forked from ccf-ai-infra/GPUCodeForces
158 lines
5.1 KiB
Plaintext
158 lines
5.1 KiB
Plaintext
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.
|
||
|
||
Circle Loss CUDA Optimization with Two-Stage Reduction
|
||
|
||
Core Algorithm
|
||
Loss Function: Implements Circle Loss for deep metric learning
|
||
|
||
Key Formula:
|
||
|
||
Positive pairs: logit_p = -α_p * (s_p - Δ_p) * γ
|
||
|
||
Negative pairs: logit_n = α_n * (s_n - Δ_n) * γ
|
||
|
||
Final loss: log(1 + exp(logsumexp_p + logsumexp_n))
|
||
|
||
CUDA Optimization Techniques
|
||
1. Two-Stage Parallel Reduction
|
||
Stage 1: Find global maximum of logits for numerical stability
|
||
|
||
Kernel: circleloss_find_max_kernel
|
||
|
||
Block-level max reduction with shared memory
|
||
|
||
Grid-stride loops for load balancing
|
||
|
||
Stage 2: Compute sum of exponentials with stability
|
||
|
||
Kernel: circleloss_sum_exp_diff_kernel
|
||
|
||
Subtract global max before exponentiation
|
||
|
||
Block-level sum reduction
|
||
|
||
2. Memory Access Optimization
|
||
Coalesced Memory Access: Sequential memory access patterns
|
||
|
||
Shared Memory Utilization:
|
||
|
||
s_data_p[BLOCK_SIZE] for positive pairs
|
||
|
||
s_data_n[BLOCK_SIZE] for negative pairs
|
||
|
||
Contiguous Tensors: Ensure input tensors are contiguous
|
||
|
||
3. Numerical Stability Features
|
||
Log-Sum-Exp Trick: Subtract maximum before exponentiation
|
||
|
||
Stable Softplus: max(0,x) + log(1 + exp(-abs(x)))
|
||
|
||
Float Safety: Use FLT_MAX for initialization
|
||
|
||
4. Parallelization Strategy
|
||
Grid-Stride Loops: Handle arbitrary input sizes
|
||
|
||
Block Size: 256 threads per block for optimal occupancy
|
||
|
||
Dynamic Grid Size: (n_elements + BLOCK_SIZE - 1) / BLOCK_SIZE
|
||
|
||
5. Kernel Design Patterns
|
||
|
||
6. Performance Optimizations
|
||
Minimal CPU-GPU Synchronization: Avoid unnecessary .item() calls
|
||
|
||
Batch Matrix Multiplication: Precompute similarity matrix
|
||
|
||
Memory Pre-allocation: Pre-allocate output buffers
|
||
|
||
7. Implementation Features
|
||
Template-Free Design: Optimized for float32 precision
|
||
|
||
Error Checking: Comprehensive tensor validation
|
||
|
||
Edge Case Handling: Empty tensor and size validation
|
||
|
||
PyTorch Integration: Seamless tensor passing and gradient support
|
||
|
||
Key CUDA Concepts Used
|
||
Shared Memory Reduction for parallel statistics computation
|
||
|
||
Grid-Stride Loops for workload distribution
|
||
|
||
Memory Coalescing for efficient global memory access
|
||
|
||
Numerical Stability through careful floating-point handling
|
||
|
||
Kernel Fusion combining multiple operations in single kernels
|
||
|
||
Expected Performance Benefits
|
||
2-5x speedup over PyTorch implementation for large batch sizes
|
||
|
||
Better numerical stability with log-sum-exp trick
|
||
|
||
Scalable performance with increasing batch dimensions
|
||
|
||
GPU utilization through optimized block and grid sizing
|
||
|
||
This implementation provides a production-ready, numerically stable Circle Loss with significant performance improvements through careful CUDA optimization and parallel reduction patterns.
|
||
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
|
||
|
||
# 定义常量
|
||
BATCH_SIZE = 256
|
||
FEATURE_DIM = 512
|
||
MARGIN = 0.25
|
||
GAMMA = 256
|
||
|
||
class Model(nn.Module):
|
||
|
||
def __init__(self):
|
||
super().__init__()
|
||
self.margin = MARGIN
|
||
self.gamma = GAMMA
|
||
|
||
def forward(self, features: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
|
||
|
||
# Circle Loss的 PyTorch 实现
|
||
# 假设 features 已经是 L2 归一化的
|
||
# (B, D) @ (D, B) -> (B, B)
|
||
similarities = torch.matmul(features, features.t())
|
||
|
||
# 创建正负样本对的掩码
|
||
mask_positive = labels.unsqueeze(1) == labels.unsqueeze(0)
|
||
mask_negative = labels.unsqueeze(1) != labels.unsqueeze(0)
|
||
|
||
# 收集正样本对和负样本对的相似度
|
||
# .masked_select() 会将张量展平
|
||
sp = similarities[mask_positive]
|
||
sn = similarities[mask_negative]
|
||
|
||
# 计算 Circle Loss 的 logits
|
||
# .detach() 用于停止梯度反向传播
|
||
ap = torch.clamp_min(-sp.detach() + 1 + self.margin, min=0.)
|
||
an = torch.clamp_min(sn.detach() + self.margin, min=0.)
|
||
|
||
delta_p = 1 - self.margin
|
||
delta_n = self.margin
|
||
|
||
logit_p = -ap * (sp - delta_p) * self.gamma
|
||
logit_n = an * (sn - delta_n) * self.gamma
|
||
|
||
# 使用 logsumexp 和 softplus 计算最终的 loss
|
||
# 这是 "unified" 版本的 loss
|
||
loss = F.softplus(torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0))
|
||
|
||
return loss
|
||
|
||
def get_inputs():
|
||
|
||
# 特征需要 L2 归一化
|
||
features = F.normalize(torch.randn(BATCH_SIZE, FEATURE_DIM, dtype=torch.float32), p=2, dim=1)
|
||
labels = torch.randint(0, 10, (BATCH_SIZE,), dtype=torch.long) # 假设有 10 个类别
|
||
return [features, labels]
|
||
|
||
def get_init_inputs():
|
||
return [] |