[TEST][EASY] Remove test_chain_reduce from triton's test_core.py (#8038)

As it is a duplicate with the reduce2d mode of `test_reduce_layouts` in gluon's `test_lowerings.py`

Author: Jokeren

python/test/unit/language/test_core.py

@@ -3079,82 +3079,6 @@ def kernel(
     assert compiled_kernel.asm["ttgir"].count('"tt.reduce"') == 1, "we shouldn't rematerialize tt.reduce"
 
 
-layouts = [
-    BlockedLayout([1, 4], [1, THREADS_PER_WARP], [4, 1], [1, 0], [1, 1], [1, 1], [0, 1]),
-    BlockedLayout([1, 4], [1, THREADS_PER_WARP], [2, 2], [1, 0], [1, 1], [1, 1], [0, 1]),
-    # [HIP] TO DO: some tests are flaky with the layout, so turn off them for now.
-    # BlockedLayout([1, 4], [1, THREADS_PER_WARP], [1, 4], [1, 0], [1, 1], [1, 1], [0, 1]),
-    BlockedLayout([1, 4], [THREADS_PER_WARP // 32, 32], [1, 4], [1, 0], [1, 1], [1, 1], [0, 1]),
-    BlockedLayout([1, 4], [8, THREADS_PER_WARP // 8], [2, 2], [0, 1], [1, 1], [1, 1], [0, 1])
-]
-
-
-@pytest.mark.parametrize("M, N", [[128, 128], [256, 128], [256, 256], [128, 256]])
-@pytest.mark.parametrize("src_layout", layouts)
-@pytest.mark.parametrize("op", ["sum", "max"])
-@pytest.mark.parametrize("first_axis", [0, 1])
-def test_chain_reduce(M, N, src_layout, op, device, first_axis, tmp_path: pathlib.Path):
-
-    op_str = ""
-    if op == "sum":
-        op_str = """
-        %13 = arith.addi %arg2, %arg3 : i32
-        tt.reduce.return %13 : i32"""
-    elif op == "max":
-        op_str = """
-        %13 = arith.cmpi "sgt", %arg2, %arg3 : i32
-        %14 = arith.select %13, %arg2, %arg3 : i32
-        tt.reduce.return %14 : i32"""
-    ir = f"""
-    #src = {src_layout}
-    module attributes {{"{GPU_DIALECT}.num-warps" = 4 : i32, "ttg.num-ctas" = 1 : i32, "ttg.threads-per-warp" = {THREADS_PER_WARP} : i32}} {{
-    tt.func public @sum_kernel_0d1d(%arg0: !tt.ptr<i32> {{tt.divisibility = 16 : i32}}, %arg1: !tt.ptr<i32> {{tt.divisibility = 16 : i32}}) {{
-        %cst = arith.constant dense<{N}> : tensor<{M}x1xi32, #src>
-        %0 = tt.make_range {{end = {M} : i32, start = 0 : i32}} : tensor<{M}xi32, #{GPU_DIALECT}.slice<{{dim = 1, parent = #src}}>>
-        %1 = tt.expand_dims %0 {{axis = 1 : i32}} : tensor<{M}xi32, #{GPU_DIALECT}.slice<{{dim = 1, parent = #src}}>> -> tensor<{M}x1xi32, #src>
-        %2 = arith.muli %1, %cst : tensor<{M}x1xi32, #src>
-        %3 = tt.make_range {{end = {N} : i32, start = 0 : i32}} : tensor<{N}xi32, #{GPU_DIALECT}.slice<{{dim = 0, parent = #src}}>>
-        %4 = tt.expand_dims %3 {{axis = 0 : i32}} : tensor<{N}xi32, #{GPU_DIALECT}.slice<{{dim = 0, parent = #src}}>> -> tensor<1x{N}xi32, #src>
-        %5 = tt.broadcast %2 : tensor<{M}x1xi32, #src> -> tensor<{M}x{N}xi32, #src>
-        %6 = tt.broadcast %4 : tensor<1x{N}xi32, #src> -> tensor<{M}x{N}xi32, #src>
-        %7 = arith.addi %5, %6 : tensor<{M}x{N}xi32, #src>
-        %8 = tt.splat %arg0 : !tt.ptr<i32> -> tensor<{M}x{N}x!tt.ptr<i32>, #src>
-        %9 = tt.addptr %8, %7 : tensor<{M}x{N}x!tt.ptr<i32>, #src>, tensor<{M}x{N}xi32, #src>
-        %10 = tt.load %9 : tensor<{M}x{N}x!tt.ptr<i32>, #src>
-        %11 = "tt.reduce"(%10) ({{
-        ^bb0(%arg2: i32, %arg3: i32):
-        {op_str}
-        }}) {{axis = {first_axis} : i32}} : (tensor<{M}x{N}xi32, #src>) -> tensor<{M if first_axis == 1 else N}xi32, #{GPU_DIALECT}.slice<{{dim = {first_axis}, parent = #src}}>>
-        %12 = "tt.reduce"(%11) ({{
-        ^bb0(%arg2: i32, %arg3: i32):
-        {op_str}
-        }}) {{axis = 0 : i32}} : (tensor<{M if first_axis == 1 else N}xi32, #{GPU_DIALECT}.slice<{{dim = {first_axis}, parent = #src}}>>) -> i32
-        tt.store %arg1, %12 : !tt.ptr<i32>
-        tt.return
-    }}
-    }}
-    """
-    temp_file = tmp_path / "test_chain_reduce.ttgir"
-    temp_file.write_text(ir)
-    kernel = triton.compile(str(temp_file))
-
-    rs = RandomState(17)
-    x = rs.randint(0, 4, (M, N)).astype('int32')
-
-    z = np.zeros((1, )).astype('int32')
-
-    x_tri = torch.tensor(x, device=device)
-    z_tri = torch.tensor(z, device=device)
-
-    pgm = kernel[(1, 1, 1)](x_tri, z_tri)
-    if op == "sum":
-        z_ref = np.sum(x)
-    elif op == "max":
-        z_ref = np.max(x)
-
-    np.testing.assert_allclose(z_ref, z_tri.cpu().numpy(), rtol=0.01, atol=1e-3)
-
-
 @triton.jit
 def _welford_combine(mean_1, m2_1, weight_1, mean_2, m2_2, weight_2):
     delta = mean_2 - mean_1