add copy_tile_init() needed before eltwise max operand copies (#1495)

tenstorrent · Dec 4, 2024 · d837ac7 · d837ac7
1 parent 7f6046e
commit d837ac7
Show file tree

Hide file tree

Showing 2 changed files with 15 additions and 1 deletion.
diff --git a/lib/Conversion/TTIRToTTMetal/TTIRToTTMetal.cpp b/lib/Conversion/TTIRToTTMetal/TTIRToTTMetal.cpp
@@ -1007,11 +1007,12 @@ class TTIRToTTMetalDispatchRewriter : public OpRewritePattern<ttir::GenericOp> {
     builder.create<ttkernel::TileRegsAcquireOp>(location);
     {
       // copy inCB0[inCB0TileIndex] and inCB1[inCB1TileIndex] to DST:
+      builder.create<ttkernel::CopyTileInitOp>(location);
       builder.create<ttkernel::CopyTileOp>(location, inCB0, inCB0TileIndex,
                                            dstLhsTileIndex);
       builder.create<ttkernel::CopyTileOp>(location, inCB1, inCB1TileIndex,
                                            dstRhsTileIndex);
-      // SFPU ooperates on DST tiles:
+      // SFPU operates on DST tiles:
       builder.create<TTKernelTilesOp>(location, dstLhsTileIndex,
                                       dstRhsTileIndex);
     }

diff --git a/test/ttmlir/Silicon/TTMetal/simple_max.mlir b/test/ttmlir/Silicon/TTMetal/simple_max.mlir
@@ -0,0 +1,13 @@
+// RUN: ttmlir-opt --ttir-to-ttmetal-backend-pipeline="system-desc-path=%system_desc_path%" %s > %t.mlir
+// RUN: FileCheck %s --input-file=%t.mlir
+// RUN: ttmlir-translate --ttmetal-to-flatbuffer %t.mlir > %t.ttm
+
+#any_device = #tt.operand_constraint<dram|l1|scalar|tile|any_device|any_device_tile>
+
+func.func @maximum(%arg0: tensor<64x128xf32>, %arg1: tensor<64x128xf32>) -> tensor<64x128xf32> {
+  // CHECK: %[[C:.*]] = "ttmetal.alloc"[[C:.*]]
+  %0 = tensor.empty() : tensor<64x128xf32>
+  // CHECK: %[[C:.*]] = "ttmetal.dispatch"[[C:.*]]
+  %1 = "ttir.maximum"(%arg0, %arg1, %0) <{operandSegmentSizes = array<i32: 2, 1>, operand_constraints = [#any_device, #any_device, #any_device]}> : (tensor<64x128xf32>, tensor<64x128xf32>, tensor<64x128xf32>) -> tensor<64x128xf32>
+  return %1 : tensor<64x128xf32>
+}