You write custom CUDA kernels to replace the PyTorch operators in the given EvoNorm architecture to get speedups.
You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining normalization+affine_transform+nonlinear_gating), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.

Technologies Used :

PyTorch: Deep learning framework

CUDA: GPU acceleration for parallel computing

C++/CUDA C++: High-performance kernel programming

Inline C++/CUDA Extension (torch.utils.cpp_extension.load_inline): Just-In-Time (JIT) compilation of custom operators

Jaccard Similarity (Intersection over Union): Similarity measure between sets or vectors

Vectorized Memory Access (float4): Uses 128-bit wide loads (4 floats) to improve memory bandwidth utilization

Instruction-Level Parallelism (ILP): Processes multiple float4 elements per loop iteration to hide instruction latency

Warp-Level Primitives: Uses __shfl_down_sync for efficient intra-warp reduction

Two-Stage Parallel Reduction: Combines warp-level reduction with shared memory and block-level reduction

Multi-Dimensional Grid Layout: Uses dim3(splits, N) for parallel processing across splits and batches

Dynamic Kernel Configuration: Calculates optimal split count based on GPU SM count and data size

Fused Kernel Design: Computes dot product and squared sums simultaneously in single kernel

Atomic Operations (atomicAdd): Safely accumulates results from multiple thread blocks to temporary buffer

Temporary Buffer Strategy: Uses pre-allocated buffer [N, 3] to store intermediate results (dot, sq1, sq2)

Constant Memory/__ldg: Uses read-only data cache for improved memory access patterns

Fast Math Operations: Uses FMA operations with --use_fast_math compiler flag

Memory Coalescing: Optimized memory access patterns through contiguous tensor layout

Pointer Chasing Loop: Efficient main loop with ILP-unrolled memory access patterns

Tail Processing: Handles remaining elements after main vectorized loop

Numerical Stability: Adds epsilon (eps) to prevent division by zero in final calculation

Buffer Zeroing: Clears temporary buffer before each forward pass

Device Query API: Uses cudaGetDevice and cudaDeviceGetAttribute for optimal kernel configuration

Three-Accumulator Design: Maintains separate accumulators for dot product, x1 squared, and x2 squared

Efficient Union Calculation: Computes Jaccard similarity using algebraic identity: union = sum(sq1) + sum(sq2) - sum(dot)

Shared Memory for Warp Results: Uses separate shared memory arrays for each reduction variable

Boundary Checking: Handles data size variations and split boundaries safely


Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
import torch
import torch.nn as nn
import torch.nn.functional as F

N, C, H, W = 32, 64, 56, 56
EPS = 1e-6


class JaccardSimilarity(nn.Module):

    def __init__(self, eps=1e-6):
        super().__init__()
        self.eps = eps

    def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
        intersection = torch.sum(x1 * x2, dim=[1, 2, 3])

        sum_sq1 = torch.sum(x1 * x1, dim=[1, 2, 3])
        sum_sq2 = torch.sum(x2 * x2, dim=[1, 2, 3])

        union = sum_sq1 + sum_sq2 - intersection

        return intersection / (union + self.eps)


class Model(nn.Module):
    def __init__(self):
        super().__init__()
        self.op = JaccardSimilarity(EPS)

    def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
        return self.op(x1, x2)


def get_inputs():
    x1 = torch.randn(N, C, H, W, dtype=torch.float32)
    x2 = torch.randn(N, C, H, W, dtype=torch.float32)
    return [x1, x2]


def get_init_inputs():
    return []