The AI CUDA Engineer 👷

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch as th


def module_fn(
    x: torch.Tensor,
    clusters: torch.Tensor,
    clusters2: torch.Tensor,
    bn_weight: torch.Tensor,
    bn_bias: torch.Tensor,
    bn_running_mean: torch.Tensor,
    bn_running_var: torch.Tensor,
    feature_size: int,
    cluster_size: int,
    is_training: bool,
) -> torch.Tensor:
    """
    Functional version of the NetVLAD without ghost clusters

    Args:
        x: Input tensor of shape (batch_size, num_features, feature_size)
        clusters: Weight tensor for cluster assignments
        clusters2: Weight tensor for visual words
        bn_weight: BatchNorm weight
        bn_bias: BatchNorm bias
        bn_running_mean: BatchNorm running mean
        bn_running_var: BatchNorm running var
        feature_size: Size of each feature
        cluster_size: Number of clusters (excluding ghost clusters)
        is_training: Whether in training mode

    Returns:
        Output tensor of shape (batch_size, cluster_size * feature_size)
    """
    max_sample = x.size()[1]
    x = x.view(-1, feature_size)  # B x N x D -> BN x D

    if x.device != clusters.device:
        msg = f"x.device {x.device} != cluster.device {clusters.device}"
        raise ValueError(msg)

    assignment = th.matmul(x, clusters)  # (BN x D) x (D x (K+G)) -> BN x (K+G)
    assignment = F.batch_norm(
        assignment,
        bn_running_mean,
        bn_running_var,
        bn_weight,
        bn_bias,
        training=is_training,
    )

    assignment = F.softmax(assignment, dim=1)  # BN x (K+G) -> BN x (K+G)
    # remove ghost assigments
    assignment = assignment[:, :cluster_size]
    assignment = assignment.view(-1, max_sample, cluster_size)  # -> B x N x K
    a_sum = th.sum(assignment, dim=1, keepdim=True)  # B x N x K -> B x 1 x K
    a = a_sum * clusters2

    assignment = assignment.transpose(1, 2)  # B x N x K -> B x K x N

    x = x.view(-1, max_sample, feature_size)  # BN x D -> B x N x D
    vlad = th.matmul(assignment, x)  # (B x K x N) x (B x N x D) -> B x K x D
    vlad = vlad.transpose(1, 2)  # -> B x D x K
    vlad = vlad - a

    # L2 intra norm
    vlad = F.normalize(vlad)

    # flattening + L2 norm
    vlad = vlad.reshape(-1, cluster_size * feature_size)  # -> B x DK
    vlad = F.normalize(vlad)
    return vlad  # B x DK


class Model(nn.Module):
    def __init__(self, cluster_size, feature_size, ghost_clusters):
        super(Model, self).__init__()

        self.feature_size = feature_size
        self.cluster_size = cluster_size
        self.ghost_clusters = ghost_clusters

        init_sc = 1 / math.sqrt(feature_size)
        clusters = cluster_size + ghost_clusters

        # The `clusters` weights are the `(w,b)` in the paper
        self.clusters = nn.Parameter(init_sc * th.randn(feature_size, clusters))

        # Extract batchnorm parameters
        bn = nn.BatchNorm1d(clusters)
        self.bn_weight = nn.Parameter(bn.weight.data.clone())
        self.bn_bias = nn.Parameter(bn.bias.data.clone())
        self.bn_running_mean = nn.Parameter(bn.running_mean.data.clone())
        self.bn_running_var = nn.Parameter(bn.running_var.data.clone())

        # The `clusters2` weights are the visual words `c_k` in the paper
        self.clusters2 = nn.Parameter(init_sc * th.randn(1, feature_size, cluster_size))
        self.out_dim = self.cluster_size * feature_size

    def forward(self, x, fn=module_fn):
        return fn(
            x,
            self.clusters,
            self.clusters2,
            self.bn_weight,
            self.bn_bias,
            self.bn_running_mean,
            self.bn_running_var,
            self.feature_size,
            self.cluster_size,
            self.training,
        )


batch_size = 32
num_features = 100
num_clusters = 32
feature_size = 512
ghost_clusters = 0


def get_inputs():
    return [torch.randn(batch_size, num_features, feature_size)]


def get_init_inputs():
    return [num_clusters, feature_size, ghost_clusters]

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

#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)

constexpr int TILE_SIZE = 128;

__global__ void fused_assignment_kernel(
    const float* __restrict__ x,
    const float* __restrict__ clusters,
    const float* bn_weight,
    const float* bn_bias,
    const float* bn_mean,
    const float* bn_var,
    float* output,
    int64_t BxN,
    int64_t D,
    int64_t KplusG,
    bool is_training) {
    
    int row = blockIdx.x * blockDim.y + threadIdx.y;
    int tid = threadIdx.x;
    int col = threadIdx.y;
    
    __shared__ float smem_max[TILE_SIZE];
    __shared__ float smem_sum[TILE_SIZE];
    
    if (row >= BxN) return;
    
    extern __shared__ float shared[];
    float* row_cache = shared;
    
    // Compute matmul row
    float sum = 0.0f;
    for (int i = tid; i < D; i += TILE_SIZE) {
        sum += x[row * D + i] * clusters[i * KplusG + col];
    }
    atomicAdd(&row_cache[col], sum);
    
    __syncthreads();
    
    // Apply BN
    float val = row_cache[col];
    if (!is_training) {
        val = (val - bn_mean[col]) * bn_weight[col] / sqrtf(bn_var[col] + 1e-5f) + bn_bias[col];
    }
    
    // Softmax reduction
    float max_val = -INFINITY;
    for (int i = tid; i < KplusG; i += TILE_SIZE)
        max_val = fmaxf(max_val, row_cache[i]);
    smem_max[tid] = max_val;
    
    __syncthreads();
    
    for (int s = blockDim.x/2; s > 0; s >>= 1) {
        if (tid < s)
            smem_max[tid] = fmaxf(smem_max[tid], smem_max[tid + s]);
        __syncthreads();
    }
    
    max_val = smem_max[0];
    
    float sum_exp = 0.0f;
    val = __expf(val - max_val);
    for (int i = tid; i < KplusG; i += TILE_SIZE)
        sum_exp += row_cache[i];
    smem_sum[tid] = sum_exp;
    
    __syncthreads();
    
    for (int s = blockDim.x/2; s > 0; s >>= 1) {
        if (tid < s)
            smem_sum[tid] += smem_sum[tid + s];
        __syncthreads();
    }
    
    output[row * KplusG + col] = val / smem_sum[0];
}

torch::Tensor forward(
    torch::Tensor x,
    torch::Tensor clusters,
    torch::Tensor clusters2,
    torch::Tensor bn_weight,
    torch::Tensor bn_bias,
    torch::Tensor bn_running_mean,
    torch::Tensor bn_running_var,
    int64_t feature_size,
    int64_t cluster_size,
    bool is_training) {
    
    CHECK_INPUT(x);
    CHECK_INPUT(clusters);
    CHECK_INPUT(clusters2);
    CHECK_INPUT(bn_weight);
    CHECK_INPUT(bn_bias);
    CHECK_INPUT(bn_running_mean);
    CHECK_INPUT(bn_running_var);

    int64_t B = x.size(0);
    int64_t N = x.size(1);
    int64_t D = feature_size;
    int64_t K = cluster_size;
    int64_t KplusG = clusters.size(1);
    int64_t BxN = B * N;

    x = x.reshape({-1, D});
    auto assignment = torch::empty({BxN, KplusG}, x.options());

    dim3 block(TILE_SIZE, TILE_SIZE);
    dim3 grid((BxN + TILE_SIZE - 1) / TILE_SIZE);
    size_t shared_mem = KplusG * sizeof(float);
    
    fused_assignment_kernel<<<grid, block, shared_mem>>>(
        x.data_ptr<float>(),
        clusters.data_ptr<float>(),
        bn_weight.data_ptr<float>(),
        bn_bias.data_ptr<float>(),
        bn_running_mean.data_ptr<float>(),
        bn_running_var.data_ptr<float>(),
        assignment.data_ptr<float>(),
        BxN,
        D,
        KplusG,
        is_training);

    assignment = assignment.narrow(1, 0, K).reshape({B, N, K});
    
    auto a_sum = assignment.sum(1, true);
    clusters2 = clusters2.expand({B, D, K});
    auto a = clusters2 * a_sum;

    assignment = assignment.transpose(1, 2);
    x = x.reshape({B, N, D});
    auto vlad = torch::bmm(assignment, x).transpose(1, 2) - a;

    vlad = torch::nn::functional::normalize(
        vlad, torch::nn::functional::NormalizeFuncOptions().p(2).dim(1));
    vlad = vlad.reshape({B, D * K});
    vlad = torch::nn::functional::normalize(
        vlad, torch::nn::functional::NormalizeFuncOptions().p(2).dim(1));

    return vlad;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "NetVLAD forward");
}