forked from ccf-ai-infra/GPUCodeForces
80 lines
3.5 KiB
Plaintext
80 lines
3.5 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.
|
|
|
|
Technical Overview: CUDA-Optimized Triplet Margin Loss with L2 Distance
|
|
This implementation provides a high-performance CUDA kernel for computing triplet loss, designed for deep metric learning applications with optimized parallel computation and memory access patterns.
|
|
Key Features:
|
|
Architecture:
|
|
Custom CUDA kernel with inline compilation using PyTorch C++ extensions
|
|
Optimized for NVIDIA GPUs with warp-level parallelism and shared memory utilization
|
|
Supports 4-element vectorization (float4) for memory coalescing
|
|
Implements both standard and "swap" variants of triplet loss
|
|
Performance Optimizations:
|
|
Vectorized Memory Access: Uses float4 data type to load 4 elements per instruction
|
|
Coalesced Memory Reads: Contiguous memory access through __ldgintrinsic
|
|
Warp Reduction: Efficient warp-level reduction operations using __shfl_down_sync
|
|
Shared Memory: Intermediate results stored in shared memory for block-level reduction
|
|
Branch Optimization: Conditional swap computation handled efficiently
|
|
Kernel Specifications:
|
|
Block size: 256 threads
|
|
Warp size: 32 threads
|
|
Grid dimension: N (batch size)
|
|
Input requirement: Embedding dimension D must be divisible by 4
|
|
|
|
|
|
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, D = 32, 128
|
|
|
|
|
|
class TripletMarginWithDistanceLoss(nn.Module):
|
|
|
|
def __init__(self, distance_function=None, margin=1.0, swap=False, reduction='mean'):
|
|
super().__init__()
|
|
self.distance_function = distance_function if distance_function is not None else nn.PairwiseDistance()
|
|
self.margin = margin
|
|
self.swap = swap
|
|
self.reduction = reduction
|
|
|
|
def forward(self, anchor: torch.Tensor, positive: torch.Tensor, negative: torch.Tensor) -> torch.Tensor:
|
|
|
|
d_ap = self.distance_function(anchor, positive)
|
|
|
|
d_an = self.distance_function(anchor, negative)
|
|
|
|
if self.swap:
|
|
d_pn = self.distance_function(positive, negative)
|
|
d_an = torch.min(d_an, d_pn)
|
|
|
|
loss = torch.clamp(d_ap - d_an + self.margin, min=0.0)
|
|
|
|
if self.reduction == 'mean':
|
|
return loss.mean()
|
|
elif self.reduction == 'sum':
|
|
return loss.sum()
|
|
else: # 'none'
|
|
return loss
|
|
|
|
|
|
class Model(nn.Module):
|
|
def __init__(self, margin=1.0, swap=False):
|
|
super().__init__()
|
|
|
|
self.op = TripletMarginWithDistanceLoss(distance_function=nn.PairwiseDistance(), margin=margin, swap=swap)
|
|
|
|
def forward(self, a: torch.Tensor, p: torch.Tensor, n: torch.Tensor) -> torch.Tensor:
|
|
return self.op(a, p, n)
|
|
|
|
|
|
def get_inputs():
|
|
anchor = torch.randn(N, D, dtype=torch.float32)
|
|
positive = torch.randn(N, D, dtype=torch.float32)
|
|
negative = torch.randn(N, D, dtype=torch.float32)
|
|
return [anchor, positive, negative]
|
|
|
|
|
|
def get_init_inputs():
|
|
return [1.0, False] # margin, swap |