Add NoC Mcast Ops in D2M (#1100)

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

@@ -625,6 +625,68 @@ def TTKernel_CastToL1PtrOp : TTKernel_Op<"reinterpret_cast<volatile tt_l1_ptr ui
     let results = (outs TTKernel_L1AddrPtr:$l1_ptr);
 }
 
+//===----------------------------------------------------------------------===//
+// TTKernel Multicast NoC operations
+//===----------------------------------------------------------------------===//
+
+def TTKernel_GetNocMulticastAddrOp : TTKernel_Op<"get_noc_multicast_addr"> {
+  let summary = "GetNocMulticastAddr";
+  let description = [{
+    GetNocMulticastAddr
+  }];
+
+  let arguments = (ins I32:$noc_x_start, I32:$noc_y_start, I32:$noc_x_end, I32:$noc_y_end, I32:$addr, Optional<I8>:$noc);
+  let results = (outs TTKernel_NocAddr:$mcastNocAddr);
+}
+
+def TTKernel_NocAsyncWriteMulticastOnePacketOp : TTKernel_Op<"noc_async_write_multicast_one_packet"> {
+  let summary = "NocAsyncWriteMulticastOnePacket";
+  let description = [{
+    NocAsyncWriteMulticastOnePacket
+    this issues only a single packet with size <= NOC_MAX_BURST_SIZE (ie maximum packet size)
+  }];
+
+  let arguments = (ins I32:$srcLocalL1Addr, TTKernel_NocAddr:$dstNocAddrMulticast, I32:$size, I32:$num_dests, OptionalAttr<BoolAttr>:$linked, OptionalAttr<BoolAttr>:$multicast_path_reserve, Optional<I8>:$noc);
+}
+
+def TTKernel_NocAsyncWriteMulticastOp : TTKernel_Op<"noc_async_write_multicast"> {
+  let summary = "NocAsyncWriteMulticast";
+  let description = [{
+    Initiates an asynchronous write from a source address in L1 memory on the
+    Tensix core executing this function call to a rectangular destination grid.
+    The destinations are specified using a uint64_t encoding referencing an
+    on-chip grid of nodes located at NOC coordinate range
+    (x_start,y_start,x_end,y_end) and a local address created using
+    *get_noc_multicast_addr* function. Also, *see noc_async_write_barrier*.
+
+    The destination nodes can only be a set of Tensix cores + L1 memory address.
+    The destination nodes must form a rectangular grid. The destination L1
+    memory address must be the same on all destination nodes.
+
+    With this API, the multicast sender cannot be part of the multicast
+    destinations. If the multicast sender has to be in the multicast
+    destinations (i.e. must perform a local L1 write), the other API variant
+    *noc_async_write_multicast_loopback_src* can be used.
+
+    Note: The number of destinations needs to be non-zero. Besides that,
+    there is no restriction on the number of destinations, i.e. the
+    multicast destinations can span the full chip. However, as mentioned
+    previously, the multicast source cannot be part of the destinations. So, the
+    maximum number of destinations is 119.
+  }];
+
+  let arguments = (ins I32:$srcLocalL1Addr, TTKernel_NocAddr:$dstNocAddrMulticast, I32:$size, I32:$num_dests, OptionalAttr<BoolAttr>:$linked, OptionalAttr<BoolAttr>:$multicast_path_reserve, Optional<I8>:$noc);
+}
+
+def TTKernel_NocAsyncWriteMulticastLoopbackSrcOp : TTKernel_Op<"noc_async_write_multicast_loopback_src"> {
+  let summary = "NocAsyncWriteMulticastLoopbackSrc";
+  let description = [{
+    NocAsyncWriteMulticastLoopbackSrc
+  }];
+
+  let arguments = (ins I32:$srcLocalL1Addr, TTKernel_NocAddr:$dstNocAddrMulticast, I32:$size, I32:$num_dests, OptionalAttr<BoolAttr>:$linked, OptionalAttr<BoolAttr>:$multicast_path_reserve, Optional<I8>:$noc);
+}
+
 //===----------------------------------------------------------------------===//
 // TTKernel Misc operations
 //===----------------------------------------------------------------------===//

lib/Conversion/TTKernelToEmitC/TTKernelToEmitC.cpp

@@ -247,7 +247,8 @@ class TTMetalToEmitCOpaqueRewriter : public OpConversionPattern<SourceOp> {
     } else if constexpr (std::is_same_v<SourceOp, ttkernel::GetArgValOp>) {
       SmallVector<Attribute, 1> template_args;
 
-      template_args.push_back(emitc::OpaqueAttr::get(op.getContext(), "uint32_t"));
+      template_args.push_back(
+          emitc::OpaqueAttr::get(op.getContext(), "uint32_t"));
 
       return ArrayAttr::get(op.getContext(), template_args);
     }
@@ -369,6 +370,12 @@ class ConvertTTKernelToEmitCPass
           TTMetalToEmitCOpaqueRewriter<ttkernel::NocAsyncReadBarrierOp>,
           TTMetalToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteOp>,
           TTMetalToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteBarrierOp>,
+          TTMetalToEmitCOpaqueRewriter<ttkernel::GetNocMulticastAddrOp>,
+          TTMetalToEmitCOpaqueRewriter<
+              ttkernel::NocAsyncWriteMulticastOnePacketOp>,
+          TTMetalToEmitCOpaqueRewriter<ttkernel::NocAsyncWriteMulticastOp>,
+          TTMetalToEmitCOpaqueRewriter<
+              ttkernel::NocAsyncWriteMulticastLoopbackSrcOp>,
           TTMetalToEmitCOpaqueRewriter<ttkernel::UnaryOpInitCommonOp>,
           TTMetalToEmitCOpaqueRewriter<ttkernel::CopyTileOp>,
           TTMetalToEmitCOpaqueRewriter<ttkernel::ExpTileInitOp>,