D2M: Add experimental get_noc_multicast_addr API that flips start/end coords on noc1 (#3855)

### Ticket
No mlir issue, part of overarching matmul objective.

### Problem description
tt-metal does not flip start and end coordinates in
`get_noc_multicast_addr`
This is required because the NoC requires the start coordinate to be
reached first.



### What's changed
This PR introduces an experimental API to reverse start and end
coordinates for multicast on noc1.
Note: This is not to do with flipping coordinate values in terms of the
noc layout. This is simply reversing the start & end coordinates.

### Checklist
- [ ] New/Existing tests provide coverage for changes

Author: jdesousa-TT

include/ttmlir/Dialect/TTKernel/IR/TTKernelOps.td

@@ -986,7 +986,17 @@ def TTKernel_MyYOp : TTKernel_Op<"my_y"> {
 def TTKernel_GetNocMulticastAddrOp : TTKernel_Op<"get_noc_multicast_addr"> {
   let summary = "GetNocMulticastAddr";
   let description = [{
-    GetNocMulticastAddr
+    Default tt-metal get_noc_multicast_addr
+  }];
+
+  let arguments = (ins IndexLike:$noc_x_start, IndexLike:$noc_y_start, IndexLike:$noc_x_end, IndexLike:$noc_y_end, AnyTypeOf<[I32, TTKernel_L1Addr, TTKernel_Semaphore]>:$addr, Optional<I8>:$noc);
+  let results = (outs TTKernel_NocAddr:$mcastNocAddr);
+}
+
+def TTKernel_ExperimentalGetNocMulticastAddrOp : TTKernel_Op<"experimental::get_noc_multicast_addr"> {
+  let summary = "Experimental GetNocMulticastAddr";
+  let description = [{
+    Default tt-metal get_noc_multicast_addr, but flips mcast start and end coordinates on NOC1.
   }];
 
   let arguments = (ins IndexLike:$noc_x_start, IndexLike:$noc_y_start, IndexLike:$noc_x_end, IndexLike:$noc_y_end, AnyTypeOf<[I32, TTKernel_L1Addr, TTKernel_Semaphore]>:$addr, Optional<I8>:$noc);

include/ttmlir/Target/TTKernel/LLKs/CMakeLists.txt

@@ -5,6 +5,7 @@ include(GenerateRawStringHeader)
 set(LLK_HEADERS
     ${CMAKE_SOURCE_DIR}/include/ttmlir/Target/TTKernel/LLKs/experimental_tilize_llks.h
     ${CMAKE_SOURCE_DIR}/include/ttmlir/Target/TTKernel/LLKs/experimental_untilize_llks.h
+    ${CMAKE_SOURCE_DIR}/include/ttmlir/Target/TTKernel/LLKs/experimental_dataflow_api.h
 )
 
 # Set the output directory for generated headers

include/ttmlir/Target/TTKernel/LLKs/experimental_dataflow_api.h

@@ -0,0 +1,34 @@
+// SPDX-FileCopyrightText: (c) 2025 Tenstorrent AI ULC
+//
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef TTMLIR_TARGET_TTKERNEL_LLKS_EXPERIMENTAL_DATAFLOW_API_H
+#define TTMLIR_TARGET_TTKERNEL_LLKS_EXPERIMENTAL_DATAFLOW_API_H
+
+namespace experimental {
+
+FORCE_INLINE
+std::uint64_t
+get_noc_multicast_addr(std::uint32_t noc_x_start, std::uint32_t noc_y_start,
+                       std::uint32_t noc_x_end, std::uint32_t noc_y_end,
+                       std::uint32_t addr, uint8_t noc = noc_index) {
+  /*
+      Get an encoding which contains tensix core and address you want to
+      read from/write to via the noc
+  */
+  if (noc) {
+    // noc 1
+    return NOC_MULTICAST_ADDR(
+        DYNAMIC_NOC_X(noc, noc_x_end), DYNAMIC_NOC_Y(noc, noc_y_end),
+        DYNAMIC_NOC_X(noc, noc_x_start), DYNAMIC_NOC_Y(noc, noc_y_start), addr);
+  } else {
+    // noc 0
+    return NOC_MULTICAST_ADDR(
+        DYNAMIC_NOC_X(noc, noc_x_start), DYNAMIC_NOC_Y(noc, noc_y_start),
+        DYNAMIC_NOC_X(noc, noc_x_end), DYNAMIC_NOC_Y(noc, noc_y_end), addr);
+  }
+}
+
+} // namespace experimental
+
+#endif

lib/Conversion/TTIRToTTKernel/TTIRToTTKernel.cpp

@@ -638,9 +638,10 @@ class TTIRDMARewriter : public OpConversionPattern<ttir::DMAOp> {
             op.getLoc(), op.getMcastShape()[0], op.getMcastShape()[1]);
         auto numDests = rewriter.create<arith::IndexCastOp>(
             op.getLoc(), rewriter.getI32Type(), numDestsIdx);
-        auto mcastAddr = rewriter.create<ttkernel::GetNocMulticastAddrOp>(
-            op.getLoc(), virtX, virtY, mcastEndX, mcastEndY, dstL1Start,
-            nullptr);
+        auto mcastAddr =
+            rewriter.create<ttkernel::ExperimentalGetNocMulticastAddrOp>(
+                op.getLoc(), virtX, virtY, mcastEndX, mcastEndY, dstL1Start,
+                nullptr);
         if (adaptor.getSrc() == adaptor.getDst()) {
           // If src and dst refer to the same memref, we do not loopback mcast
           // Dests are one less because the sender core is not included
@@ -989,9 +990,10 @@ class TTIRSemaphoreUpdateRewriter : public OpConversionPattern<ConcreteOp> {
           op.getLoc(), op.getMcastShape()[0], op.getMcastShape()[1]);
       Value numDests = rewriter.create<arith::IndexCastOp>(
           op.getLoc(), rewriter.getI32Type(), numDestsIdx);
-      auto mcastAddr = rewriter.create<ttkernel::GetNocMulticastAddrOp>(
-          op.getLoc(), virtX, virtY, mcastEndX, mcastEndY, semaphoreAddr,
-          nullptr);
+      auto mcastAddr =
+          rewriter.create<ttkernel::ExperimentalGetNocMulticastAddrOp>(
+              op.getLoc(), virtX, virtY, mcastEndX, mcastEndY, semaphoreAddr,
+              nullptr);
 
       auto semaphorePtr =
           rewriter.create<ttkernel::CastToL1PtrOp>(op.getLoc(), semaphoreAddr);

lib/Conversion/TTKernelToEmitC/TTKernelToEmitC.cpp

@@ -659,6 +659,8 @@ class ConvertTTKernelToEmitCPass
         TTKernelToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteTileOp>,
         TTKernelToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteBarrierOp>,
         TTKernelToEmitCOpaqueRewriter<ttkernel::GetNocMulticastAddrOp>,
+        TTKernelToEmitCOpaqueRewriter<
+            ttkernel::ExperimentalGetNocMulticastAddrOp>,
         TTKernelToEmitCOpaqueRewriter<
             ttkernel::NocAsyncWriteMulticastOnePacketOp>,
         TTKernelToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteMulticastOp>,

lib/Target/TTKernel/TTKernelToCpp.cpp

@@ -6,6 +6,7 @@
 
 #include "ttmlir/Dialect/TTKernel/IR/TTKernelOpsTypes.h"
 
+#include "ttmlir/Target/TTKernel/LLKs/experimental_dataflow_api_generated.h"
 #include "ttmlir/Target/TTKernel/LLKs/experimental_tilize_llks_generated.h"
 #include "ttmlir/Target/TTKernel/LLKs/experimental_untilize_llks_generated.h"
 
@@ -43,6 +44,7 @@ class ScopedModuleHelper {
 
       builder->create<emitc::IncludeOp>(loc, "dataflow_api.h",
                                         /*isStandard=*/false);
+      emitExperimentalLLKs();
     }
     if (threadType == ThreadType::Compute) {
       builder->create<emitc::IncludeOp>(loc, "llk_defs.h",
@@ -165,6 +167,13 @@ void dprint(Arg &&arg, ArgV&&... argv) {
                     experimental_untilize_llks_generated_len);
       builder->create<emitc::VerbatimOp>(loc, experimentalUntilizeLLKs);
     }
+
+    if (hasCall("experimental::get_noc_multicast_addr")) {
+      auto experimentalDataflowLLKs =
+          StringRef(experimental_dataflow_api_generated,
+                    experimental_dataflow_api_generated_len);
+      builder->create<emitc::VerbatimOp>(loc, experimentalDataflowLLKs);
+    }
   }
 
   bool hasCall(StringRef name) {

test/ttmlir/Conversion/TTIRToTTKernel/dma_lowering.mlir

@@ -42,7 +42,7 @@ func.func @test_local_to_remote_multicast_regular(%arg0: memref<2x2x!ttcore.tile
   %c4 = arith.constant 4 : index
   // CHECK: ttkernel.get_read_ptr
   // CHECK: ttkernel.get_write_ptr
-  // CHECK: ttkernel.get_noc_multicast_addr
+  // CHECK: ttkernel.experimental::get_noc_multicast_addr
   // CHECK: ttkernel.noc_async_write_multicast
   %0 = ttir.dma %arg0[%c0, %c0], %arg0[%c0, %c0] core[%c1, %c2] mcast[%c3, %c4] : (memref<2x2x!ttcore.tile<32x32, f32>, #l1_>, memref<2x2x!ttcore.tile<32x32, f32>, #l1_>) -> !ttir.mem_tx
   ttir.dma_wait %0
@@ -59,7 +59,7 @@ func.func @test_local_to_remote_multicast_loopback(%arg0: memref<2x2x!ttcore.til
   %c4 = arith.constant 4 : index
   // CHECK: ttkernel.get_read_ptr
   // CHECK: ttkernel.get_write_ptr
-  // CHECK: ttkernel.get_noc_multicast_addr
+  // CHECK: ttkernel.experimental::get_noc_multicast_addr
   // CHECK: ttkernel.noc_async_write_multicast_loopback_src
   %0 = ttir.dma %arg0[%c0, %c0], %arg1[%c0, %c0] core[%c1, %c2] mcast[%c3, %c4] : (memref<2x2x!ttcore.tile<32x32, f32>, #l1_>, memref<2x2x!ttcore.tile<32x32, f32>, #l1_>) -> !ttir.mem_tx
   ttir.dma_wait %0