forked from ccf-ai-infra/GPUCodeForces
fixes SobelFilter3d #91
This commit is contained in:
parent
cc73715277
commit
55b341fdae
|
|
@ -0,0 +1,58 @@
|
|||
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 torch.nn.functional as F
|
||||
|
||||
torch.backends.cuda.matmul.allow_tf32 = False
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super(Model, self).__init__()
|
||||
|
||||
smooth = torch.tensor([1., 2., 1.], dtype=torch.float32)
|
||||
diff = torch.tensor([-1., 0., 1.], dtype=torch.float32)
|
||||
|
||||
k_x = (smooth.view(3, 1, 1) * smooth.view(1, 3, 1) * diff.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
k_y = (smooth.view(3, 1, 1) * diff.view(1, 3, 1) * smooth.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
k_z = (diff.view(3, 1, 1) * smooth.view(1, 3, 1) * smooth.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
self.register_buffer('k_x', k_x)
|
||||
self.register_buffer('k_y', k_y)
|
||||
self.register_buffer('k_z', k_z)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
N, C, D, H, W = x.shape
|
||||
|
||||
weight_x = self.k_x.expand(C, 1, 3, 3, 3)
|
||||
weight_y = self.k_y.expand(C, 1, 3, 3, 3)
|
||||
weight_z = self.k_z.expand(C, 1, 3, 3, 3)
|
||||
|
||||
gx = F.conv3d(x, weight_x, padding=1, groups=C)
|
||||
gy = F.conv3d(x, weight_y, padding=1, groups=C)
|
||||
gz = F.conv3d(x, weight_z, padding=1, groups=C)
|
||||
|
||||
return gx*gx + gy*gy + gz*gz
|
||||
|
||||
|
||||
N = 4
|
||||
C = 4
|
||||
D = 64
|
||||
H = 128
|
||||
W = 128
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randint(0, 10, (N, C, D, H, W), device='cuda').float()
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
```
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from sobel_filter_3d_torch import Model,get_inputs,get_init_inputs
|
||||
from sobel_filter_3d_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()
|
||||
|
|
@ -0,0 +1,132 @@
|
|||
import torch
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
sobel_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
|
||||
__global__ void sobel_3d_kernel(
|
||||
const float* __restrict__ input,
|
||||
float* __restrict__ output,
|
||||
int D, int H, int W,
|
||||
long spatial_size, // H * W
|
||||
long volume_size // D * H * W
|
||||
) {
|
||||
|
||||
|
||||
int nc_idx = blockIdx.z; // Index of the (n, c) volume
|
||||
int d = blockIdx.y;
|
||||
int spatial_idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
|
||||
if (spatial_idx < spatial_size) {
|
||||
int h = spatial_idx / W;
|
||||
int w = spatial_idx % W;
|
||||
|
||||
const float* vol_in = input + nc_idx * volume_size;
|
||||
float* vol_out = output + nc_idx * volume_size;
|
||||
|
||||
float gx = 0.0f;
|
||||
float gy = 0.0f;
|
||||
float gz = 0.0f;
|
||||
|
||||
|
||||
#pragma unroll
|
||||
for (int kz = -1; kz <= 1; ++kz) {
|
||||
int in_d = d + kz;
|
||||
float wz_s = (kz == 0) ? 2.0f : 1.0f; // Smooth weight
|
||||
float wz_d = (float)kz; // Diff weight (-1, 0, 1)
|
||||
|
||||
// Y Loop
|
||||
#pragma unroll
|
||||
for (int ky = -1; ky <= 1; ++ky) {
|
||||
int in_h = h + ky;
|
||||
float wy_s = (ky == 0) ? 2.0f : 1.0f;
|
||||
float wy_d = (float)ky;
|
||||
|
||||
// X Loop
|
||||
#pragma unroll
|
||||
for (int kx = -1; kx <= 1; ++kx) {
|
||||
int in_w = w + kx;
|
||||
float wx_s = (kx == 0) ? 2.0f : 1.0f;
|
||||
float wx_d = (float)kx;
|
||||
|
||||
// Boundary Check (Zero Padding)
|
||||
float val = 0.0f;
|
||||
if (in_d >= 0 && in_d < D &&
|
||||
in_h >= 0 && in_h < H &&
|
||||
in_w >= 0 && in_w < W)
|
||||
{
|
||||
// Calculate offset manually to avoid multiplication if possible
|
||||
// But here stride is necessary
|
||||
long idx = (long)in_d * spatial_size + (long)in_h * W + in_w;
|
||||
val = __ldg(&vol_in[idx]);
|
||||
}
|
||||
|
||||
// Accumulate Gradients
|
||||
// Gx: Smooth(z) * Smooth(y) * Diff(x)
|
||||
gx += val * (wz_s * wy_s * wx_d);
|
||||
|
||||
// Gy: Smooth(z) * Diff(y) * Smooth(x)
|
||||
gy += val * (wz_s * wy_d * wx_s);
|
||||
|
||||
// Gz: Diff(z) * Smooth(y) * Smooth(x)
|
||||
gz += val * (wz_d * wy_s * wx_s);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Write Result: Squared Magnitude
|
||||
long out_idx = (long)d * spatial_size + spatial_idx;
|
||||
vol_out[out_idx] = gx * gx + gy * gy + gz * gz;
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor sobel_3d_cuda(torch::Tensor input) {
|
||||
int N = input.size(0);
|
||||
int C = input.size(1);
|
||||
int D = input.size(2);
|
||||
int H = input.size(3);
|
||||
int W = input.size(4);
|
||||
|
||||
// Output shape same as input
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
long spatial_size = H * W;
|
||||
long volume_size = D * spatial_size;
|
||||
int nc = N * C;
|
||||
|
||||
// Grid Config
|
||||
const int block = 256;
|
||||
dim3 grid((spatial_size + block - 1) / block, D, nc);
|
||||
|
||||
sobel_3d_kernel<<<grid, block>>>(
|
||||
input.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
D, H, W,
|
||||
spatial_size,
|
||||
volume_size
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
cpp_source = "torch::Tensor sobel_3d_cuda(torch::Tensor input);"
|
||||
|
||||
sobel_module = load_inline(
|
||||
name="sobel_filter_3d_extension",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=sobel_source,
|
||||
functions=["sobel_3d_cuda"],
|
||||
verbose=True,
|
||||
with_cuda=True
|
||||
)
|
||||
|
||||
class ModelNew(torch.nn.Module):
|
||||
def __init__(self):
|
||||
super(ModelNew, self).__init__()
|
||||
self.cuda_op = sobel_module
|
||||
|
||||
def forward(self, x):
|
||||
return self.cuda_op.sobel_3d_cuda(x.contiguous())
|
||||
|
|
@ -0,0 +1,50 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
torch.backends.cuda.matmul.allow_tf32 = False
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super(Model, self).__init__()
|
||||
|
||||
smooth = torch.tensor([1., 2., 1.], dtype=torch.float32)
|
||||
diff = torch.tensor([-1., 0., 1.], dtype=torch.float32)
|
||||
|
||||
k_x = (smooth.view(3, 1, 1) * smooth.view(1, 3, 1) * diff.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
k_y = (smooth.view(3, 1, 1) * diff.view(1, 3, 1) * smooth.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
k_z = (diff.view(3, 1, 1) * smooth.view(1, 3, 1) * smooth.view(1, 1, 3)).unsqueeze(0).unsqueeze(0)
|
||||
|
||||
self.register_buffer('k_x', k_x)
|
||||
self.register_buffer('k_y', k_y)
|
||||
self.register_buffer('k_z', k_z)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
N, C, D, H, W = x.shape
|
||||
|
||||
weight_x = self.k_x.expand(C, 1, 3, 3, 3)
|
||||
weight_y = self.k_y.expand(C, 1, 3, 3, 3)
|
||||
weight_z = self.k_z.expand(C, 1, 3, 3, 3)
|
||||
|
||||
gx = F.conv3d(x, weight_x, padding=1, groups=C)
|
||||
gy = F.conv3d(x, weight_y, padding=1, groups=C)
|
||||
gz = F.conv3d(x, weight_z, padding=1, groups=C)
|
||||
|
||||
return gx*gx + gy*gy + gz*gz
|
||||
|
||||
|
||||
N = 4
|
||||
C = 4
|
||||
D = 64
|
||||
H = 128
|
||||
W = 128
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randint(0, 10, (N, C, D, H, W), device='cuda').float()
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
Loading…
Reference in New Issue