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>

// Fused activation: applies Mish followed by Tanh
// Mish(x) = x * tanhf(softplus(x)) where softplus(x) = log(1 + exp(x))
__device__ __forceinline__ float fused_mish_tanh_activation(float x) {
    float softplus = logf(1.0f + expf(x));
    return tanhf(x * tanhf(softplus));
}

// Kernel using vectorized loads/stores without any shared memory barriers
// Each thread processes groups of 4 floats using __ldg() for efficient, read-only cache access.
// No __syncthreads() is needed since all threads work independently on their assigned indices.
__global__ void vec_nosync_mish_tanh_kernel(
    float* __restrict__ output,
    const float* __restrict__ input,
    int total_elements,
    int total_vec4
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;
    
    // Process vectorized data (groups of 4 floats) without any synchronization
    for (int i = idx; i < total_vec4; i += stride) {
        float4 in_val = __ldg(reinterpret_cast<const float4*>(input) + i);
        float4 out_val;
        // Process each element in the float4 vector
        float* in_ptr = reinterpret_cast<float*>(&in_val);
        float* out_ptr = reinterpret_cast<float*>(&out_val);
        #pragma unroll
        for (int j = 0; j < 4; j++) {
            out_ptr[j] = fused_mish_tanh_activation(in_ptr[j]);
        }
        reinterpret_cast<float4*>(output)[i] = out_val;
    }

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

// Module forward function: perform 3D convolution then apply the fused Mish-Tanh activation
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 optimized conv3d
    auto x_conv = at::conv3d(
        x,
        conv_weight,
        conv_bias,
        {stride, stride, stride},
        {padding, padding, padding}
    );
    
    auto output = torch::empty_like(x_conv);
    int total_elements = x_conv.numel();
    int total_vec4 = total_elements / 4;
    
    const int block_size = 256;
    int num_blocks = (total_vec4 + block_size - 1) / block_size;
    // Limit grid size to avoid oversubscribing
    if (num_blocks > 2048) {
        num_blocks = 2048;
    }
    
    vec_nosync_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, "Vectorized no-sync Mish-Tanh activation fused with conv3d (CUDA)");
}