forked from ccf-ai-infra/GPUCodeForces
147 lines
5.1 KiB
Plaintext
147 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.
|
|
|
|
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
|
|
|
|
Negative Log-Likelihood Loss (NLLLoss): Classification loss function for probability distributions
|
|
|
|
Dual-Kernel Strategy: Separate kernels for "none" reduction vs "mean"/"sum" reduction
|
|
|
|
Grid-Stride Loops: Efficiently processes data of arbitrary size using fixed thread blocks
|
|
|
|
Shared Memory Reduction: Uses __shared__ arrays for block-level parallel reduction
|
|
|
|
Tree Reduction Pattern: Binary tree reduction within thread blocks using __syncthreads()
|
|
|
|
Device Function: compute_nll_loss_item helper function shared between kernels
|
|
|
|
Conditional Weight Handling: Supports optional class weights with null pointer checking
|
|
|
|
Ignore Index Support: Filters out specified target indices from loss calculation
|
|
|
|
Multi-Dimensional Tensor Support: Handles 2D+ inputs with spatial dimensions
|
|
|
|
Tensor Flattening: Converts multi-dimensional tensors to flat views for kernel processing
|
|
|
|
Two-Stage Reduction: Block-level partial reduction followed by host-side final reduction
|
|
|
|
Boundary Checking: Validates target indices and handles out-of-range values
|
|
|
|
Memory Coalescing: Ensures contiguous tensor layout for optimal memory access
|
|
|
|
Fast Math Operations: Uses --use_fast_math compiler flag
|
|
|
|
Comprehensive Input Validation: Checks tensor dimensions, types, and device placement
|
|
|
|
Zero-Size Tensor Handling: Returns zero loss for empty inputs
|
|
|
|
Numerical Stability: Handles zero total weight case for mean reduction
|
|
|
|
Flexible Reduction Modes: Supports "none", "mean", and "sum" reduction strategies
|
|
|
|
Optional Tensor Handling: Uses c10::optional for optional weight parameter
|
|
|
|
|
|
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
|
|
|
|
# -------------------------------------------------------------
|
|
# 常量定义
|
|
# -------------------------------------------------------------
|
|
N, C, H, W = 8, 10, 16, 16 # (N, C, H, W)
|
|
|
|
# 损失函数参数
|
|
WEIGHT = torch.rand(C, dtype=torch.float32) # (C,)
|
|
IGNORE_INDEX = -100
|
|
REDUCTION = 'mean'
|
|
|
|
|
|
# -------------------------------------------------------------
|
|
|
|
class Model(nn.Module):
|
|
"""
|
|
nn.NLLLoss 的纯 PyTorch 基准实现
|
|
(K-dim, 2D-example)
|
|
"""
|
|
|
|
def __init__(self, weight=None, size_average=None, ignore_index=-100,
|
|
reduce=None, reduction='mean'):
|
|
super().__init__()
|
|
|
|
# 处理已弃用的 size_average 和 reduce
|
|
if size_average is not None or reduce is not None:
|
|
# (省略... 遵循 torch.nn.modules.loss)
|
|
pass
|
|
|
|
self.reduction = reduction
|
|
self.ignore_index = ignore_index
|
|
|
|
# 确保 weight 在正确的设备上
|
|
if weight is not None:
|
|
self.register_buffer('weight', weight)
|
|
else:
|
|
self.weight = None
|
|
|
|
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
|
|
|
|
input_flat = input.view(N, C, -1)
|
|
|
|
target_flat = target.view(N, -1)
|
|
|
|
|
|
loss_unreduced = input_flat.gather(dim=1, index=target_flat.unsqueeze(1))
|
|
loss_unreduced = -loss_unreduced.squeeze(1) # (N, H*W)
|
|
|
|
|
|
if self.weight is not None:
|
|
|
|
weights_applied = self.weight[target_flat]
|
|
loss_unreduced = loss_unreduced * weights_applied
|
|
else:
|
|
|
|
weights_applied = torch.ones_like(target_flat, dtype=input.dtype)
|
|
|
|
|
|
mask = (target_flat != self.ignore_index)
|
|
loss_unreduced = loss_unreduced * mask
|
|
weights_applied = weights_applied * mask
|
|
|
|
|
|
if self.reduction == 'mean':
|
|
|
|
total_weight = weights_applied.sum()
|
|
if total_weight == 0:
|
|
return torch.tensor(0.0, device=input.device, dtype=input.dtype)
|
|
return loss_unreduced.sum() / total_weight
|
|
|
|
elif self.reduction == 'sum':
|
|
return loss_unreduced.sum()
|
|
|
|
else:
|
|
return loss_unreduced.view_as(target)
|
|
|
|
|
|
def get_inputs():
|
|
|
|
input_log_probs = F.log_softmax(torch.randn(N, C, H, W, dtype=torch.float32), dim=1)
|
|
target = torch.empty(N, H, W, dtype=torch.long).random_(0, C)
|
|
|
|
|
|
target.view(-1)[::10] = IGNORE_INDEX
|
|
|
|
return [input_log_probs, target]
|
|
|
|
|
|
def get_init_inputs():
|
|
|
|
return [WEIGHT, None, IGNORE_INDEX, None, REDUCTION] |