forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish localresponsenorm Operator #8' (#28) from ZZZJ/GPUCodeForces:localresponsenorm into main
This commit is contained in:
commit
5f3df5b940
|
|
@ -0,0 +1,98 @@
|
|||
# localresponsenorm_cuda.py
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
from localresponsenorm_torch import N, C, H, W, SIZE, ALPHA, BETA, K
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self._compile_cuda_kernel()
|
||||
|
||||
def _compile_cuda_kernel(self):
|
||||
half_size = SIZE // 2
|
||||
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
torch::Tensor lrn_forward_cuda(torch::Tensor input, int size, int half_size, float alpha, float beta, float k);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <cuda_runtime.h>
|
||||
#include <device_launch_parameters.h>
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
|
||||
__global__ void lrn_fused_kernel(
|
||||
const float* __restrict__ x,
|
||||
float* __restrict__ y,
|
||||
int N, int C, int H, int W,
|
||||
int size, int half_size, float alpha, float beta, float k
|
||||
) {
|
||||
int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
if (idx >= N * C * H * W) return;
|
||||
|
||||
int n = idx / (C * H * W);
|
||||
int rem = idx % (C * H * W);
|
||||
int c = rem / (H * W);
|
||||
int hw = rem % (H * W);
|
||||
|
||||
int base_idx = n * C * H * W + hw;
|
||||
float sum_sq = 0.0f;
|
||||
|
||||
int start_c = max(0, c - half_size);
|
||||
int end_c = min(C - 1, c + half_size);
|
||||
|
||||
#pragma unroll 4
|
||||
for (int window_c = start_c; window_c <= end_c; ++window_c) {
|
||||
float val = x[base_idx + window_c * H * W];
|
||||
sum_sq = __fmaf_rn(val, val, sum_sq);
|
||||
}
|
||||
|
||||
float alpha_over_n = alpha / (float)size;
|
||||
float scale = __fmaf_rn(alpha_over_n, sum_sq, k);
|
||||
|
||||
|
||||
float norm_factor = __powf(scale, beta);
|
||||
float input_val = x[idx];
|
||||
y[idx] = __fdividef(input_val, norm_factor);
|
||||
}
|
||||
|
||||
torch::Tensor lrn_forward_cuda(torch::Tensor input, int size, int half_size, float alpha, float beta, float k) {
|
||||
const int N = input.size(0);
|
||||
const int C = input.size(1);
|
||||
const int H = input.size(2);
|
||||
const int W = input.size(3);
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
const int n_elements = N * C * H * W;
|
||||
const int blocks = (n_elements + BLOCK_SIZE - 1) / BLOCK_SIZE;
|
||||
|
||||
lrn_fused_kernel<<<blocks, BLOCK_SIZE>>>(
|
||||
input.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
N, C, H, W,
|
||||
size, half_size, alpha, beta, k
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
self.lrn_op = load_inline(
|
||||
name="lrn_correct_formula",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=["lrn_forward_cuda"],
|
||||
extra_cuda_cflags=[
|
||||
"-O3",
|
||||
"--use_fast_math",
|
||||
"--fmad=true"
|
||||
],
|
||||
verbose=True
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
half_size = SIZE // 2
|
||||
return self.lrn_op.lrn_forward_cuda(x.contiguous(), SIZE, half_size, ALPHA, BETA, K)
|
||||
|
|
@ -0,0 +1,27 @@
|
|||
# localresponsenorm_torch.py
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
|
||||
N, C, H, W = 16, 64, 32, 32
|
||||
|
||||
SIZE = 5 # 窗口大小 (n)
|
||||
ALPHA = 1e-4 # 缩放因子 (alpha)
|
||||
BETA = 0.75 # 幂指数 (beta)
|
||||
K = 2.0 # 偏置项 (k)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.lrn = nn.LocalResponseNorm(SIZE, alpha=ALPHA, beta=BETA, k=K)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.lrn(x)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,34 @@
|
|||
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:
|
||||
|
||||
```python
|
||||
# localresponsenorm_torch.py
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
|
||||
N, C, H, W = 16, 64, 32, 32
|
||||
|
||||
SIZE = 5 # 窗口大小 (n)
|
||||
ALPHA = 1e-4 # 缩放因子 (alpha)
|
||||
BETA = 0.75 # 幂指数 (beta)
|
||||
K = 2.0 # 偏置项 (k)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.lrn = nn.LocalResponseNorm(SIZE, alpha=ALPHA, beta=BETA, k=K)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.lrn(x)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,88 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from localresponsenorm_torch import Model, get_inputs, get_init_inputs
|
||||
from localresponsenorm_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 (matmul + relu) 平均执行时间: {torch_time:.6f} 秒")
|
||||
print(f"自定义 CUDA ReLU 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