zizi05
/
GPUCodeForces
forked from ccf-ai-infra/GPUCodeForces
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request 'finish multilabelmarginloss #44' (#73) from hli28146/GPUCodeForces:MultiLabelMarginLoss into main

This commit is contained in:
Kuohais 2025-11-13 11:35:05 +08:00
parent c335d71e2b d6c100a2f5
commit 3a164824d6
4 changed files with 344 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

156
S1/44/multilabelmarginloss_cuda.py Normal file
View File

@ -0,0 +1,156 @@
import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline
cpp_source = """
#include <torch/extension.h>
#include <string>
// 函数前向声明
torch::Tensor multi_label_margin_loss_cuda_forward(
const torch::Tensor& input,
const torch::Tensor& target,
const std::string& reduction
);
"""
cuda_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>
#define BLOCK_SIZE 256
// 设置一个合理的单个样本最大正类标签数，用于共享内存数组
#define MAX_POSITIVE_LABELS 32
template <typename T>
__global__ void multi_label_margin_loss_kernel(
T* output, // (N)
const T* input, // (N, C)
const long* target, // (N, C)
const int N,
const int C)
{
int sample_idx = blockIdx.x;
if (sample_idx >= N) return;
__shared__ long positive_indices[MAX_POSITIVE_LABELS];
__shared__ int num_positives;
// 线程 0 初始化共享内存计数器
if (threadIdx.x == 0) {
num_positives = 0;
}
__syncthreads();
for (int i = threadIdx.x; i < C; i += blockDim.x) {
long label = target[sample_idx * C + i];
if (label != -1) {
int index = atomicAdd(&num_positives, 1);
if (index < MAX_POSITIVE_LABELS) {
positive_indices[index] = label;
}
}
}
__syncthreads();
__shared__ T sdata[BLOCK_SIZE];
int tid = threadIdx.x;
T my_sum = 0.0f;
const T* input_row = input + sample_idx * C;
for (int neg_class_idx = tid; neg_class_idx < C; neg_class_idx += blockDim.x) {
// 检查当前类别是否为正类
bool is_positive = false;
for (int j = 0; j < num_positives; ++j) {
if (positive_indices[j] == neg_class_idx) {
is_positive = true;
break;
}
}
// 如果是负类，则计算与所有正类的损失
if (!is_positive) {
T x_neg = input_row[neg_class_idx];
for (int j = 0; j < num_positives; ++j) {
long pos_class_idx = positive_indices[j];
T x_pos = input_row[pos_class_idx];
T loss_term = 1.0f - (x_pos - x_neg);
if (loss_term > 0) {
my_sum += loss_term;
}
}
}
}
sdata[tid] = my_sum;
__syncthreads();
for (int s = blockDim.x / 2; s > 0; s >>= 1) {
if (tid < s) {
sdata[tid] += sdata[tid + s];
}
__syncthreads();
}
if (tid == 0) {
output[sample_idx] = sdata[0] / C;
}
}
torch::Tensor multi_label_margin_loss_cuda_forward(
const torch::Tensor& input,
const torch::Tensor& target,
const std::string& reduction)
{
TORCH_CHECK(input.is_cuda() && target.is_cuda(), "Tensors must be on CUDA");
TORCH_CHECK(input.dim() == 2, "Input must be 2D");
TORCH_CHECK(target.dim() == 2, "Target must be 2D");
TORCH_CHECK(input.size(0) == target.size(0), "Batch sizes must match");
TORCH_CHECK(input.is_contiguous() && target.is_contiguous(), "Tensors must be contiguous");
const int N = input.size(0);
const int C = input.size(1);
auto options = torch::TensorOptions().device(input.device()).dtype(input.dtype());
auto sample_losses = torch::empty({N}, options);
dim3 grid(N);
dim3 block(BLOCK_SIZE);
AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "multi_label_margin_loss_kernel", ([&] {{
multi_label_margin_loss_kernel<scalar_t><<<grid, block>>>(
sample_losses.data_ptr<scalar_t>(),
input.data_ptr<scalar_t>(),
target.data_ptr<long>(),
N, C
);
}}));
if (reduction == "none") {
return sample_losses;
} else if (reduction == "sum") {
return sample_losses.sum();
} else {{ // "mean"
return sample_losses.mean();
}}
}
"""
class ModelNew(nn.Module):
def __init__(self, reduction='mean'):
super(ModelNew, self).__init__()
self.reduction = reduction
self.op = load_inline(
name='multi_label_margin_loss_op',
cpp_sources=cpp_source,
cuda_sources=cuda_source,
functions=['multi_label_margin_loss_cuda_forward'],
verbose=False
)
def forward(self, input_tensor: torch.Tensor, target_tensor: torch.Tensor) -> torch.Tensor:
return self.op.multi_label_margin_loss_cuda_forward(
input_tensor,
target_tensor,
self.reduction
)

34
S1/44/multilabelmarginloss_torch.py Normal file
View File

@ -0,0 +1,34 @@
import torch
import torch.nn as nn
import numpy as np
BATCH_SIZE = 512
NUM_CLASSES = 1024
REDUCTION = 'mean'
# 每个样本的正类标签数量范围
MIN_LABELS = 1
MAX_LABELS = 10
class Model(nn.Module):
def __init__(self, reduction='mean'):
super(Model, self).__init__()
self.loss_fn = nn.MultiLabelMarginLoss(reduction=reduction)
def forward(self, input_tensor: torch.Tensor, target_tensor: torch.Tensor) -> torch.Tensor:
return self.loss_fn(input_tensor, target_tensor)
def get_inputs():
input_tensor = torch.randn(BATCH_SIZE, NUM_CLASSES, dtype=torch.float32)
# target每行包含正类索引，并用 -1 填充
target_np = np.full((BATCH_SIZE, NUM_CLASSES), -1, dtype=np.int64)
for i in range(BATCH_SIZE):
num_labels = np.random.randint(MIN_LABELS, MAX_LABELS + 1)
labels = np.random.choice(NUM_CLASSES, num_labels, replace=False)
target_np[i, :num_labels] = labels
target_tensor = torch.from_numpy(target_np)
return [input_tensor.contiguous(), target_tensor.contiguous()]
def get_init_inputs():
return [REDUCTION]

66
S1/44/prompt.txt Normal file
View File

@ -0,0 +1,66 @@
Write a custom CUDA kernel to optimize `torch.nn.MultiLabelMarginLoss`.
The original operation is defined by the formula:
`loss(x, y) = sum_{j,i} max(0, 1 - (x[y[j]] - x[i])) / (x.size(0) * x.size(1))`
where `y[j]` are the positive class indices for a sample and `i` are the negative class indices. This is computed per sample and then reduced.
**Problem Analysis:**
`MultiLabelMarginLoss` is notoriously difficult to vectorize efficiently in PyTorch. The performance bottlenecks are severe:
1. **Irregular Data Access**: Each sample has a variable number of positive labels defined in `y`, which are padded with -1. Identifying the set of positive and negative classes for each sample requires complex, non-vectorized logic (e.g., masks, loops, or boolean indexing), which is slow.
2. **Massive Intermediate Tensors**: A naive vectorized approach would require gathering scores for positive classes and broadcasting them for subtraction against scores of negative classes. This would create huge intermediate tensors and is highly memory-inefficient.
3. **Complex Nested Loop Logic**: The core formula is a nested loop (`for each positive class`, `for each negative class`) for every sample, which is antithetical to efficient GPU execution without a custom kernel.
**Optimization Strategy: Fused Block-Level Parallelism with Shared Memory Caching**
The strategy is to implement the entire complex logic within a single CUDA kernel, using a block-per-sample parallelization model.
1. **Parallelization Model**: A grid of `N` blocks is launched, where `N` is the batch size. Each thread block is assigned to compute the total loss for one sample.
2. **Shared Memory Caching**: For each sample (block), the kernel first collaboratively reads the list of positive class indices from the `target` tensor. These indices (and their count) are cached in **shared memory**. This makes the critical metadata for the sample instantly accessible to all threads in the block.
3. **Fused Computation Loop**: The threads within a block then work together to iterate through all `C` possible classes. For each class `i`, a thread checks if it's a positive or negative class using the cached shared memory data.
* If `i` is a negative class, the thread then iterates through the *positive class indices cached in shared memory*.
* For each positive-negative pair, it calculates the hinge loss term `max(0, 1 - (x_pos - x_neg))` and accumulates it into a thread-local register. This fuses the nested loops, indexing, subtraction, and `max` operations.
4. **Efficient Intra-Block Reduction**: Once all classes are processed, a fast parallel reduction is performed using shared memory to sum the partial results from all threads within the block into a single total loss for that sample.
5. **Finalization**: The first thread of each block performs the final division and writes the result to the output tensor. The kernel directly produces the per-sample losses (`reduction='none'`). The final batch reduction (`'mean'` or `'sum'`) is efficiently handled by a single PyTorch call on the small 1D output tensor.
This approach transforms the complex, memory-bound, and hard-to-vectorize PyTorch operation into a single, efficient, compute-focused CUDA kernel.
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
```python
import torch
import torch.nn as nn
import numpy as np
BATCH_SIZE = 512
NUM_CLASSES = 1024
REDUCTION = 'mean'
# 每个样本的正类标签数量范围
MIN_LABELS = 1
MAX_LABELS = 10
class Model(nn.Module):
def __init__(self, reduction='mean'):
super(Model, self).__init__()
self.loss_fn = nn.MultiLabelMarginLoss(reduction=reduction)
def forward(self, input_tensor: torch.Tensor, target_tensor: torch.Tensor) -> torch.Tensor:
return self.loss_fn(input_tensor, target_tensor)
def get_inputs():
input_tensor = torch.randn(BATCH_SIZE, NUM_CLASSES, dtype=torch.float32)
# target每行包含正类索引，并用 -1 填充
target_np = np.full((BATCH_SIZE, NUM_CLASSES), -1, dtype=np.int64)
for i in range(BATCH_SIZE):
num_labels = np.random.randint(MIN_LABELS, MAX_LABELS + 1)
labels = np.random.choice(NUM_CLASSES, num_labels, replace=False)
target_np[i, :num_labels] = labels
target_tensor = torch.from_numpy(target_np)
return [input_tensor.contiguous(), target_tensor.contiguous()]
def get_init_inputs():
return [REDUCTION]

88
S1/44/run_code.py Normal file
View File

@ -0,0 +1,88 @@
###########################################################
# 性能和精度验证程序
###########################################################
import torch
import torch.nn as nn
import time
from multilabelmarginloss_torch import Model, get_inputs, get_init_inputs
from multilabelmarginloss_cuda 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)
# 更严格的精度检查
abs_diff = (output_torch - output_cuda).abs()
max_diff = abs_diff.max().item()
mean_diff = abs_diff.mean().item()
print(f"最大差异: {max_diff:.6f}")
print(f"平均差异: {mean_diff:.6f}")
precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-05, atol=1e-05)
if precision_flag:
print("✅ 精度对齐：两个模型的输出结果非常接近。")
else:
print("❌ 精度不一致！")
print("\n-------------------- 性能加速比测试 --------------------")
num_iterations = 1000 # 增加迭代次数以获得更准确的时间测量
# Warm up
for _ in range(100):
_ = 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 MultiLabelMarginLoss 平均执行时间: {torch_time:.6f} 秒")
print(f"自定义 CUDA MultiLabelMarginLoss 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): {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()