[NVPTX] Add TMA Bulk Copy Intrinsics

durga4github · durga4github · commit 359413acce58 · 2025-05-06T17:25:50.000+05:30
This patch adds a new variant of TMA Bulk Copy
intrinsics introduced in sm100+. This variant
has an additional byte_mask to select the bytes
for the copy operation.

Signed-off-by: Durgadoss R &lt;durgadossr@nvidia.com&gt;
diff --git a/llvm/docs/NVPTXUsage.rst b/llvm/docs/NVPTXUsage.rst
@@ -616,6 +616,7 @@ Syntax:
 .. code-block:: llvm
 
   declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i64 %ch, i1 %flag_ch)
+  declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global.bytemask(..., i32 %size, i16 %mask, i64 %ch, i1 %flag_ch)
 
 Overview:
 """""""""
@@ -624,7 +625,10 @@ The '``@llvm.nvvm.cp.async.bulk.shared.cta.to.global``' intrinsic
 corresponds to the ``cp.async.bulk.global.shared::cta.*`` set of PTX
 instructions. These instructions initiate an asynchronous copy from
 shared::cta to global memory. The 32-bit operand ``%size`` specifies
-the amount of memory to be copied and it must be a multiple of 16.
+the amount of memory to be copied (in bytes) and it must be a multiple
+of 16. For the ``.bytemask`` variant, the 16-bit wide mask operand
+specifies whether the i-th byte of each 16-byte wide chunk of source
+data is copied to the destination.
 
 * The last argument to these intrinsics is a boolean flag
   indicating support for cache_hint. This flag argument must
diff --git a/llvm/include/llvm/IR/IntrinsicsNVVM.td b/llvm/include/llvm/IR/IntrinsicsNVVM.td
@@ -5323,6 +5323,20 @@ def int_nvvm_cp_async_bulk_shared_cta_to_global
        NoCapture<ArgIndex<0>>, NoCapture<ArgIndex<1>>,
        ImmArg<ArgIndex<4>>]>;
 
+// From Shared CTA to Global memory with bytemask
+def int_nvvm_cp_async_bulk_shared_cta_to_global_bytemask
+  : DefaultAttrsIntrinsic<[],
+      [llvm_global_ptr_ty, // dst_gmem_ptr
+       llvm_shared_ptr_ty, // src_smem_ptr
+       llvm_i32_ty,        // copy_size
+       llvm_i16_ty,        // byte_mask
+       llvm_i64_ty,        // cache_hint
+       llvm_i1_ty],        // Flag for cache_hint
+      [IntrConvergent, IntrArgMemOnly,
+       WriteOnly<ArgIndex<0>>, ReadOnly<ArgIndex<1>>,
+       NoCapture<ArgIndex<0>>, NoCapture<ArgIndex<1>>,
+       ImmArg<ArgIndex<5>>]>;
+
 // Intrinsics for Bulk Copy Prefetch L2
 def int_nvvm_cp_async_bulk_prefetch_L2
   : DefaultAttrsIntrinsic<[],
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.cpp
@@ -2720,28 +2720,46 @@ void NVPTXDAGToDAGISel::SelectCpAsyncBulkTensorReduceCommon(SDNode *N,
   ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops));
 }
 
-void NVPTXDAGToDAGISel::SelectCpAsyncBulkS2G(SDNode *N) {
+void NVPTXDAGToDAGISel::SelectCpAsyncBulkS2GCommon(SDNode *N, bool HasMask) {
   // We have {Chain, Intrinsic-ID} followed by the actual intrisic args:
-  // dst, src, size, cache_hint, cache_hint_flag
+  // dst, src, size, mask, cache_hint, cache_hint_flag
   // NumOperands = {Chain, IID} + {Actual intrinsic args}
-  //             = {2}          + {5}
+  //             = {2}          + {6}
   size_t NumOps = N->getNumOperands();
   bool IsCacheHint = N->getConstantOperandVal(NumOps - 1) == 1;
-  size_t NumArgs = IsCacheHint ? 4 : 3; // src, dst, size, cache_hint
+  size_t CacheHintIdx = NumOps - 2;
 
   SDLoc DL(N);
-  SmallVector<SDValue, 8> Ops(N->ops().slice(2, NumArgs));
-  Ops.push_back(N->getOperand(0)); // Chain operand
+  SDValue Offset, Base;
+  SelectADDR(N->getOperand(3), Base, Offset); // src
 
-  bool IsShared32 =
-      CurDAG->getDataLayout().getPointerSizeInBits(ADDRESS_SPACE_SHARED) == 32;
-  unsigned Opcode;
+  SmallVector<SDValue, 8> Ops;
+  // BaseArgs: {dst, src, size}
+  Ops.push_back(N->getOperand(2)); // dst
+  Ops.push_back(Base);             // src
+  Ops.push_back(Offset);           // src
+  Ops.push_back(N->getOperand(4)); // size
+
+  // Push Mask operand, if available
+  if (HasMask)
+    Ops.push_back(N->getOperand(CacheHintIdx - 1));
+
+  // Push CacheHint operand, if available
   if (IsCacheHint)
-    Opcode = IsShared32 ? NVPTX::CP_ASYNC_BULK_S2G_SHARED32_CH
-                        : NVPTX::CP_ASYNC_BULK_S2G_CH;
-  else
-    Opcode = IsShared32 ? NVPTX::CP_ASYNC_BULK_S2G_SHARED32
-                        : NVPTX::CP_ASYNC_BULK_S2G;
+    Ops.push_back(N->getOperand(CacheHintIdx));
+
+  // Finally, the chain operand
+  Ops.push_back(N->getOperand(0));
+
+  unsigned Opcode = [&]() {
+    if (HasMask && IsCacheHint)
+      return NVPTX::CP_ASYNC_BULK_S2G_BM_CH;
+    if (HasMask)
+      return NVPTX::CP_ASYNC_BULK_S2G_BM;
+    if (IsCacheHint)
+      return NVPTX::CP_ASYNC_BULK_S2G_CH;
+    return NVPTX::CP_ASYNC_BULK_S2G;
+  }();
   ReplaceNode(N, CurDAG->getMachineNode(Opcode, DL, N->getVTList(), Ops));
 }
 
@@ -2928,7 +2946,10 @@ bool NVPTXDAGToDAGISel::tryIntrinsicVoid(SDNode *N) {
     SelectCpAsyncBulkG2S(N);
     return true;
   case Intrinsic::nvvm_cp_async_bulk_shared_cta_to_global:
-    SelectCpAsyncBulkS2G(N);
+    SelectCpAsyncBulkS2GCommon(N);
+    return true;
+  case Intrinsic::nvvm_cp_async_bulk_shared_cta_to_global_bytemask:
+    SelectCpAsyncBulkS2GCommon(N, /*HasMask=*/true);
     return true;
   case Intrinsic::nvvm_cp_async_bulk_prefetch_L2:
     SelectCpAsyncBulkPrefetchL2(N);
diff --git a/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h b/llvm/lib/Target/NVPTX/NVPTXISelDAGToDAG.h
@@ -93,7 +93,7 @@ class LLVM_LIBRARY_VISIBILITY NVPTXDAGToDAGISel : public SelectionDAGISel {
   void SelectV2I64toI128(SDNode *N);
   void SelectI128toV2I64(SDNode *N);
   void SelectCpAsyncBulkG2S(SDNode *N);
-  void SelectCpAsyncBulkS2G(SDNode *N);
+  void SelectCpAsyncBulkS2GCommon(SDNode *N, bool HasMask = false);
   void SelectCpAsyncBulkPrefetchL2(SDNode *N);
   void SelectCpAsyncBulkTensorG2SCommon(SDNode *N, bool IsIm2Col = false);
   void SelectCpAsyncBulkTensorS2GCommon(SDNode *N, bool IsIm2Col = false);
diff --git a/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td b/llvm/lib/Target/NVPTX/NVPTXIntrinsics.td
@@ -516,6 +516,9 @@ class CpAsyncBulkStr<bit mc, bit ch> {
   string S2G = "cp.async.bulk.global.shared::cta.bulk_group"
                # !if(ch, ".L2::cache_hint", "");
 
+  // Shared to Global memory with bytemask
+  string S2G_BM = S2G # ".cp_mask";
+
   // Global to Shared cluster memory
   string G2S = "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes"
                # !if(mc, ".multicast::cluster", "")
@@ -525,18 +528,26 @@ class CpAsyncBulkStr<bit mc, bit ch> {
   string C2C = "cp.async.bulk.shared::cluster.shared::cta.mbarrier::complete_tx::bytes";
 }
 
-multiclass CP_ASYNC_BULK_S2G<NVPTXRegClass rc> {
-  def NAME: NVPTXInst<(outs),
-            (ins Int64Regs:$dst, rc:$src, Int32Regs:$size),
+def CP_ASYNC_BULK_S2G : NVPTXInst<(outs),
+            (ins Int64Regs:$dst, ADDR:$src, Int32Regs:$size),
             !strconcat(CpAsyncBulkStr<0, 0>.S2G, " [$dst], [$src], $size;"), []>,
             Requires<[hasPTX<80>, hasSM<90>]>;
-  def NAME # _CH: NVPTXInst<(outs),
-                  (ins Int64Regs:$dst, rc:$src, Int32Regs:$size, Int64Regs:$ch),
-                  !strconcat(CpAsyncBulkStr<0, 1>.S2G, " [$dst], [$src], $size, $ch;"), []>,
-                  Requires<[hasPTX<80>, hasSM<90>]>;
-}
-defm CP_ASYNC_BULK_S2G : CP_ASYNC_BULK_S2G<Int64Regs>;
-defm CP_ASYNC_BULK_S2G_SHARED32 : CP_ASYNC_BULK_S2G<Int32Regs>;
+
+def CP_ASYNC_BULK_S2G_CH : NVPTXInst<(outs),
+            (ins Int64Regs:$dst, ADDR:$src, Int32Regs:$size, Int64Regs:$ch),
+            !strconcat(CpAsyncBulkStr<0, 1>.S2G, " [$dst], [$src], $size, $ch;"), []>,
+            Requires<[hasPTX<80>, hasSM<90>]>;
+
+// Variants with bytemask
+def CP_ASYNC_BULK_S2G_BM : NVPTXInst<(outs),
+                (ins Int64Regs:$dst, ADDR:$src, Int32Regs:$size, Int16Regs:$mask),
+                !strconcat(CpAsyncBulkStr<0, 0>.S2G_BM, " [$dst], [$src], $size, $mask;"), []>,
+                Requires<[hasPTX<86>, hasSM<100>]>;
+
+def CP_ASYNC_BULK_S2G_BM_CH : NVPTXInst<(outs),
+                   (ins Int64Regs:$dst, ADDR:$src, Int32Regs:$size, Int16Regs:$mask, Int64Regs:$ch),
+                   !strconcat(CpAsyncBulkStr<0, 1>.S2G_BM, " [$dst], [$src], $size, $ch, $mask;"), []>,
+                   Requires<[hasPTX<86>, hasSM<100>]>;
 
 multiclass CP_ASYNC_BULK_G2S<NVPTXRegClass rc> {
   def NAME: NVPTXInst<(outs),
diff --git a/llvm/test/CodeGen/NVPTX/cp-async-bulk-s2g-sm100.ll b/llvm/test/CodeGen/NVPTX/cp-async-bulk-s2g-sm100.ll
@@ -0,0 +1,46 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_100 -mattr=+ptx86| FileCheck --check-prefixes=CHECK-PTX64 %s
+; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_100 -mattr=+ptx86 --nvptx-short-ptr| FileCheck --check-prefixes=CHECK-PTX-SHARED32 %s
+; RUN: %if ptxas-12.8 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_100 -mattr=+ptx86| %ptxas-verify -arch=sm_100 %}
+; RUN: %if ptxas-12.8 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_100 -mattr=+ptx86 --nvptx-short-ptr| %ptxas-verify -arch=sm_100 %}
+
+target triple = "nvptx64-nvidia-cuda"
+
+declare void @llvm.nvvm.cp.async.bulk.shared.cta.to.global.bytemask(ptr addrspace(1), ptr addrspace(3), i32, i16, i64, i1)
+
+define void @cp_async_bulk_s2g_bytemask(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i16 %mask, i64 %ch) {
+; CHECK-PTX64-LABEL: cp_async_bulk_s2g_bytemask(
+; CHECK-PTX64:       {
+; CHECK-PTX64-NEXT:    .reg .b16 %rs<2>;
+; CHECK-PTX64-NEXT:    .reg .b32 %r<2>;
+; CHECK-PTX64-NEXT:    .reg .b64 %rd<4>;
+; CHECK-PTX64-EMPTY:
+; CHECK-PTX64-NEXT:  // %bb.0:
+; CHECK-PTX64-NEXT:    ld.param.u64 %rd1, [cp_async_bulk_s2g_bytemask_param_0];
+; CHECK-PTX64-NEXT:    ld.param.u64 %rd2, [cp_async_bulk_s2g_bytemask_param_1];
+; CHECK-PTX64-NEXT:    ld.param.u32 %r1, [cp_async_bulk_s2g_bytemask_param_2];
+; CHECK-PTX64-NEXT:    ld.param.u16 %rs1, [cp_async_bulk_s2g_bytemask_param_3];
+; CHECK-PTX64-NEXT:    ld.param.u64 %rd3, [cp_async_bulk_s2g_bytemask_param_4];
+; CHECK-PTX64-NEXT:    cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint.cp_mask [%rd1], [%rd2], %r1, %rd3, %rs1;
+; CHECK-PTX64-NEXT:    cp.async.bulk.global.shared::cta.bulk_group.cp_mask [%rd1], [%rd2], %r1, %rs1;
+; CHECK-PTX64-NEXT:    ret;
+;
+; CHECK-PTX-SHARED32-LABEL: cp_async_bulk_s2g_bytemask(
+; CHECK-PTX-SHARED32:       {
+; CHECK-PTX-SHARED32-NEXT:    .reg .b16 %rs<2>;
+; CHECK-PTX-SHARED32-NEXT:    .reg .b32 %r<3>;
+; CHECK-PTX-SHARED32-NEXT:    .reg .b64 %rd<3>;
+; CHECK-PTX-SHARED32-EMPTY:
+; CHECK-PTX-SHARED32-NEXT:  // %bb.0:
+; CHECK-PTX-SHARED32-NEXT:    ld.param.u64 %rd1, [cp_async_bulk_s2g_bytemask_param_0];
+; CHECK-PTX-SHARED32-NEXT:    ld.param.u32 %r1, [cp_async_bulk_s2g_bytemask_param_1];
+; CHECK-PTX-SHARED32-NEXT:    ld.param.u32 %r2, [cp_async_bulk_s2g_bytemask_param_2];
+; CHECK-PTX-SHARED32-NEXT:    ld.param.u16 %rs1, [cp_async_bulk_s2g_bytemask_param_3];
+; CHECK-PTX-SHARED32-NEXT:    ld.param.u64 %rd2, [cp_async_bulk_s2g_bytemask_param_4];
+; CHECK-PTX-SHARED32-NEXT:    cp.async.bulk.global.shared::cta.bulk_group.L2::cache_hint.cp_mask [%rd1], [%r1], %r2, %rd2, %rs1;
+; CHECK-PTX-SHARED32-NEXT:    cp.async.bulk.global.shared::cta.bulk_group.cp_mask [%rd1], [%r1], %r2, %rs1;
+; CHECK-PTX-SHARED32-NEXT:    ret;
+  tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.global.bytemask(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i16 %mask, i64 %ch, i1 1)
+  tail call void @llvm.nvvm.cp.async.bulk.shared.cta.to.global.bytemask(ptr addrspace(1) %dst, ptr addrspace(3) %src, i32 %size, i16 %mask, i64 0, i1 0)
+  ret void
+}
