The AI CUDA Engineer 👷

import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(x: torch.Tensor) -> torch.Tensor:
    """
    Applies HardSigmoid activation to the input tensor.

    Args:
        x (torch.Tensor): Input tensor of any shape.

    Returns:
        torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
    """
    return F.hardsigmoid(x)


class Model(nn.Module):
    """
    Simple model that performs a HardSigmoid activation.
    """

    def __init__(self):
        super(Model, self).__init__()

    def forward(self, x: torch.Tensor, fn=module_fn) -> torch.Tensor:
        """
        Applies HardSigmoid activation to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of any shape.

        Returns:
            torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
        """
        return fn(x)


batch_size = 16
dim = 16384


def get_inputs():
    x = torch.randn(batch_size, dim)
    return [x]


def get_init_inputs():
    return []  # No special initialization inputs needed

#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <type_traits>

// This kernel evenly distributes workloads by precomputing the number of complete vectorized chunks
// and then assigning them uniformly among threads. It processes remaining tail elements separately.

template <typename scalar_t, int VEC_SIZE>
__global__ void hardsigmoid_kernel(const scalar_t* __restrict__ input,
                                   scalar_t* __restrict__ output,
                                   size_t numel) {
    // Each chunk holds VEC_SIZE elements
    size_t num_chunks = numel / VEC_SIZE;
    int tid = threadIdx.x + blockIdx.x * blockDim.x;
    int total_threads = blockDim.x * gridDim.x;

    constexpr scalar_t three = static_cast<scalar_t>(3.0);
    constexpr scalar_t sixth = static_cast<scalar_t>(1.0/6.0);

    // Select vector type based on precision: for float use float4, for double use double2
    using vec_t = typename std::conditional<
        std::is_same<scalar_t, float>::value,
        float4,
        double2
    >::type;

    // Evenly process full vectorized chunks
    for (size_t idx = tid; idx < num_chunks; idx += total_threads) {
        size_t base = idx * VEC_SIZE;
        vec_t chunk = *reinterpret_cast<const vec_t*>(&input[base]);
        scalar_t elems[VEC_SIZE];
        *reinterpret_cast<vec_t*>(elems) = chunk;

        #pragma unroll
        for (int i = 0; i < VEC_SIZE; i++) {
            scalar_t x = elems[i];
            x = (x + three) * sixth;  // computes (x + 3) / 6
            x = (x < static_cast<scalar_t>(0)) ? static_cast<scalar_t>(0) :
                (x > static_cast<scalar_t>(1) ? static_cast<scalar_t>(1) : x);
            elems[i] = x;
        }

        *reinterpret_cast<vec_t*>(&output[base]) = *reinterpret_cast<vec_t*>(elems);
    }

    // Process any remaining tail elements
    size_t tail_start = num_chunks * VEC_SIZE;
    for (size_t i = tail_start + tid; i < numel; i += total_threads) {
        scalar_t x = input[i];
        x = (x + three) * sixth;
        x = (x < static_cast<scalar_t>(0)) ? static_cast<scalar_t>(0) :
            (x > static_cast<scalar_t>(1) ? static_cast<scalar_t>(1) : x);
        output[i] = x;
    }
}

torch::Tensor forward(torch::Tensor input) {
    TORCH_CHECK(input.is_cuda(), "Input tensor must be on CUDA");
    auto output = torch::empty_like(input);
    size_t numel = input.numel();

    const int threads = 256;

    AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "evenly_distributed_hardsigmoid_cuda", ([&] {
        // Choose vector size depending on type: 4 for float (float4), 2 for double (double2)
        constexpr int vec_size = std::is_same<scalar_t, float>::value ? 4 : 2;
        size_t num_chunks = numel / vec_size;
        int blocks = (num_chunks + threads - 1) / threads;
        if (blocks == 0) {
            blocks = 1;
        }
        hardsigmoid_kernel<scalar_t, vec_size><<<blocks, threads>>>(
            input.data_ptr<scalar_t>(),
            output.data_ptr<scalar_t>(),
            numel);
    }));

    cudaError_t err = cudaGetLastError();
    TORCH_CHECK(err == cudaSuccess, "CUDA kernel failed: ", cudaGetErrorString(err));
    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("forward", &forward, "Evenly Distributed HardSigmoid activation forward (CUDA)");
}