using a
heuristic approach.
---------

Co-authored-by: Lixun Zhang <[email protected]>

Author: leeliu103

include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td

@@ -286,25 +286,10 @@ When vec=2, elements are swizzled in pairs of 2.  In other words, the element at
         }
 
         // ---- begin WMMA ----
-        if (mlir::isa<AMDWmmaEncodingAttr>(dotOpEnc.getParent())) {
-          if (dotOpEnc.getOpIdx() == 0) {
-            const int numBanks = 32;
-            const int bankBitWidth = 32;
-
-            // number of inner dimension rows per one pattern repeat
-            int innerDimLength = shape[order[0]];
-            int elemsPerOneBanksRow = (numBanks * bankBitWidth) / typeWidthInBit;
-
-            int perPhase = std::max(1, elemsPerOneBanksRow / innerDimLength);
-            int vecSize = ((typeWidthInBit == 16) ? 64 : 32 ) / typeWidthInBit;
-            int maxPhase = 16 / perPhase;
-
-            return get(context, vecSize, perPhase, maxPhase, order, CTALayout);
-          } else {
-            // Do not swizzle in case k dimension is not innermost.
-            // In this case accesses will go in different banks even without swizzling.
-            return get(context, 1, 1, 1, order, CTALayout);
-          }
+        if (auto wmmaEnc = mlir::dyn_cast<AMDWmmaEncodingAttr>(dotOpEnc.getParent())) {
+          return wmmaEnc.composeSharedLayoutForOperand(
+              CTALayout, dotOpEnc.getOpIdx(), shape, order, dotOpEnc.getKWidth(),
+              typeWidthInBit, needTrans);
         }
 
 
@@ -1230,6 +1215,13 @@ Row |
                                           Type elemType, int kWidth, int kDim, int opIdx) const;
     SmallVector<unsigned> getRepOrderForOperand(int opIdx) const;
     static SmallVector<unsigned> getMNKDimPerInstr();
+
+    // Returns a swizzled shared layout matching this WMMA layout for the
+    // dot operand at the given |operandIdx| with |operandShape|.
+    SwizzledSharedEncodingAttr composeSharedLayoutForOperand(
+        CTALayoutAttr ctaLayout, int operandIdx, ArrayRef<int64_t> operandShape,
+        ArrayRef<unsigned> sharedOrder, unsigned kWidth,
+        unsigned elemBitWidth, bool needTrans) const;
   }];
 }
 

lib/Dialect/TritonGPU/IR/Dialect.cpp

@@ -2031,6 +2031,42 @@ SmallVector<unsigned> AMDWmmaEncodingAttr::getMNKDimPerInstr() {
   return {16, 16, 16};
 }
 
+SwizzledSharedEncodingAttr AMDWmmaEncodingAttr::composeSharedLayoutForOperand(
+    CTALayoutAttr ctaLayout, int operandIdx, ArrayRef<int64_t> operandShape,
+    ArrayRef<unsigned> sharedOrder, unsigned kWidth, unsigned elemBitWidth,
+    bool needTrans) const {
+  int kDimIndex = operandIdx == 0 ? 1 : 0;
+  bool isKContig = sharedOrder[0] == kDimIndex;
+
+  if (!isKContig) {
+    // Do not swizzle. In this case accesses will go in different banks even
+    // without swizzling.
+    return SwizzledSharedEncodingAttr::get(getContext(), 1, 1, 1, sharedOrder,
+                                           ctaLayout);
+  }
+
+  // max vectorization size for ds_load is 128 bits
+  int vectorSize = std::min(kWidth * elemBitWidth, 128u) / elemBitWidth;
+
+  const int numBanks = 32;
+  const int bankBitWidth = 32;
+
+  // Number of inner dimension rows per one pattern repeat
+  int innerDimLength = operandShape[sharedOrder[0]];
+  int elemsPerOneBanksRow = (numBanks * bankBitWidth) / elemBitWidth;
+
+  int perPhase = std::max(1, elemsPerOneBanksRow / innerDimLength);
+  // for both RDNA3 and RDNA4, the M/N dimension of wmma is 16
+  // This represents the max number of rows that can be accessed
+  // at the same time
+  int mDim = getMNKDimPerInstr()[0];
+  int maxPhase =
+      std::max(std::min(mDim / perPhase, innerDimLength / vectorSize), 1);
+
+  return SwizzledSharedEncodingAttr::get(getContext(), vectorSize, perPhase,
+                                         maxPhase, sharedOrder, ctaLayout);
+}
+
 //===----------------------------------------------------------------------===//
 // Mma encoding
 //===----------------------------------------------------------------------===//