mlx/mlx/backend/cuda/device/config.h
2025-06-16 18:49:32 -07:00

13 lines
388 B
C

// Copyright © 2025 Apple Inc.
// This file is used by both CUDA kernel code and host-only C++ code.
#pragma once
// The maximum dimensions of shape/strides passed as kernel parameters.
#define MAX_NDIM 10
// All existing NVIDIA hardware has a fixed 32 warp size. Though a built-in
// warpSize variable exists, using it would prevent compile-time optimizations.
#define WARP_SIZE 32