forked from ccf-ai-infra/GPUCodeForces
finish robustscalehuber #107
This commit is contained in:
parent
10eed82956
commit
a29733e771
|
|
@ -0,0 +1,94 @@
|
|||
You write custom CUDA kernels to replace the pytorch operators in the given GeGLU 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 chunk+gelu+elementwise_mul), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.
|
||||
|
||||
# Technologies Used in This Code
|
||||
|
||||
## Core Libraries
|
||||
- **PyTorch**: Deep learning framework
|
||||
- **CUDA**: NVIDIA GPU parallel computing
|
||||
- **C++**: Kernel implementation with math.h
|
||||
|
||||
## CUDA Components
|
||||
- **CUDA kernel**: `robust_scale_huber_kernel`
|
||||
- **CUDA math functions**: `fabsf()` for absolute value
|
||||
- **Element-wise parallelism**: One thread per element
|
||||
- **Conditional branching**: Huber loss piecewise logic
|
||||
|
||||
## Mathematical Operations
|
||||
1. **Robust standardization**: `(x - center) / scale`
|
||||
2. **Huber loss function**: Piecewise quadratic/linear
|
||||
- Quadratic: `0.5 * y²` when |y| ≤ delta
|
||||
- Linear: `delta * (|y| - 0.5*delta)` when |y| > delta
|
||||
3. **Absolute value**: `fabsf(y)` for condition check
|
||||
|
||||
## Architecture
|
||||
- **Standard 1D grid**: Simple block/grid configuration
|
||||
- **Element-wise computation**: Independent processing per element
|
||||
- **Branching logic**: Condition based on delta threshold
|
||||
|
||||
## Robust Statistics Features
|
||||
- **Outlier resistance**: Huber loss reduces influence of outliers
|
||||
- **Parameterized**: User-defined center, scale, and delta
|
||||
- **Piecewise behavior**: Smooth transition at delta boundary
|
||||
- **Scale invariance**: Sensitive to scale parameter
|
||||
|
||||
## Performance Features
|
||||
- **GPU acceleration**: Parallel computation across all elements
|
||||
- **Simple operations**: Basic arithmetic and conditional logic
|
||||
- **Memory efficiency**: Direct input-output mapping
|
||||
- **Low computational cost**: Moderate operations per element
|
||||
|
||||
## Numerical Considerations
|
||||
- **Scale requirement**: scale ≠ 0 (no protection in code)
|
||||
- **Delta parameter**: Controls quadratic/linear transition point
|
||||
- **Continuity**: Huber loss is C¹ continuous at |y| = delta
|
||||
- **Absolute value**: Uses `fabsf()` for efficiency
|
||||
|
||||
## Use Case Applications
|
||||
- **Robust regression**: Loss function resistant to outliers
|
||||
- **Error metric**: Combines L1 and L2 loss properties
|
||||
- **Standardized input**: Pre-scales data before loss computation
|
||||
- **Custom loss function**: Parameterized Huber loss implementation
|
||||
|
||||
## Mathematical Properties
|
||||
- **Convex**: Huber loss is convex for optimization
|
||||
- **Differentiable**: Smooth derivative everywhere
|
||||
- **Robustness**: Less sensitive to outliers than pure L2 loss
|
||||
- **Parameter tuning**: Delta controls robustness vs sensitivity
|
||||
|
||||
|
||||
|
||||
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
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, center, scale, delta):
|
||||
super(Model, self).__init__()
|
||||
self.center = center
|
||||
self.scale = scale
|
||||
self.delta = delta
|
||||
|
||||
def forward(self, x):
|
||||
y = (x - self.center) / self.scale
|
||||
abs_y = torch.abs(y)
|
||||
|
||||
quad = 0.5 * y * y
|
||||
linear = self.delta * (abs_y - 0.5 * self.delta)
|
||||
|
||||
return torch.where(abs_y <= self.delta, quad, linear)
|
||||
|
||||
|
||||
batch_size = 1024
|
||||
dim = 1024
|
||||
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(batch_size, dim) * 10.0
|
||||
return [x]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return [0.0, 1.0, 1.0]
|
||||
|
|
@ -0,0 +1,78 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <math.h>
|
||||
|
||||
__global__ void robust_scale_huber_kernel(
|
||||
const float* __restrict__ input,
|
||||
float* __restrict__ output,
|
||||
float center,
|
||||
float scale,
|
||||
float delta,
|
||||
int size
|
||||
) {
|
||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
if (idx < size) {
|
||||
float x = input[idx];
|
||||
|
||||
// Robust Standardization: Y = (X - center) / scale
|
||||
float y = (x - center) / scale;
|
||||
float abs_y = fabsf(y);
|
||||
|
||||
if (abs_y <= delta) {
|
||||
// Quadratic: 0.5 * Y^2
|
||||
output[idx] = 0.5f * y * y;
|
||||
} else {
|
||||
// Linear: delta * (|Y| - 0.5 * delta)
|
||||
output[idx] = delta * (abs_y - 0.5f * delta);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor robust_scale_huber_cuda(torch::Tensor input, float center, float scale, float delta) {
|
||||
auto output = torch::empty_like(input);
|
||||
int size = input.numel();
|
||||
|
||||
const int block_size = 256;
|
||||
int num_blocks = (size + block_size - 1) / block_size;
|
||||
|
||||
robust_scale_huber_kernel<<<num_blocks, block_size>>>(
|
||||
input.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
center,
|
||||
scale,
|
||||
delta,
|
||||
size
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
cpp_source = """
|
||||
torch::Tensor robust_scale_huber_cuda(torch::Tensor input, float center, float scale, float delta);
|
||||
"""
|
||||
|
||||
module = load_inline(
|
||||
name="robust_scale_huber",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=["robust_scale_huber_cuda"],
|
||||
verbose=True
|
||||
)
|
||||
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, center, scale, delta):
|
||||
super(ModelNew, self).__init__()
|
||||
self.center = center
|
||||
self.scale = scale
|
||||
self.delta = delta
|
||||
self.module = module
|
||||
|
||||
def forward(self, x):
|
||||
return self.module.robust_scale_huber_cuda(x, self.center, self.scale, self.delta)
|
||||
|
|
@ -0,0 +1,32 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, center, scale, delta):
|
||||
super(Model, self).__init__()
|
||||
self.center = center
|
||||
self.scale = scale
|
||||
self.delta = delta
|
||||
|
||||
def forward(self, x):
|
||||
y = (x - self.center) / self.scale
|
||||
abs_y = torch.abs(y)
|
||||
|
||||
quad = 0.5 * y * y
|
||||
linear = self.delta * (abs_y - 0.5 * self.delta)
|
||||
|
||||
return torch.where(abs_y <= self.delta, quad, linear)
|
||||
|
||||
|
||||
batch_size = 1024
|
||||
dim = 1024
|
||||
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(batch_size, dim) * 10.0
|
||||
return [x]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return [0.0, 1.0, 1.0]
|
||||
|
|
@ -0,0 +1,77 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from robustscalehuber_torch import Model, get_inputs, get_init_inputs
|
||||
from robustscalehuber_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)
|
||||
|
||||
precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-03)
|
||||
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 = 100
|
||||
|
||||
# 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 torch.relu 平均执行时间: {torch_time:.6f} 秒")
|
||||
print(f"自定义 CUDA 内核 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()
|
||||
Loading…
Reference in New Issue