[Mlir-commits] [mlir] [mlir][Vector] Add a rewrite pattern for better low-precision ext(bit… (PR #65774)

[llvmlistbot at llvm.org]

================
@@ -155,6 +170,256 @@ struct ConvertVectorTransferRead final
 };
 } // end anonymous namespace
 
+//===----------------------------------------------------------------------===//
+// RewriteExtOfBitCast
+//===----------------------------------------------------------------------===//
+
+/// Create a vector of bit masks: `idx .. idx + step - 1` and broadcast it
+/// `numOccurrences` times.
+/// `step` must divide `bitwidth` evenly.
+/// Example: (4, 2, 3) -> [0xc, 0x3, 0xc, 0x3, 0xc, 0x3].
+static SmallVector<Attribute> computeExtOfBitCastMasks(MLIRContext *ctx,
+                                                       int64_t bitwidth,
+                                                       int64_t step,
+                                                       int64_t numOccurrences) {
+  assert(bitwidth % step == 0 && "step must divide bitwidth evenly");
+  IntegerType interimIntType = IntegerType::get(ctx, bitwidth);
+  SmallVector<Attribute> tmpMasks;
+  tmpMasks.reserve(bitwidth / step);
+  // Create a vector of bit masks: `idx .. idx + step - 1`.
+  for (int64_t idx = 0; idx < bitwidth; idx += step) {
+    LDBG("Mask bits " << idx << " .. " << idx + step - 1 << " out of "
+                      << bitwidth);
+    IntegerAttr mask = IntegerAttr::get(
+        interimIntType, llvm::APInt::getBitsSet(bitwidth, idx, idx + step));
+    tmpMasks.push_back(mask);
+  }
+  // Replicate the vector of bit masks to the desired size.
+  SmallVector<Attribute> masks;
+  masks.reserve(numOccurrences * tmpMasks.size());
+  for (int64_t idx = 0; idx < numOccurrences; ++idx)
+    llvm::append_range(masks, tmpMasks);
+  return masks;
+}
+
+/// Create a vector of bit shifts by `k * idx` and broadcast it `numOccurrences`
+/// times.
+/// `step` must divide `bitwidth` evenly.
+/// Example: (4, 2, 3) -> [0x0, 0x2, 0x0, 0x2, 0x0, 0x2].
+static SmallVector<Attribute>
+computeExtOfBitCastShifts(MLIRContext *ctx, int64_t bitwidth, int64_t step,
+                          int64_t numOccurrences) {
+  assert(bitwidth % step == 0 && "step must divide bitwidth evenly");
+  IntegerType interimIntType = IntegerType::get(ctx, bitwidth);
+  SmallVector<Attribute> tmpShifts;
+  for (int64_t idx = 0; idx < bitwidth; idx += step) {
+    IntegerAttr shift = IntegerAttr::get(interimIntType, idx);
+    tmpShifts.push_back(shift);
+  }
+  SmallVector<Attribute> shifts;
+  for (int64_t idx = 0; idx < numOccurrences; ++idx)
+    llvm::append_range(shifts, tmpShifts);
+  return shifts;
+}
+
+/// Create a vector of bit shuffles: `numOccurrences * idx` and broadcast it
+/// `bitwidth/step` times.
+/// `step` must divide `bitwidth` evenly.
+/// Example: (4, 2, 3) -> [0x0, 0x1, 0x0, 0x1, 0x0, 0x1].
----------------
qcolombet wrote:

I have a hard time reconciling the comment and the example.

For instance if we broadcast `bitwidth/step` times, we should only have 2 reps, not 3.

https://github.com/llvm/llvm-project/pull/65774