The AI CUDA Engineer 👷

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


def module_fn(
    x: torch.Tensor,
    stride: int,
    padding: int,
    conv_weight: torch.Tensor,
    conv_bias: torch.Tensor,
) -> torch.Tensor:
    """
    Applies 3D convolution followed by Mish and Tanh activations.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_channels, D, H, W)
        stride (int): Stride of the convolution
        padding (int): Padding of the convolution
        conv_weight (torch.Tensor): Convolution weight tensor of shape
            (out_channels, in_channels, kernel_size, kernel_size, kernel_size)
        conv_bias (torch.Tensor): Bias tensor for convolution of shape (out_channels)

    Returns:
        torch.Tensor: Output tensor after applying convolution, Mish and Tanh activations
    """
    x = F.conv3d(x, conv_weight, bias=conv_bias, stride=stride, padding=padding)
    x = F.mish(x)
    x = torch.tanh(x)
    return x


class Model(nn.Module):
    """
    Model that performs a 3D convolution, applies Mish activation, and then applies Tanh activation.
    """

    def __init__(self, in_channels, out_channels, kernel_size, stride, padding):
        super(Model, self).__init__()
        conv = nn.Conv3d(
            in_channels, out_channels, kernel_size, stride=stride, padding=padding
        )
        self.conv_weight = nn.Parameter(conv.weight)
        self.conv_bias = nn.Parameter(
            conv.bias
            + torch.randn(
                conv.bias.shape, device=conv.bias.device, dtype=conv.bias.dtype
            )
            * 0.02
        )

    def forward(self, x, stride, padding, fn=module_fn):
        return fn(x, stride, padding, self.conv_weight, self.conv_bias)


batch_size = 16
in_channels = 3
out_channels = 16
D, H, W = 16, 32, 32
kernel_size = 3
stride = 1
padding = 0


def get_inputs():
    return [torch.randn(batch_size, in_channels, D, H, W), stride, padding]


def get_init_inputs():
    return [in_channels, out_channels, kernel_size, stride, padding]

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

// Device function: fused Mish and Tanh activation
// Computes: tanh(x * tanh(log(1 + exp(x))))
__device__ __forceinline__ float fused_mish_tanh_activation(float x) {
    float softplus = logf(1.0f + expf(x));
    float mish = x * tanhf(softplus);
    return tanhf(mish);
}

// Kernel that uses __ldg() for read-only global memory loads and vectorized (128-bit) accesses
// Processes most of the data in chunks of 4 floats and handles remaining elements separately.
__global__ void aligned_ldg_mish_tanh_kernel(
    float* __restrict__ output,
    const float* __restrict__ input,
    const int total_elements,
    const int total_vec4  // Number of 4-float groups
) {
    int global_threads = blockDim.x * gridDim.x;
    
    // Process vectorized loads and stores (128-bit aligned, i.e., 4 floats at a time)
    for (int i = threadIdx.x + blockIdx.x * blockDim.x; i < total_vec4; i += global_threads) {
        // Use __ldg() to load a float4 from aligned global memory
        float4 in_val = __ldg(reinterpret_cast<const float4*>(input) + i);
        float4 out_val;
        out_val.x = fused_mish_tanh_activation(in_val.x);
        out_val.y = fused_mish_tanh_activation(in_val.y);
        out_val.z = fused_mish_tanh_activation(in_val.z);
        out_val.w = fused_mish_tanh_activation(in_val.w);
        
        // Store the result back to global memory (assumed to be 16-byte aligned)
        reinterpret_cast<float4*>(output)[i] = out_val;
    }

    // Process remaining elements that don't fit into a vectorized load
    int vec4_total = total_vec4 * 4;
    for (int i = vec4_total + threadIdx.x + blockIdx.x * blockDim.x; i < total_elements; i += global_threads) {
        float in_val = __ldg(input + i);
        output[i] = fused_mish_tanh_activation(in_val);
    }
}

// Forward function: Perform 3D convolution then apply the fused activation via custom CUDA kernel
// This implementation leverages __ldg() for optimized read-only loads and 128-bit aligned vectorized memory accesses
torch::Tensor module_fn_forward(
    torch::Tensor x,
    int64_t stride,
    int64_t padding,
    torch::Tensor conv_weight,
    torch::Tensor conv_bias
) {
    TORCH_CHECK(x.is_cuda(), "Input tensor x must be a CUDA tensor");
    TORCH_CHECK(conv_weight.is_cuda(), "Convolution weight must be a CUDA tensor");
    TORCH_CHECK(conv_bias.is_cuda(), "Convolution bias must be a CUDA tensor");

    // Perform 3D convolution using PyTorch's highly optimized conv3d
    auto x_conv = at::conv3d(
        x, 
        conv_weight, 
        conv_bias, 
        {stride, stride, stride},
        {padding, padding, padding}
    );

    // Prepare output tensor for activation result
    auto output = torch::empty_like(x_conv);
    const int total_elements = x_conv.numel();
    
    // Determine how many complete groups of 4 floats can be processed
    const int total_vec4 = total_elements / 4;  // 4 floats per 128-bit vectorized load

    // Set execution configuration
    const int block_size = 256;
    // Launch enough blocks to cover the vectorized part; leftover elements are handled within the same kernel
    const int num_blocks = (total_vec4 + block_size - 1) / block_size;

    aligned_ldg_mish_tanh_kernel<<<num_blocks, block_size>>>(
        output.data_ptr<float>(),
        x_conv.data_ptr<float>(),
        total_elements,
        total_vec4
    );

    return output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &module_fn_forward, "Fused convolution with Mish and Tanh activations using aligned __ldg() (CUDA)");
}