forked from ccf-ai-infra/GPUCodeForces
feat:add high performance focallossfuse #31
This commit is contained in:
parent
f989885dde
commit
2e2b036748
|
|
@ -0,0 +1,158 @@
|
|||
import torch
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
import os
|
||||
|
||||
# CUDA C++ 源代码字符串
|
||||
focal_loss_cpp_source = """
|
||||
torch::Tensor focal_loss_forward_backward_cuda(
|
||||
torch::Tensor logits,
|
||||
torch::Tensor targets,
|
||||
float alpha,
|
||||
float gamma
|
||||
);
|
||||
"""
|
||||
|
||||
# CUDA 源代码字符串
|
||||
focal_loss_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <cmath>
|
||||
|
||||
__global__ void focal_loss_forward_backward_kernel(
|
||||
const float* logits,
|
||||
const float* targets,
|
||||
float* loss, // 输出：损失值
|
||||
float* d_logits, // 输出：梯度
|
||||
int batch_size,
|
||||
float alpha,
|
||||
float gamma
|
||||
) {
|
||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
if (idx < batch_size) {
|
||||
float logit = logits[idx];
|
||||
float target = targets[idx];
|
||||
|
||||
// --- 1. 前向计算部分 ---
|
||||
float prob; // p
|
||||
if (logit > 0) {
|
||||
prob = 1.0f / (1.0f + expf(-logit));
|
||||
} else {
|
||||
float exp_logit = expf(logit);
|
||||
prob = exp_logit / (1.0f + exp_logit);
|
||||
}
|
||||
prob = fmaxf(fminf(prob, 1.0f - 1e-7f), 1e-7f);
|
||||
|
||||
float pt = (target > 0.5f) ? prob : (1.0f - prob);
|
||||
float bce = -(target * logf(prob) + (1.0f - target) * logf(1.0f - prob));
|
||||
float focal_weight = alpha * powf(1.0f - pt, gamma);
|
||||
loss[idx] = focal_weight * bce;
|
||||
|
||||
// --- 2. 反向梯度计算部分 ---
|
||||
// dL/dlogit = alpha * gamma * (1-pt)^(gamma-1) * (1-p) * p * log(pt) - alpha * (1-pt)^gamma * (p - target)
|
||||
float one_minus_pt = 1.0f - pt;
|
||||
float one_minus_pt_pow_gamma_minus_1 = powf(one_minus_pt, gamma - 1.0f);
|
||||
|
||||
float term1 = alpha * gamma * one_minus_pt_pow_gamma_minus_1 * (1.0f - prob) * prob * logf(pt);
|
||||
float term2 = focal_weight * (prob - target);
|
||||
|
||||
d_logits[idx] = term1 - term2;
|
||||
}
|
||||
}
|
||||
|
||||
// C++ 封装函数
|
||||
torch::Tensor focal_loss_forward_backward_cuda(
|
||||
torch::Tensor logits,
|
||||
torch::Tensor targets,
|
||||
float alpha,
|
||||
float gamma
|
||||
) {
|
||||
logits = logits.contiguous().to(torch::kFloat32);
|
||||
targets = targets.contiguous().to(torch::kFloat32);
|
||||
|
||||
auto batch_size = logits.numel();
|
||||
|
||||
// 创建一个输出张量，前半部分存loss，后半部分存梯度
|
||||
auto output = torch::empty({2 * batch_size}, logits.options());
|
||||
|
||||
float* loss_ptr = output.data_ptr<float>();
|
||||
float* d_logits_ptr = loss_ptr + batch_size;
|
||||
|
||||
const int block_size = 256;
|
||||
int num_blocks = (batch_size + block_size - 1) / block_size;
|
||||
|
||||
focal_loss_forward_backward_kernel<<<num_blocks, block_size>>>(
|
||||
logits.data_ptr<float>(),
|
||||
targets.data_ptr<float>(),
|
||||
loss_ptr,
|
||||
d_logits_ptr,
|
||||
batch_size,
|
||||
alpha,
|
||||
gamma
|
||||
);
|
||||
|
||||
return output; // 返回包含loss和grads的组合张量
|
||||
}
|
||||
"""
|
||||
|
||||
# 编译CUDA代码
|
||||
build_dir = './cuda_build_fused'
|
||||
os.makedirs(build_dir, exist_ok=True)
|
||||
|
||||
focal_loss_module = load_inline(
|
||||
name="focal_loss_fused",
|
||||
cpp_sources=focal_loss_cpp_source,
|
||||
cuda_sources=focal_loss_source,
|
||||
functions=["focal_loss_forward_backward_cuda"],
|
||||
verbose=True,
|
||||
build_directory=build_dir,
|
||||
extra_cuda_cflags=["-O3"]
|
||||
)
|
||||
|
||||
class FocalLossFunction(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(ctx, logits, targets, alpha, gamma):
|
||||
# 调用融合内核
|
||||
output = focal_loss_module.focal_loss_forward_backward_cuda(logits, targets, alpha, gamma)
|
||||
|
||||
batch_size = logits.numel()
|
||||
loss = output.narrow(0, 0, batch_size)
|
||||
d_logits = output.narrow(0, batch_size, batch_size)
|
||||
|
||||
# 为反向传播保存梯度和超参数
|
||||
ctx.save_for_backward(d_logits)
|
||||
ctx.alpha = alpha
|
||||
ctx.gamma = gamma
|
||||
return loss
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, grad_output):
|
||||
# 从上下文中恢复梯度
|
||||
d_logits, = ctx.saved_tensors
|
||||
|
||||
# grad_output是上游传来的梯度，需要相乘
|
||||
return d_logits * grad_output, None, None, None
|
||||
|
||||
class ModelNew(torch.nn.Module):
|
||||
def __init__(self, alpha=0.25, gamma=2.0, reduction='mean'):
|
||||
super(ModelNew, self).__init__()
|
||||
self.alpha = alpha
|
||||
self.gamma = gamma
|
||||
self.reduction = reduction
|
||||
self.focal_loss_fn = FocalLossFunction.apply
|
||||
|
||||
def forward(self, logits, targets):
|
||||
logits = logits.to(torch.float32)
|
||||
targets = targets.to(torch.float32)
|
||||
|
||||
if logits.dim() == 2 and logits.size(1) == 1:
|
||||
logits = logits.squeeze(1)
|
||||
|
||||
# 调用自定义的autograd函数
|
||||
loss = self.focal_loss_fn(logits, targets, self.alpha, self.gamma)
|
||||
|
||||
if self.reduction == 'mean':
|
||||
return loss.mean()
|
||||
elif self.reduction == 'sum':
|
||||
return loss.sum()
|
||||
else:
|
||||
return loss
|
||||
|
|
@ -0,0 +1,69 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
class Model(nn.Module):
|
||||
"""
|
||||
Focal Loss implementation for binary classification.
|
||||
Focal Loss = -α * (1-pt)^γ * log(pt)
|
||||
where pt = p if target=1, else (1-p)
|
||||
"""
|
||||
def __init__(self, alpha=0.25, gamma=2.0, reduction='mean'):
|
||||
super(Model, self).__init__()
|
||||
self.alpha = alpha
|
||||
self.gamma = gamma
|
||||
self.reduction = reduction
|
||||
|
||||
def forward(self, inputs: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Compute Focal Loss between inputs and targets.
|
||||
|
||||
Args:
|
||||
inputs (torch.Tensor): Predicted logits of shape (batch_size, num_classes)
|
||||
targets (torch.Tensor): Ground truth labels of shape (batch_size,)
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Computed focal loss
|
||||
"""
|
||||
# 确保输入输出类型完全一致
|
||||
inputs = inputs.to(torch.float32)
|
||||
targets = targets.to(torch.float32)
|
||||
|
||||
# 确保targets的形状与inputs匹配
|
||||
if inputs.dim() == 2 and inputs.size(1) == 1:
|
||||
targets = targets.unsqueeze(1) # 将targets从[batch_size]变为[batch_size, 1]
|
||||
|
||||
# Convert logits to probabilities
|
||||
probs = torch.sigmoid(inputs)
|
||||
|
||||
# Compute pt
|
||||
pt = torch.where(targets == 1, probs, 1 - probs)
|
||||
|
||||
# Compute focal weight
|
||||
focal_weight = self.alpha * torch.pow(1 - pt, self.gamma)
|
||||
|
||||
# Compute binary cross entropy
|
||||
bce = F.binary_cross_entropy_with_logits(inputs, targets, reduction='none')
|
||||
|
||||
# Apply focal weight
|
||||
focal_loss = focal_weight * bce
|
||||
|
||||
if self.reduction == 'mean':
|
||||
return focal_loss.mean()
|
||||
elif self.reduction == 'sum':
|
||||
return focal_loss.sum()
|
||||
else:
|
||||
return focal_loss
|
||||
|
||||
batch_size = 1024 # 增大批次以更好地观察性能差异
|
||||
num_classes = 1 # Binary classification
|
||||
|
||||
def get_inputs():
|
||||
# Generate random logits - 明确指定float32
|
||||
inputs = torch.randn(batch_size, num_classes, dtype=torch.float32)
|
||||
# Generate random binary targets (0 or 1) - 明确指定float32
|
||||
targets = torch.randint(0, 2, (batch_size,), dtype=torch.float32)
|
||||
return [inputs, targets]
|
||||
|
||||
def get_init_inputs():
|
||||
return [] # No special initialization inputs needed
|
||||
|
|
@ -0,0 +1,102 @@
|
|||
You write custom CUDA kernels to replace the pytorch operators in the given 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 matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.
|
||||
|
||||
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
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self) -> None:
|
||||
super().__init__()
|
||||
|
||||
def forward(self, a, b):
|
||||
return a + b
|
||||
|
||||
def get_inputs():
|
||||
a = torch.randn(1, 128).cuda()
|
||||
b = torch.randn(1, 128).cuda()
|
||||
return [a, b]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
The example new arch with custom CUDA kernels looks like this:
|
||||
|
||||
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self) -> None:
|
||||
super().__init__()
|
||||
|
||||
def forward(self, a, b):
|
||||
return a + b
|
||||
|
||||
def get_inputs():
|
||||
a = torch.randn(1, 128).cuda()
|
||||
b = torch.randn(1, 128).cuda()
|
||||
return [a, b]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
You are given the following architecture:
|
||||
|
||||
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
class Model(nn.Module):
|
||||
"""
|
||||
Baseline Focal Loss implementation using fundamental PyTorch ops.
|
||||
This version avoids any built-in fused functions like `binary_cross_entropy_with_logits`
|
||||
to provide a fair comparison against a custom fused kernel.
|
||||
"""
|
||||
def __init__(self, alpha=0.25, gamma=2.0, reduction='mean'):
|
||||
super(Model, self).__init__()
|
||||
self.alpha = alpha
|
||||
self.gamma = gamma
|
||||
self.reduction = reduction
|
||||
|
||||
def forward(self, inputs: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
|
||||
# 1. Manually implement Sigmoid
|
||||
probs = torch.sigmoid(inputs)
|
||||
|
||||
# 2. Manually implement Binary Cross Entropy
|
||||
eps = 1e-7
|
||||
bce = -(targets * torch.log(probs + eps) + (1 - targets) * torch.log(1 - probs + eps))
|
||||
|
||||
# 3. Compute pt
|
||||
pt = torch.where(targets == 1, probs, 1 - probs)
|
||||
|
||||
# 4. Compute Focal Weight
|
||||
focal_weight = self.alpha * torch.pow(1 - pt, self.gamma)
|
||||
|
||||
# 5. Apply Focal Weight
|
||||
focal_loss = focal_weight * bce
|
||||
|
||||
if self.reduction == 'mean':
|
||||
return focal_loss.mean()
|
||||
elif self.reduction == 'sum':
|
||||
return focal_loss.sum()
|
||||
else:
|
||||
return focal_loss
|
||||
|
||||
batch_size = 1024
|
||||
num_classes = 1
|
||||
|
||||
def get_inputs():
|
||||
inputs = torch.randn(batch_size, num_classes, dtype=torch.float32)
|
||||
targets = torch.randint(0, 2, (batch_size,), dtype=torch.float32)
|
||||
return [inputs, targets]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
IMPORTANT: The baseline architecture is a step-by-step implementation of Focal Loss, involving multiple separate operations (sigmoid, log, pow, etc.). This creates several intermediate tensors (probs, bce, pt, focal_weight) and requires multiple kernel launches during the forward pass. Furthermore, during backpropagation, PyTorch’s autograd engine will launch separate kernels for each operation’s gradient calculation. The primary optimization goal is operator fusion across the forward and backward passes: combine the entire forward computation and the entire backward gradient computation into a single, highly efficient CUDA kernel. This fusion should eliminate all intermediate tensors, drastically reduce memory traffic, and minimize kernel launch overhead. Focus on creating a novel torch.autograd.Function that encapsulates this fused logic, providing a seamless drop-in replacement that significantly accelerates the training process.
|
||||
|
|
@ -0,0 +1,86 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
|
||||
# --- 修改点 1: 更新 import 语句，指向我们新的文件 ---
|
||||
from focalloss_fused_torchcode import Model, get_inputs, get_init_inputs
|
||||
from focalloss_fused_cudacode import ModelNew
|
||||
|
||||
def run_benchmark():
|
||||
# 检查 CUDA 是否可用
|
||||
if not torch.cuda.is_available():
|
||||
print("CUDA 不可用，请确保您有可用的 NVIDIA GPU 并已正确安装 PyTorch CUDA 版本。")
|
||||
return
|
||||
else:
|
||||
device = torch.device("cuda")
|
||||
|
||||
# 初始化模型
|
||||
init_inputs = get_init_inputs()
|
||||
init_inputs = [
|
||||
x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in init_inputs
|
||||
]
|
||||
inputs = get_inputs()
|
||||
inputs = [
|
||||
x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in inputs
|
||||
]
|
||||
|
||||
torch_model = Model(*init_inputs).cuda()
|
||||
cuda_model = ModelNew(*init_inputs).cuda()
|
||||
|
||||
torch_model.eval()
|
||||
cuda_model.eval()
|
||||
|
||||
print("-------------------- 精度对齐验证 --------------------")
|
||||
with torch.no_grad():
|
||||
output_torch = torch_model(*inputs)
|
||||
output_cuda = cuda_model(*inputs)
|
||||
|
||||
precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-03, atol=1e-05)
|
||||
max_diff = torch.max(torch.abs(output_torch - output_cuda)).item()
|
||||
mean_diff = torch.mean(torch.abs(output_torch - output_cuda)).item()
|
||||
|
||||
if precision_flag:
|
||||
print(f"✅ 精度对齐：两个模型的输出结果非常接近。")
|
||||
print(f"最大误差: {max_diff:.8f}, 平均误差: {mean_diff:.8f}")
|
||||
else:
|
||||
print(f"❌ 精度不一致！最大误差: {max_diff:.8f}, 平均误差: {mean_diff:.8f}")
|
||||
|
||||
print("\n-------------------- 性能加速比测试 --------------------")
|
||||
num_iterations = 100
|
||||
|
||||
# GPU 预热
|
||||
for _ in range(10):
|
||||
_ = torch_model(*inputs)
|
||||
_ = cuda_model(*inputs)
|
||||
|
||||
# PyTorch 模型计时
|
||||
torch.cuda.synchronize()
|
||||
start_time = time.time()
|
||||
for _ in range(num_iterations):
|
||||
_ = torch_model(*inputs)
|
||||
torch.cuda.synchronize()
|
||||
torch_time = (time.time() - start_time) / num_iterations
|
||||
|
||||
# 自定义 CUDA 内核计时
|
||||
torch.cuda.synchronize()
|
||||
start_time = time.time()
|
||||
for _ in range(num_iterations):
|
||||
_ = cuda_model(*inputs)
|
||||
torch.cuda.synchronize()
|
||||
cuda_time = (time.time() - start_time) / num_iterations
|
||||
|
||||
print(f"PyTorch 基准 Focal Loss 平均执行时间: {torch_time:.6f} 秒")
|
||||
print(f"自定义 Fused Focal Loss 平均执行时间: {cuda_time:.6f} 秒")
|
||||
speedup = 0
|
||||
if cuda_time > 0:
|
||||
speedup = torch_time / cuda_time
|
||||
print(f"加速比: {speedup:.2f}x")
|
||||
else:
|
||||
print("CUDA 内核执行时间为0，无法计算加速比。")
|
||||
return precision_flag, speedup
|
||||
|
||||
if __name__ == "__main__":
|
||||
precision_flag, speedup = run_benchmark()
|
||||
Loading…
Reference in New Issue