[Language] Support tile operator T.cumsum

src/op/reduce.cc

@@ -240,5 +240,55 @@ TIR_REGISTER_TL_OP(ReduceOp, reduce)
     .set_attr<TCallEffectKind>("TCallEffectKind",
                                Integer(CallEffectKind::kOpaque));
 
+CumSumOp::CumSumOp(Array<PrimExpr> args, BufferMap vmap) {
+  /*
+    CumSum arguments:
+      src: input buffer
+      dst: output buffer
+      dim: dimension to cumsum
+      reverse: whether to cumsum in reverse order
+   */
+  CHECK_EQ(args.size(), 4);
+  src = vmap[GetVarFromAccessPtr(args[0])];
+  dst = vmap[GetVarFromAccessPtr(args[1])];
+  dim = args[2].as<IntImm>().value()->value;
+  reverse = args[3].as<Bool>().value();
+  CHECK_LT(dim, static_cast<int>(src->shape.size()));
+}
+
+Stmt CumSumOp::Lower(const LowerArgs &T, arith::Analyzer *analyzer) const {
+  if (this->src.scope() == "local.fragment" &&
+      this->dst.scope() == "local.fragment") {
+    LOG(FATAL) << "CumSum for fragment not implemented, please raise an issue "
+                  "if you need this feature.";
+  } else if (this->src.scope() == "shared.dyn" ||
+             this->src.scope() == "shared") {
+    ICHECK(this->dst.scope() == "shared.dyn" || this->dst.scope() == "shared");
+    std::stringstream ss;
+    auto threads = T.thread_bounds->extent - T.thread_bounds->min;
+    ss << "tl::CumSum2D<" << threads << ", " << dim << ", "
+       << (reverse ? "true" : "false") << ">::run";
+    Array<PrimExpr> args = {StringImm(ss.str()), src.access_ptr(1),
+                            dst.access_ptr(3)};
+    for (int i = 0; i < src->shape.size(); i++) {
+      args.push_back(src->shape[i]);
+    }
+    return Evaluate(Call(dst->dtype, builtin::call_extern(), args));
+  } else {
+    ICHECK(false) << "Cannot lower cumsum for " << this->src.scope() << " and "
+                  << this->dst.scope();
+  }
+
+  return Stmt();
+}
+
+LayoutMap CumSumOp::InferLayout(const LayoutInferArgs &T, InferLevel level) {
+  return {};
+}
+
+TIR_REGISTER_TL_OP(CumSumOp, cumsum)
+    .set_num_inputs(4)
+    .set_attr<TCallEffectKind>("TCallEffectKind",
+                               Integer(CallEffectKind::kOpaque));
 } // namespace tl
 } // namespace tvm

src/op/reduce.h

@@ -41,6 +41,19 @@ class ReduceOp : public Operator {
   std::string MakeCodegenReducer() const;
 };
 
+class CumSumOp : public Operator {
+public:
+  CumSumOp(Array<PrimExpr> args, BufferMap vmap);
+  Stmt Lower(const LowerArgs &T, arith::Analyzer *analyzer) const final;
+  LayoutMap InferLayout(const LayoutInferArgs &T, InferLevel level) final;
+  static const Op &Get();
+
+private:
+  tir::Buffer src, dst;
+  int dim;
+  bool reverse;
+};
+
 } // namespace tl
 } // namespace tvm
 

src/tl_templates/cuda/reduce.h

@@ -67,4 +67,83 @@ struct AllReduce {
   }
 };
 
+template <int threads, int Axis = 0, bool reverse = false> struct CumSum2D {
+  static_assert(threads == 1024 or threads == 512 or threads == 256 or
+                threads == 128 or threads == 64 or threads == 32);
+  template <typename T, int SEG = 32>
+  static TL_DEVICE T run(const T *__restrict__ src, T *__restrict__ dst, int H,
+                         int W) {
+
+    constexpr int TILE_H = threads / SEG;
+    constexpr unsigned MASK = 0xffffffff;
+    const int num_blocks = (H + TILE_H - 1) / TILE_H;
+    const int tid = threadIdx.x;
+    const int lane = tid % 32;
+    const int row = tid / 32;
+
+    for (int b = 0; b < num_blocks; ++b) {
+      const int gRow = b * TILE_H + row;
+      if (gRow >= H)
+        return;
+
+      T carry = (T)0;
+
+      if (reverse) {
+        // Start from the last segment for reverse mode
+        for (int seg = (W + SEG - 1) / SEG - 1; seg >= 0; --seg) {
+          const int col = seg * SEG + lane;
+
+          const int real_row = Axis == 1 ? gRow : col;
+          const int real_col = Axis == 1 ? col : gRow;
+
+          T val = (col < W) ? src[real_row * W + real_col] : (T)0;
+
+#pragma unroll
+          for (int off = 1; off < SEG; off <<= 1) {
+            T n = (T)__shfl_down_sync(MASK, val, off);
+            if (lane < SEG - off)
+              val += n;
+          }
+
+          val += carry;
+
+          if (real_col < W)
+            dst[real_row * W + real_col] = val;
+
+          T segSum = (T)__shfl_sync(MASK, val, (T)0);
+          if (lane == 0)
+            carry = segSum;
+          carry = (T)__shfl_sync(MASK, carry, (T)0);
+        }
+      } else {
+        for (int seg = 0; seg * SEG < W; ++seg) {
+          const int col = seg * SEG + lane;
+
+          const int real_row = Axis == 1 ? gRow : col;
+          const int real_col = Axis == 1 ? col : gRow;
+
+          T val = (col < W) ? src[real_row * W + real_col] : (T)0;
+
+#pragma unroll
+          for (int off = 1; off < SEG; off <<= 1) {
+            T n = (T)__shfl_up_sync(MASK, val, off);
+            if (lane >= off)
+              val += n;
+          }
+
+          val += carry;
+
+          if (real_col < W)
+            dst[real_row * W + real_col] = val;
+
+          T segSum = (T)__shfl_sync(MASK, val, SEG - 1);
+          if (lane == SEG - 1)
+            carry = segSum;
+          carry = (T)__shfl_sync(MASK, carry, SEG - 1);
+        }
+      }
+    }
+  }
+};
+
 } // namespace tl

testing/python/language/test_tilelang_language_cumsum.py

@@ -0,0 +1,96 @@
+# Copyright (c) Tile-AI Corporation.
+# Licensed under the MIT License.
+
+from tilelang import tvm as tvm
+import tilelang.testing
+import tilelang as tl
+import torch
+
+
+def cumsum_smem_test(M, N, block_M, block_N, dim=0, reverse=False, dtype="float16"):
+    import tilelang.language as T
+
+    @T.prim_func
+    def cumsum(
+            A: T.Tensor((M, N), dtype),
+            B: T.Tensor((M, N), dtype),
+    ):
+        # Initialize Kernel Context
+        with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=256) as (bx, by):
+            A_shared = T.alloc_shared((block_M, block_N), dtype)
+
+            T.copy(A[by * block_M, bx * block_N], A_shared)
+            T.cumsum(src=A_shared, dim=dim, reverse=reverse)
+            T.copy(A_shared, B[by * block_M, bx * block_N])
+
+    return cumsum
+
+
+def cumsum_fragment_test(M, N, block_M, block_N, dim=0, reverse=False, dtype="float16"):
+    import tilelang.language as T
+
+    @T.prim_func
+    def cumsum(
+            A: T.Tensor((M, N), dtype),
+            B: T.Tensor((M, N), dtype),
+    ):
+        # Initialize Kernel Context
+        with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=256) as (bx, by):
+            A_shared = T.alloc_shared((block_M, block_N), dtype)
+            A_fragment = T.alloc_fragment((block_M, block_N), dtype)
+
+            T.copy(A[by * block_M, bx * block_N], A_shared)
+            T.copy(A_shared, A_fragment)
+            T.cumsum(src=A_fragment, dim=dim, reverse=reverse)
+            T.copy(A_fragment, B[by * block_M, bx * block_N])
+
+    return cumsum
+
+
+def run_cumsum(M, N, block_M, block_N, dim=0, reverse=False, dtype="float16", scope="smem"):
+    if scope == "smem":
+        program = cumsum_smem_test(M, N, block_M, block_N, dim, reverse, dtype)
+    elif scope == "fragment":
+        program = cumsum_fragment_test(M, N, block_M, block_N, dim, reverse, dtype)
+    jit_kernel = tl.compile(program, out_idx=-1)
+    profiler = jit_kernel.get_profiler(tensor_supply_type=tl.TensorSupplyType.One)
+
+    def ref_program(A):
+        ref_b = torch.empty_like(A)
+        for i in range(M // block_M):
+            for j in range(N // block_N):
+                ref_b[i * block_M:(i + 1) * block_M,
+                      j * block_N:(j + 1) * block_N] = A[i * block_M:(i + 1) * block_M, j *
+                                                         block_N:(j + 1) * block_N].cumsum(dim=dim)
+                if reverse:
+                    ref_b[i * block_M:(i + 1) * block_M, j * block_N:(j + 1) *
+                          block_N] = ref_b[i * block_M:(i + 1) * block_M,
+                                           j * block_N:(j + 1) * block_N].flip(dims=[dim])
+        return ref_b
+
+    profiler.assert_allclose(ref_program)
+
+
+def test_cumsum_smem():
+    # Test different sizes
+    run_cumsum(1024, 1024, 128, 128)
+    run_cumsum(1024, 1024, 128, 128, dim=1)
+    run_cumsum(1024, 1024, 128, 128, dim=1, reverse=True)
+
+    # Test different dtypes
+    run_cumsum(256, 256, 128, 128, dtype="float32")
+    run_cumsum(256, 256, 128, 128, dtype="float16")
+
+
+def test_cumsum_fragment():
+    run_cumsum(1024, 1024, 128, 128, scope="fragment")
+    run_cumsum(1024, 1024, 128, 128, dim=1, scope="fragment")
+    run_cumsum(1024, 1024, 128, 128, dim=1, reverse=True, scope="fragment")
+
+    # Test different dtypes
+    run_cumsum(256, 256, 128, 128, dtype="float32", scope="fragment")
+    run_cumsum(256, 256, 128, 128, dtype="float16", scope="fragment")
+
+
+if __name__ == "__main__":
+    tilelang.testing.main()

testing/python/language/test_tilelang_language_reduce_sum.py

@@ -0,0 +1,57 @@
+# Copyright (c) Tile-AI Corporation.
+# Licensed under the MIT License.
+
+from tilelang import tvm as tvm
+import tilelang.testing
+import tilelang as tl
+
+tilelang.disable_cache()
+
+
+def reduce_sum_test(M, N, dtype="float16"):
+    import tilelang.language as T
+
+    @T.prim_func
+    def main(
+            A: T.Tensor((M, N), dtype),
+            B: T.Tensor((M,), dtype),
+    ):
+        with T.Kernel(1) as _:
+            A_local = T.alloc_fragment((M, N), dtype)
+            B_local = T.alloc_fragment((M,), dtype)
+
+            # Copy input to local
+            T.copy(A, A_local)
+            # Perform reduce_sum operation
+            T.reduce_sum(A_local, B_local, dim=0)
+            # Copy result back
+            T.copy(B_local, B)
+
+    return main
+
+
+def run_reduce_sum(M, N, dtype="float16"):
+    program = reduce_sum_test(M, N, dtype)
+    jit_kernel = tl.compile(program, out_idx=-1)
+    print(jit_kernel.get_kernel_source())
+    profiler = jit_kernel.get_profiler()
+
+    def ref_program(A):
+        return A.sum(dim=0)
+
+    profiler.assert_allclose(ref_program, atol=1e-2, rtol=1e-2)
+
+
+def test_reduce_sum():
+    # Test different sizes
+    run_reduce_sum(256, 256)
+    run_reduce_sum(512, 128)
+    run_reduce_sum(128, 512)
+
+    # Test different dtypes
+    run_reduce_sum(256, 256, "float32")
+    run_reduce_sum(256, 256, "float16")
+
+
+if __name__ == "__main__":
+    tilelang.testing.main()

tilelang/language/__init__.py

@@ -50,6 +50,7 @@
     reduce_sum,  # noqa: F401
     reduce_abssum,  # noqa: F401
     reduce_absmax,  # noqa: F401
+    cumsum,  # noqa: F401
 )
 from .print import print  # noqa: F401
 from .customize import (

tilelang/language/reduce.py

@@ -3,6 +3,8 @@
 """The language interface for tl programs."""
 
 from tvm import tir
+from typing import Optional
+from tilelang.language import copy, macro, alloc_shared
 
 
 def reduce(buffer: tir.Buffer, out: tir.Buffer, reduce_type: str, dim: int, clear: bool):
@@ -106,3 +108,33 @@ def reduce_absmax(buffer: tir.Buffer, out: tir.Buffer, dim: int):
         tir.Call: Handle to the reduction operation
     """
     return reduce(buffer, out, "absmax", dim, True)
+
+
+@macro
+def cumsum_fragment(src: tir.Buffer, dst: tir.Buffer, dim: int, reverse: bool) -> tir.PrimExpr:
+    cumsum_smem = alloc_shared(src.shape, src.dtype, "shared.dyn")
+    copy(src, cumsum_smem)
+    tir.call_intrin(
+        "handle",
+        tir.op.Op.get("tl.cumsum"),
+        cumsum_smem.access_ptr("r"),
+        cumsum_smem.access_ptr("w"),
+        dim,
+        reverse,
+    )
+    copy(cumsum_smem, dst)
+
+
+def cumsum(src: tir.Buffer, dst: Optional[tir.Buffer] = None, dim: int = 0, reverse: bool = False):
+    if dst is None:
+        dst = src
+    if src.scope() == "local.fragment":
+        return cumsum_fragment(src, dst, dim, reverse)
+    return tir.call_intrin(
+        "handle",
+        tir.op.Op.get("tl.cumsum"),
+        src.access_ptr("r"),
+        dst.access_ptr("w"),
+        dim,
+        reverse,
+    )

tilelang/utils/tensor.py

@@ -295,6 +295,8 @@ def torch_assert_close(
             f"({max_mismatched_ratio * 100:.2f}% allowed, but get {num_mismatched / total_elements * 100:.2f}%)."
             f"{mismatch_info}"
             f"\nGreatest absolute difference: {diff.max().item()}, "
-            f"Greatest relative difference: {(diff / (torch.abs(tensor_b) + 1e-12)).max().item()}.")
+            f"Greatest relative difference: {(diff / (torch.abs(tensor_b) + 1e-12)).max().item()}"
+            f"\nLHS: {tensor_a}"
+            f"\nRHS: {tensor_b}")
     else:
         return True