[Mlir-commits] [mlir] [MLIR][NVVM] Fix undef in cp.async.bulk.tensor.reduce Op (PR #157423)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Mon Sep 8 03:50:25 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
@llvm/pr-subscribers-mlir-llvm
Author: Srinivasa Ravi (Wolfram70)
<details>
<summary>Changes</summary>
This change moves the LLVMIR lowering code of the NVVM dialect cp.async.bulk.tensor.reduce Op to `NVVMDialect.cpp` and fixes the usage of `undef` in the lowering since it is now deprecated.
The tests are updated accordingly.
---
Patch is 37.81 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/157423.diff
3 Files Affected:
- (modified) mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td (+6-24)
- (modified) mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp (+60-29)
- (modified) mlir/test/Target/LLVMIR/nvvm/tma_store_reduce.mlir (+64-64)
``````````diff
diff --git a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
index 9d93b4efe7a5b..f4fa8b1c7bfae 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
@@ -2795,35 +2795,17 @@ def NVVM_CpAsyncBulkTensorReduceOp :
}];
let extraClassDeclaration = [{
- static llvm::Intrinsic::ID getIntrinsicID(int tensorDims,
- NVVM::TMAReduxKind kind,
- bool isIm2Col);
+ static mlir::NVVM::IDArgPair
+ getIntrinsicIDAndArgs(Operation &op,
+ LLVM::ModuleTranslation &mt, llvm::IRBuilderBase& builder);
}];
let hasVerifier = 1;
string llvmBuilder = [{
- // Arguments to the intrinsic:
- // shared_mem_ptr, tmaDesc, tensorDims
- // cache_hint(if applicable) and flag(boolean)
- llvm::SmallVector<llvm::Value *> translatedOperands;
- translatedOperands.push_back($srcMem);
- translatedOperands.push_back($tmaDescriptor);
-
- for (auto v : op.getCoordinates())
- translatedOperands.push_back(moduleTranslation.lookupValue(v));
-
- llvm::LLVMContext &ctx = moduleTranslation.getLLVMContext();
- auto *i64Undef = llvm::UndefValue::get(llvm::IntegerType::get(ctx, 64));
-
- bool isCacheHint = op.getL2CacheHint() ? true : false;
- translatedOperands.push_back(isCacheHint ? $l2CacheHint : i64Undef);
- translatedOperands.push_back(builder.getInt1(isCacheHint));
-
- auto intId = NVVM::CpAsyncBulkTensorReduceOp::getIntrinsicID(
- op.getCoordinates().size(), $redKind,
- (op.getMode() == NVVM::TMAStoreMode::IM2COL));
- createIntrinsicCall(builder, intId, translatedOperands);
+ auto [id, args] = NVVM::CpAsyncBulkTensorReduceOp::getIntrinsicIDAndArgs(
+ *op, moduleTranslation, builder);
+ createIntrinsicCall(builder, id, args);
}];
}
diff --git a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
index 376e3c3e1fcbe..6cfe36ebcf35a 100644
--- a/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
+++ b/mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp
@@ -1641,46 +1641,77 @@ CpAsyncBulkTensorSharedCTAToGlobalOp::getIntrinsicIDAndArgs(
is_im2col ? CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, im2col) \
: CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, dim, tile)
-#define GET_CP_ASYNC_BULK_TENSOR_ID(op, dims, is_im2col) \
- [&]() -> auto { \
- switch (dims) { \
- case 1: \
- return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 1, tile); \
- case 2: \
- return CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 2, tile); \
- case 3: \
- return CP_ASYNC_BULK_TENSOR_REDUCE(op, 3, is_im2col); \
- case 4: \
- return CP_ASYNC_BULK_TENSOR_REDUCE(op, 4, is_im2col); \
- case 5: \
- return CP_ASYNC_BULK_TENSOR_REDUCE(op, 5, is_im2col); \
- default: \
- llvm_unreachable("Invalid TensorDim in CpAsyncBulkTensorReduceOp."); \
- } \
- }()
+#define GET_CP_ASYNC_BULK_TENSOR_ID(iid, op, dims, is_im2col) \
+ switch (dims) { \
+ case 1: \
+ iid = CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 1, tile); \
+ break; \
+ case 2: \
+ iid = CP_ASYNC_BULK_TENSOR_REDUCE_MODE(op, 2, tile); \
+ break; \
+ case 3: \
+ iid = CP_ASYNC_BULK_TENSOR_REDUCE(op, 3, is_im2col); \
+ break; \
+ case 4: \
+ iid = CP_ASYNC_BULK_TENSOR_REDUCE(op, 4, is_im2col); \
+ break; \
+ case 5: \
+ iid = CP_ASYNC_BULK_TENSOR_REDUCE(op, 5, is_im2col); \
+ break; \
+ default: \
+ llvm_unreachable("Invalid TensorDim in CpAsyncBulkTensorReduceOp."); \
+ break; \
+ } \
+ break;
+
+NVVM::IDArgPair CpAsyncBulkTensorReduceOp::getIntrinsicIDAndArgs(
+ Operation &op, LLVM::ModuleTranslation &mt, llvm::IRBuilderBase &builder) {
+ auto thisOp = cast<NVVM::CpAsyncBulkTensorReduceOp>(op);
+ llvm::LLVMContext &ctx = mt.getLLVMContext();
+
+ llvm::SmallVector<llvm::Value *> args;
+
+ // Arguments to the intrinsic:
+ // shared_mem_ptr, tmaDesc, tensorDims
+ // cache_hint(if applicable) and flag(boolean)
+ args.push_back(mt.lookupValue(thisOp.getSrcMem()));
+ args.push_back(mt.lookupValue(thisOp.getTmaDescriptor()));
+
+ for (auto v : thisOp.getCoordinates())
+ args.push_back(mt.lookupValue(v));
+
+ mlir::Value cacheHint = thisOp.getL2CacheHint();
+ const bool hasCacheHint = static_cast<bool>(cacheHint);
+ llvm::Value *i64Unused =
+ llvm::ConstantInt::get(llvm::Type::getInt64Ty(ctx), 0);
+ args.push_back(hasCacheHint ? mt.lookupValue(cacheHint) : i64Unused);
+ args.push_back(builder.getInt1(hasCacheHint));
+
+ llvm::Intrinsic::ID iid;
+ int tensorDims = thisOp.getCoordinates().size();
+ bool isIm2Col = thisOp.getMode() == NVVM::TMAStoreMode::IM2COL;
-llvm::Intrinsic::ID CpAsyncBulkTensorReduceOp::getIntrinsicID(
- int tensorDims, NVVM::TMAReduxKind kind, bool isIm2Col) {
using RedTy = NVVM::TMAReduxKind;
- switch (kind) {
+ switch (thisOp.getRedKind()) {
case RedTy::ADD:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_add, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_add, tensorDims, isIm2Col);
case RedTy::MIN:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_min, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_min, tensorDims, isIm2Col);
case RedTy::MAX:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_max, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_max, tensorDims, isIm2Col);
case RedTy::INC:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_inc, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_inc, tensorDims, isIm2Col);
case RedTy::DEC:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_dec, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_dec, tensorDims, isIm2Col);
case RedTy::AND:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_and, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_and, tensorDims, isIm2Col);
case RedTy::OR:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_or, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_or, tensorDims, isIm2Col);
case RedTy::XOR:
- return GET_CP_ASYNC_BULK_TENSOR_ID(reduce_xor, tensorDims, isIm2Col);
+ GET_CP_ASYNC_BULK_TENSOR_ID(iid, reduce_xor, tensorDims, isIm2Col);
}
- llvm_unreachable("Invalid Reduction Op for CpAsyncBulkTensorReduceOp");
+
+ return {iid, std::move(args)};
}
#define _none
diff --git a/mlir/test/Target/LLVMIR/nvvm/tma_store_reduce.mlir b/mlir/test/Target/LLVMIR/nvvm/tma_store_reduce.mlir
index 6e0b48489e8b0..2231f1dabd504 100644
--- a/mlir/test/Target/LLVMIR/nvvm/tma_store_reduce.mlir
+++ b/mlir/test/Target/LLVMIR/nvvm/tma_store_reduce.mlir
@@ -19,14 +19,14 @@ llvm.func @tma_store_reduce_1d(%src : !llvm.ptr<3>, %tma_desc : !llvm.ptr, %d0 :
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<or>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<xor>} : !llvm.ptr, !llvm.ptr<3>
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 undef, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.1d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i64 0, i1 false)
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0] {redKind = #nvvm.tma_redux_kind<add>, mode = #nvvm.tma_store_mode<tile>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0] {redKind = #nvvm.tma_redux_kind<min>, mode = #nvvm.tma_store_mode<tile>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0] {redKind = #nvvm.tma_redux_kind<max>, mode = #nvvm.tma_store_mode<tile>} : !llvm.ptr, !llvm.ptr<3>
@@ -59,14 +59,14 @@ llvm.func @tma_store_reduce_2d(%src : !llvm.ptr<3>, %tma_desc : !llvm.ptr, %d0 :
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<or>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<xor>} : !llvm.ptr, !llvm.ptr<3>
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 undef, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.2d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i64 0, i1 false)
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1] {redKind = #nvvm.tma_redux_kind<add>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1] {redKind = #nvvm.tma_redux_kind<min>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1] {redKind = #nvvm.tma_redux_kind<max>} : !llvm.ptr, !llvm.ptr<3>
@@ -99,14 +99,14 @@ llvm.func @tma_store_reduce_3d_tile(%src : !llvm.ptr<3>, %tma_desc : !llvm.ptr,
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<or>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<xor>} : !llvm.ptr, !llvm.ptr<3>
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
+ // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 0, i1 false)
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] {redKind = #nvvm.tma_redux_kind<add>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] {redKind = #nvvm.tma_redux_kind<min>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] {redKind = #nvvm.tma_redux_kind<max>} : !llvm.ptr, !llvm.ptr<3>
@@ -137,14 +137,14 @@ llvm.func @tma_store_reduce_3d_im2col(%src : !llvm.ptr<3>, %tma_desc : !llvm.ptr
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<or>, mode = #nvvm.tma_store_mode<im2col>} : !llvm.ptr, !llvm.ptr<3>
nvvm.cp.async.bulk.tensor.reduce %tma_desc, %src, box[%d0, %d1, %d2] l2_cache_hint = %ch {redKind = #nvvm.tma_redux_kind<xor>, mode = #nvvm.tma_store_mode<im2col>} : !llvm.ptr, !llvm.ptr<3>
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.3d(ptr addrspace(3) %[[SRC]], ptr %[[DST]], i32 %[[D0]], i32 %[[D1]], i32 %[[D2]], i64 undef, i1 false)
- // CHECK: call void @llvm.nvvm.cp...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/157423
More information about the Mlir-commits
mailing list