forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish RoiPool #155' (#815) from ZZZJ/GPUCodeForces:RoiPool into main
This commit is contained in:
commit
6d7e6dc429
|
|
@ -0,0 +1,49 @@
|
|||
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
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torchvision
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, output_size=(7, 7), spatial_scale=1.0):
|
||||
super().__init__()
|
||||
self.output_size = output_size
|
||||
self.spatial_scale = spatial_scale
|
||||
|
||||
def forward(self, input, rois):
|
||||
return torchvision.ops.roi_pool(
|
||||
input, rois,
|
||||
output_size=self.output_size,
|
||||
spatial_scale=self.spatial_scale
|
||||
)
|
||||
|
||||
N = 4
|
||||
C = 256
|
||||
H = 128
|
||||
W = 128
|
||||
K = 1000
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
rois = torch.zeros(K, 5, dtype=torch.float32)
|
||||
rois[:, 0] = torch.randint(0, N, (K,)).float()
|
||||
|
||||
x1 = torch.rand(K) * (W // 2)
|
||||
y1 = torch.rand(K) * (H // 2)
|
||||
x2 = x1 + torch.rand(K) * (W // 2) + 2.0
|
||||
y2 = y1 + torch.rand(K) * (H // 2) + 2.0
|
||||
|
||||
rois[:, 1] = x1
|
||||
rois[:, 2] = y1
|
||||
rois[:, 3] = x2
|
||||
rois[:, 4] = y2
|
||||
|
||||
return [x, rois]
|
||||
|
||||
def get_init_inputs():
|
||||
return [(7, 7), 1.0]
|
||||
|
|
@ -0,0 +1,132 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
cpp_src = """
|
||||
torch::Tensor roi_pool_cuda(torch::Tensor input, torch::Tensor rois, double spatial_scale, int pooled_height, int pooled_width);
|
||||
"""
|
||||
|
||||
cuda_src = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <cfloat>
|
||||
|
||||
__global__ void roi_pool_forward_kernel(
|
||||
const int nthreads,
|
||||
const float* input,
|
||||
const float* rois,
|
||||
float* output,
|
||||
const float spatial_scale,
|
||||
const int channels,
|
||||
const int height,
|
||||
const int width,
|
||||
const int pooled_height,
|
||||
const int pooled_width) {
|
||||
|
||||
int index = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
if (index >= nthreads) return;
|
||||
|
||||
int pw = index % pooled_width;
|
||||
int ph = (index / pooled_width) % pooled_height;
|
||||
int c = (index / pooled_width / pooled_height) % channels;
|
||||
int n = index / pooled_width / pooled_height / channels;
|
||||
|
||||
const float* offset_rois = rois + n * 5;
|
||||
int roi_batch_ind = offset_rois[0];
|
||||
|
||||
int roi_start_w = round(offset_rois[1] * spatial_scale);
|
||||
int roi_start_h = round(offset_rois[2] * spatial_scale);
|
||||
int roi_end_w = round(offset_rois[3] * spatial_scale);
|
||||
int roi_end_h = round(offset_rois[4] * spatial_scale);
|
||||
|
||||
int roi_width = max(roi_end_w - roi_start_w + 1, 1);
|
||||
int roi_height = max(roi_end_h - roi_start_h + 1, 1);
|
||||
|
||||
const float bin_size_h = (float)roi_height / (float)pooled_height;
|
||||
const float bin_size_w = (float)roi_width / (float)pooled_width;
|
||||
|
||||
int hstart = (int)(floor((float)(ph) * bin_size_h));
|
||||
int wstart = (int)(floor((float)(pw) * bin_size_w));
|
||||
int hend = (int)(ceil((float)(ph + 1) * bin_size_h));
|
||||
int wend = (int)(ceil((float)(pw + 1) * bin_size_w));
|
||||
|
||||
hstart = min(max(hstart + roi_start_h, 0), height);
|
||||
hend = min(max(hend + roi_start_h, 0), height);
|
||||
wstart = min(max(wstart + roi_start_w, 0), width);
|
||||
wend = min(max(wend + roi_start_w, 0), width);
|
||||
|
||||
bool is_empty = (hend <= hstart) || (wend <= wstart);
|
||||
|
||||
const float* offset_input = input + (roi_batch_ind * channels + c) * height * width;
|
||||
|
||||
float max_val = is_empty ? 0 : -FLT_MAX;
|
||||
|
||||
for (int h = hstart; h < hend; ++h) {
|
||||
for (int w = wstart; w < wend; ++w) {
|
||||
float val = offset_input[h * width + w];
|
||||
if (val > max_val) {
|
||||
max_val = val;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
output[index] = max_val;
|
||||
}
|
||||
|
||||
torch::Tensor roi_pool_cuda(torch::Tensor input, torch::Tensor rois, double spatial_scale, int pooled_height, int pooled_width) {
|
||||
int num_rois = rois.size(0);
|
||||
int channels = input.size(1);
|
||||
int height = input.size(2);
|
||||
int width = input.size(3);
|
||||
|
||||
auto output = torch::zeros({num_rois, channels, pooled_height, pooled_width}, input.options());
|
||||
|
||||
int output_size = num_rois * channels * pooled_height * pooled_width;
|
||||
|
||||
input = input.contiguous();
|
||||
rois = rois.contiguous();
|
||||
|
||||
const int block_size = 512;
|
||||
int grid_size = (output_size + block_size - 1) / block_size;
|
||||
if (grid_size > 2147483647) grid_size = 2147483647;
|
||||
|
||||
roi_pool_forward_kernel<<<grid_size, block_size>>>(
|
||||
output_size,
|
||||
input.data_ptr<float>(),
|
||||
rois.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
(float)spatial_scale,
|
||||
channels,
|
||||
height,
|
||||
width,
|
||||
pooled_height,
|
||||
pooled_width
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, output_size=(7, 7), spatial_scale=1.0):
|
||||
super().__init__()
|
||||
if isinstance(output_size, int):
|
||||
self.output_size = (output_size, output_size)
|
||||
else:
|
||||
self.output_size = output_size
|
||||
self.spatial_scale = spatial_scale
|
||||
|
||||
self.module = load_inline(
|
||||
name="roi_pool_opt",
|
||||
cpp_sources=cpp_src,
|
||||
cuda_sources=cuda_src,
|
||||
functions=["roi_pool_cuda"],
|
||||
verbose=False,
|
||||
extra_cuda_cflags=["-O3"]
|
||||
)
|
||||
|
||||
def forward(self, input, rois):
|
||||
return self.module.roi_pool_cuda(
|
||||
input, rois, self.spatial_scale,
|
||||
self.output_size[0], self.output_size[1]
|
||||
)
|
||||
|
|
@ -0,0 +1,42 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torchvision
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, output_size=(7, 7), spatial_scale=1.0):
|
||||
super().__init__()
|
||||
self.output_size = output_size
|
||||
self.spatial_scale = spatial_scale
|
||||
|
||||
def forward(self, input, rois):
|
||||
return torchvision.ops.roi_pool(
|
||||
input, rois,
|
||||
output_size=self.output_size,
|
||||
spatial_scale=self.spatial_scale
|
||||
)
|
||||
|
||||
N = 4
|
||||
C = 256
|
||||
H = 128
|
||||
W = 128
|
||||
K = 1000
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
rois = torch.zeros(K, 5, dtype=torch.float32)
|
||||
rois[:, 0] = torch.randint(0, N, (K,)).float()
|
||||
|
||||
x1 = torch.rand(K) * (W // 2)
|
||||
y1 = torch.rand(K) * (H // 2)
|
||||
x2 = x1 + torch.rand(K) * (W // 2) + 2.0
|
||||
y2 = y1 + torch.rand(K) * (H // 2) + 2.0
|
||||
|
||||
rois[:, 1] = x1
|
||||
rois[:, 2] = y1
|
||||
rois[:, 3] = x2
|
||||
rois[:, 4] = y2
|
||||
|
||||
return [x, rois]
|
||||
|
||||
def get_init_inputs():
|
||||
return [(7, 7), 1.0]
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from roi_pool_torch import Model,get_inputs,get_init_inputs
|
||||
from roi_pool_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