forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish NLReLU #78' (#697) from hli28146/GPUCodeForces:h78 into main
This commit is contained in:
commit
a9f3fc756d
|
|
@ -0,0 +1,109 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
# C++ 源代码 wrapper
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
|
||||
torch::Tensor nlrelu_cuda_forward(const torch::Tensor& input, float beta);
|
||||
"""
|
||||
|
||||
# CUDA 源代码
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <math.h>
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
|
||||
struct __align__(16) Float4 {
|
||||
float x, y, z, w;
|
||||
};
|
||||
|
||||
// NLReLU Logic
|
||||
__device__ __forceinline__ float compute_nlrelu(float x, float beta) {
|
||||
if (x >= 0.0f) {
|
||||
return __logf(beta * x + 1.0f);
|
||||
}
|
||||
return 0.0f;
|
||||
}
|
||||
|
||||
__global__ void nlrelu_kernel(
|
||||
float* __restrict__ output,
|
||||
const float* __restrict__ input,
|
||||
const int n,
|
||||
const float beta)
|
||||
{
|
||||
const int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
const int vec_n = n / 4;
|
||||
|
||||
int i = idx;
|
||||
const int stride = blockDim.x * gridDim.x;
|
||||
|
||||
// 1. Vectorized Loop
|
||||
for (; i < vec_n; i += stride) {
|
||||
Float4 in_vec = reinterpret_cast<const Float4*>(input)[i];
|
||||
Float4 out_vec;
|
||||
|
||||
out_vec.x = compute_nlrelu(in_vec.x, beta);
|
||||
out_vec.y = compute_nlrelu(in_vec.y, beta);
|
||||
out_vec.z = compute_nlrelu(in_vec.z, beta);
|
||||
out_vec.w = compute_nlrelu(in_vec.w, beta);
|
||||
|
||||
reinterpret_cast<Float4*>(output)[i] = out_vec;
|
||||
}
|
||||
|
||||
// 2. Scalar Tail
|
||||
int start_scalar = vec_n * 4;
|
||||
int global_tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
int total_threads = gridDim.x * gridDim.x;
|
||||
|
||||
int current_idx = start_scalar + global_tid;
|
||||
while (current_idx < n) {
|
||||
output[current_idx] = compute_nlrelu(input[current_idx], beta);
|
||||
current_idx += total_threads;
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor nlrelu_cuda_forward(const torch::Tensor& input, float beta) {
|
||||
TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
|
||||
TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
|
||||
|
||||
const int n = input.numel();
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
const int vec_n = n / 4;
|
||||
const int grid_size = (vec_n + BLOCK_SIZE - 1) / BLOCK_SIZE;
|
||||
|
||||
int final_grid = (grid_size < 1) ? 1 : grid_size;
|
||||
if (final_grid > 65535) final_grid = 65535;
|
||||
|
||||
nlrelu_kernel<<<final_grid, BLOCK_SIZE>>>(
|
||||
output.data_ptr<float>(),
|
||||
input.data_ptr<float>(),
|
||||
n,
|
||||
beta
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
nlrelu_op_module = load_inline(
|
||||
name='nlrelu_op',
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=['nlrelu_cuda_forward'],
|
||||
verbose=False,
|
||||
extra_cuda_cflags=['-O3', '--use_fast_math']
|
||||
)
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, beta=1.0):
|
||||
super(ModelNew, self).__init__()
|
||||
self.beta = beta
|
||||
self.op = nlrelu_op_module
|
||||
|
||||
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
|
||||
return self.op.nlrelu_cuda_forward(input_tensor.contiguous(), self.beta)
|
||||
|
|
@ -0,0 +1,41 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
BATCH_SIZE = 4096
|
||||
HIDDEN_DIM = 4096
|
||||
SHAPE = (BATCH_SIZE, HIDDEN_DIM)
|
||||
|
||||
# NLReLU 超参数 beta ，论文中建议范围 0.7-1.1
|
||||
BETA_VALUE = 1.0
|
||||
|
||||
class NLReLU(nn.Module):
|
||||
"""
|
||||
"Natural-Logarithm-Rectified Activation Function in Convolutional Neural Networks"
|
||||
Formula:
|
||||
f(x) = log(beta * x + 1.0) if x >= 0
|
||||
f(x) = 0 if x < 0
|
||||
"""
|
||||
def __init__(self, beta=1.0):
|
||||
super(NLReLU, self).__init__()
|
||||
self.beta = beta
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
x_relu = F.relu(x)
|
||||
inner = self.beta * x_relu + 1.0
|
||||
return torch.log(inner)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, beta=1.0):
|
||||
super(Model, self).__init__()
|
||||
self.act = NLReLU(beta=beta)
|
||||
|
||||
def forward(self, x):
|
||||
return self.act(x)
|
||||
|
||||
def get_inputs():
|
||||
input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0
|
||||
return [input_tensor.contiguous()]
|
||||
|
||||
def get_init_inputs():
|
||||
return [BETA_VALUE]
|
||||
|
|
@ -0,0 +1,69 @@
|
|||
Write a custom CUDA kernel to optimize `NLReLU` (Natural-Logarithm-Rectified Linear Unit).
|
||||
|
||||
Formula:
|
||||
f(x) = log(beta * x + 1.0) if x >= 0
|
||||
f(x) = 0 if x < 0
|
||||
|
||||
This is equivalent to `log(beta * max(0, x) + 1.0)`.
|
||||
|
||||
Problem Analysis:
|
||||
1. Memory Bound: As a point-wise activation, its performance is strictly limited by memory bandwidth.
|
||||
2. Operator Chaining: The PyTorch implementation `torch.log(beta * F.relu(x) + 1.0)` chains multiple kernels (`relu`, `mul`, `add`, `log`), creating high memory traffic.
|
||||
|
||||
Optimization Strategy: Fused Element-wise Kernel with Vectorization
|
||||
|
||||
1. One-Thread-per-Element: Map each element to a CUDA thread.
|
||||
|
||||
2. Vectorized Loads (float4): Use `float4` to process 128 bits per memory transaction to maximize throughput.
|
||||
|
||||
3. Fused Branching Logic:
|
||||
- For each element `x`, check `if (x >= 0)`.
|
||||
- If true, compute `__logf(beta * x + 1.0f)`.
|
||||
- If false, the result is `0.0f`.
|
||||
|
||||
4. One-Pass: Fuse all steps into a single read-compute-write 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 torch.nn.functional as F
|
||||
|
||||
BATCH_SIZE = 4096
|
||||
HIDDEN_DIM = 4096
|
||||
SHAPE = (BATCH_SIZE, HIDDEN_DIM)
|
||||
|
||||
# NLReLU 超参数 beta ，论文中建议范围 0.7-1.1
|
||||
BETA_VALUE = 1.0
|
||||
|
||||
class NLReLU(nn.Module):
|
||||
"""
|
||||
"Natural-Logarithm-Rectified Activation Function in Convolutional Neural Networks"
|
||||
Formula:
|
||||
f(x) = log(beta * x + 1.0) if x >= 0
|
||||
f(x) = 0 if x < 0
|
||||
"""
|
||||
def __init__(self, beta=1.0):
|
||||
super(NLReLU, self).__init__()
|
||||
self.beta = beta
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
x_relu = F.relu(x)
|
||||
inner = self.beta * x_relu + 1.0
|
||||
return torch.log(inner)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, beta=1.0):
|
||||
super(Model, self).__init__()
|
||||
self.act = NLReLU(beta=beta)
|
||||
|
||||
def forward(self, x):
|
||||
return self.act(x)
|
||||
|
||||
def get_inputs():
|
||||
input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0
|
||||
return [input_tensor.contiguous()]
|
||||
|
||||
def get_init_inputs():
|
||||
return [BETA_VALUE]
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from NLReLU_torch import Model,get_inputs,get_init_inputs
|
||||
from NLReLU_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