From ea46e277d26ab87907153962fec4266e8800c8bd Mon Sep 17 00:00:00 2001 From: Nandor Licker Date: Fri, 17 Apr 2026 03:54:24 +0300 Subject: [PATCH] Add `absf` and `floor` to `cute.math` (#3156) The ops are already exposed by the underlying dialect. --- python/CuTeDSL/cutlass/cute/math.py | 94 ++++++++++++++++++++++++ test/examples/CuTeDSL/test_math.py | 109 ++++++++++++++++++++++++++++ 2 files changed, 203 insertions(+) create mode 100644 test/examples/CuTeDSL/test_math.py diff --git a/python/CuTeDSL/cutlass/cute/math.py b/python/CuTeDSL/cutlass/cute/math.py index 8dab6b981..89acc70ab 100644 --- a/python/CuTeDSL/cutlass/cute/math.py +++ b/python/CuTeDSL/cutlass/cute/math.py @@ -49,6 +49,33 @@ def _math_op(func: Callable, fastmath: bool, *args, **kwargs): @dsl_user_op +def absf( + a: Union[TensorSSA, Numeric], fastmath: bool = False, *, loc=None, ip=None +) -> Union[TensorSSA, Numeric]: + """Compute element-wise absolute value of the input tensor. + + :param a: Input tensor + :type a: Union[TensorSSA, Numeric] + :param fastmath: Enable fast math optimizations, defaults to False + :type fastmath: bool, optional + :param loc: Source location information, defaults to None + :type loc: Optional[Location] + :param ip: Insertion point for IR generation, defaults to None + :type ip: Optional[InsertionPoint] + :return: Tensor containing the absolute value of each element in input tensor + :rtype: Union[TensorSSA, Numeric] + + Example: + + .. code-block:: + + x = cute.make_rmem_tensor(layout) # Create tensor + y = x.load() # Load values + z = absf(y) # Compute absolute value + """ + return _math_op(math.absf, fastmath, a, loc=loc, ip=ip) + + def acos( a: Union[TensorSSA, Numeric], fastmath: bool = False, *, loc=None, ip=None ) -> Union[TensorSSA, Numeric]: @@ -166,6 +193,43 @@ def atan2( return _math_op(math.atan2, fastmath, a, b, loc=loc, ip=ip) +@dsl_user_op +def copysign( + a: Union[TensorSSA, Numeric], + b: Union[TensorSSA, Numeric], + fastmath: bool = False, + *, + loc=None, + ip=None, +) -> Union[TensorSSA, Numeric]: + """Compute element-wise copysign of two tensors. + + Returns a value with the magnitude of ``a`` and the sign of ``b``. + + :param a: Input tensor providing magnitude + :type a: Union[TensorSSA, Numeric] + :param b: Input tensor providing sign + :type b: Union[TensorSSA, Numeric] + :param fastmath: Enable fast math optimizations, defaults to False + :type fastmath: bool, optional + :param loc: Source location information, defaults to None + :type loc: Optional[Location] + :param ip: Insertion point for IR generation, defaults to None + :type ip: Optional[InsertionPoint] + :return: Tensor where each element has the magnitude of ``a`` and the sign of ``b`` + :rtype: Union[TensorSSA, Numeric] + + Example: + + .. code-block:: + + mag = cute.make_rmem_tensor(ptr1, layout).load() # magnitudes + sgn = cute.make_rmem_tensor(ptr2, layout).load() # signs + result = copysign(mag, sgn) # Combine magnitude and sign + """ + return _math_op(math.copysign, fastmath, a, b, loc=loc, ip=ip) + + @dsl_user_op def cos( a: Union[TensorSSA, Numeric], fastmath: bool = False, *, loc=None, ip=None @@ -282,6 +346,33 @@ def exp2( @dsl_user_op +def floor( + a: Union[TensorSSA, Numeric], fastmath: bool = False, *, loc=None, ip=None +) -> Union[TensorSSA, Numeric]: + """Compute element-wise floor of the input tensor. + + :param a: Input tensor + :type a: Union[TensorSSA, Numeric] + :param fastmath: Enable fast math optimizations, defaults to False + :type fastmath: bool, optional + :param loc: Source location information, defaults to None + :type loc: Optional[Location] + :param ip: Insertion point for IR generation, defaults to None + :type ip: Optional[InsertionPoint] + :return: Tensor containing the largest integer less than or equal to each element in input tensor + :rtype: Union[TensorSSA, Numeric] + + Example: + + .. code-block:: + + x = cute.make_rmem_tensor(layout) # Create tensor + y = x.load() # Load values + z = floor(y) # Compute floor + """ + return _math_op(math.floor, fastmath, a, loc=loc, ip=ip) + + def log( a: Union[TensorSSA, Numeric], fastmath: bool = False, *, loc=None, ip=None ) -> Union[TensorSSA, Numeric]: @@ -508,14 +599,17 @@ def tanh( __all__ = [ + "absf", "acos", "asin", "atan", "atan2", + "copysign", "cos", "erf", "exp", "exp2", + "floor", "log", "log10", "log2", diff --git a/test/examples/CuTeDSL/test_math.py b/test/examples/CuTeDSL/test_math.py new file mode 100644 index 000000000..9c63900f0 --- /dev/null +++ b/test/examples/CuTeDSL/test_math.py @@ -0,0 +1,109 @@ +# Copyright (c) 2025 - 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: BSD-3-Clause + +# Redistribution and use in source and binary forms, with or without +# modification, are permitted provided that the following conditions are met: + +# 1. Redistributions of source code must retain the above copyright notice, this +# list of conditions and the following disclaimer. + +# 2. Redistributions in binary form must reproduce the above copyright notice, +# this list of conditions and the following disclaimer in the documentation +# and/or other materials provided with the distribution. + +# 3. Neither the name of the copyright holder nor the names of its +# contributors may be used to endorse or promote products derived from +# this software without specific prior written permission. + +# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +import pytest +import torch + +import cutlass +import cutlass.cute as cute +from cutlass.cute.runtime import from_dlpack + + +@cute.kernel +def _unary_ops_kernel( + absf_inp: cute.Tensor, absf_out: cute.Tensor, + floor_inp: cute.Tensor, floor_out: cute.Tensor, +): + tidx, _, _ = cute.arch.thread_idx() + absf_out[tidx] = cute.math.absf(absf_inp[tidx]) + floor_out[tidx] = cute.math.floor(floor_inp[tidx]) + + +@cute.jit +def _unary_ops_host( + absf_inp: cute.Tensor, absf_out: cute.Tensor, + floor_inp: cute.Tensor, floor_out: cute.Tensor, +): + _unary_ops_kernel(absf_inp, absf_out, floor_inp, floor_out).launch( + grid=[1, 1, 1], block=[absf_inp.shape[0], 1, 1] + ) + + +def test_unary_ops(): + absf_inp = torch.tensor([-3.5, 2.0, 0.0], device="cuda", dtype=torch.float32) + absf_expected = torch.tensor([3.5, 2.0, 0.0], device="cuda", dtype=torch.float32) + absf_out = torch.zeros(3, device="cuda", dtype=torch.float32) + floor_inp = torch.tensor([3.7, -2.3, 5.0], device="cuda", dtype=torch.float32) + floor_expected = torch.tensor([3.0, -3.0, 5.0], device="cuda", dtype=torch.float32) + floor_out = torch.zeros(3, device="cuda", dtype=torch.float32) + absf_inp_cute = from_dlpack(absf_inp) + absf_out_cute = from_dlpack(absf_out) + floor_inp_cute = from_dlpack(floor_inp) + floor_out_cute = from_dlpack(floor_out) + + args = (absf_inp_cute, absf_out_cute, floor_inp_cute, floor_out_cute) + + cute.compile(_unary_ops_host, *args)(*args) + torch.cuda.synchronize() + + assert torch.equal(absf_out, absf_expected) + assert torch.equal(floor_out, floor_expected) + + +@cute.kernel +def _binary_ops_kernel( + mag_inp: cute.Tensor, sign_inp: cute.Tensor, out: cute.Tensor, +): + tidx, _, _ = cute.arch.thread_idx() + out[tidx] = cute.math.copysign(mag_inp[tidx], sign_inp[tidx]) + + +@cute.jit +def _binary_ops_host( + mag_inp: cute.Tensor, sign_inp: cute.Tensor, out: cute.Tensor, +): + _binary_ops_kernel(mag_inp, sign_inp, out).launch( + grid=[1, 1, 1], block=[mag_inp.shape[0], 1, 1] + ) + + +def test_binary_ops(): + mag_inp = torch.tensor([3.5, -2.0, 0.0, 1.0], device="cuda", dtype=torch.float32) + sign_inp = torch.tensor([-1.0, 1.0, -1.0, 1.0], device="cuda", dtype=torch.float32) + expected = torch.tensor([-3.5, 2.0, -0.0, 1.0], device="cuda", dtype=torch.float32) + out = torch.zeros(4, device="cuda", dtype=torch.float32) + mag_inp_cute = from_dlpack(mag_inp) + sign_inp_cute = from_dlpack(sign_inp) + out_cute = from_dlpack(out) + + args = (mag_inp_cute, sign_inp_cute, out_cute) + + cute.compile(_binary_ops_host, *args)(*args) + torch.cuda.synchronize() + + assert torch.equal(out, expected)