forked from ccf-ai-infra/GPUCodeForces
fixes MedianFilter2d #105
This commit is contained in:
parent
cc73715277
commit
e0cd2f0cb0
|
|
@ -0,0 +1,174 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
|
||||
from MedianFilter2d_torch import KERNEL_SIZE, N, C, H, W
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.k = KERNEL_SIZE
|
||||
self._compile_cuda_kernel()
|
||||
|
||||
def _compile_cuda_kernel(self):
|
||||
|
||||
BLOCK_W = 32
|
||||
BLOCK_H = 8
|
||||
PAD = self.k // 2
|
||||
|
||||
macros = f"""
|
||||
#define K {self.k}
|
||||
#define PAD {PAD}
|
||||
#define BLOCK_W {BLOCK_W}
|
||||
#define BLOCK_H {BLOCK_H}
|
||||
#define SMEM_W (BLOCK_W + 2 * PAD)
|
||||
#define SMEM_H (BLOCK_H + 2 * PAD)
|
||||
#define WINDOW_SIZE (K * K)
|
||||
#define MEDIAN_IDX (WINDOW_SIZE / 2)
|
||||
"""
|
||||
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
torch::Tensor median_filter_cuda(torch::Tensor input);
|
||||
"""
|
||||
|
||||
cuda_source = f"""
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
{macros}
|
||||
|
||||
|
||||
__device__ __forceinline__ void partial_sort(float* window) {{
|
||||
// Outer loop: only needs to run until we find the median element
|
||||
#pragma unroll
|
||||
for (int i = 0; i <= MEDIAN_IDX; ++i) {{
|
||||
// Find min in window[i...WINDOW_SIZE-1]
|
||||
#pragma unroll
|
||||
for (int j = i + 1; j < WINDOW_SIZE; ++j) {{
|
||||
float a = window[i];
|
||||
float b = window[j];
|
||||
// Swap if b is smaller, bubbling the min to position i
|
||||
if (b < a) {{
|
||||
window[i] = b;
|
||||
window[j] = a;
|
||||
}}
|
||||
}}
|
||||
}}
|
||||
}}
|
||||
|
||||
|
||||
__global__ void median_filter_kernel(
|
||||
const float* __restrict__ input,
|
||||
float* __restrict__ output,
|
||||
int height, int width
|
||||
) {{
|
||||
// Shared memory for tile + halo
|
||||
__shared__ float smem[SMEM_H][SMEM_W];
|
||||
|
||||
// 1. Coordinates
|
||||
int bx = blockIdx.x;
|
||||
int by = blockIdx.y;
|
||||
int bz = blockIdx.z; // Batch * Channel
|
||||
|
||||
int tx = threadIdx.x;
|
||||
int ty = threadIdx.y;
|
||||
|
||||
// Base coordinates in global memory (output pixel coords)
|
||||
int base_x = bx * BLOCK_W;
|
||||
int base_y = by * BLOCK_H;
|
||||
|
||||
// Plane offset for (n, c)
|
||||
int plane_offset = bz * height * width;
|
||||
const float* in_plane = input + plane_offset;
|
||||
float* out_plane = output + plane_offset;
|
||||
|
||||
// 2. Collaborative Loading (Global -> Shared)
|
||||
// Load a (BLOCK_H + 2*PAD) x (BLOCK_W + 2*PAD) block
|
||||
int tid = ty * BLOCK_W + tx;
|
||||
int num_threads = BLOCK_H * BLOCK_W;
|
||||
int num_smem_elements = SMEM_H * SMEM_W;
|
||||
|
||||
for (int i = tid; i < num_smem_elements; i += num_threads) {{
|
||||
int s_y = i / SMEM_W;
|
||||
int s_x = i % SMEM_W;
|
||||
|
||||
// Map shared memory coord to global coord (applying halo offset)
|
||||
int g_y = base_y + s_y - PAD;
|
||||
int g_x = base_x + s_x - PAD;
|
||||
|
||||
// Replicate Padding Logic: Clamp to border
|
||||
g_y = max(0, min(g_y, height - 1));
|
||||
g_x = max(0, min(g_x, width - 1));
|
||||
|
||||
smem[s_y][s_x] = in_plane[g_y * width + g_x];
|
||||
}}
|
||||
|
||||
__syncthreads();
|
||||
|
||||
// 3. Compute Median
|
||||
int out_x = base_x + tx;
|
||||
int out_y = base_y + ty;
|
||||
|
||||
// Only compute if valid output pixel
|
||||
if (out_x < width && out_y < height) {{
|
||||
float window[WINDOW_SIZE];
|
||||
|
||||
// Read from Shared Memory
|
||||
// Center of window in smem is at [ty + PAD][tx + PAD]
|
||||
int w_idx = 0;
|
||||
#pragma unroll
|
||||
for (int dy = 0; dy < K; ++dy) {{
|
||||
#pragma unroll
|
||||
for (int dx = 0; dx < K; ++dx) {{
|
||||
window[w_idx++] = smem[ty + dy][tx + dx];
|
||||
}}
|
||||
}}
|
||||
|
||||
// Sort partially to find median
|
||||
partial_sort(window);
|
||||
|
||||
// Write result
|
||||
out_plane[out_y * width + out_x] = window[MEDIAN_IDX];
|
||||
}}
|
||||
}}
|
||||
|
||||
torch::Tensor median_filter_cuda(torch::Tensor input) {{
|
||||
TORCH_CHECK(input.is_cuda(), "Input must be CUDA");
|
||||
TORCH_CHECK(input.dim() == 4, "Input must be (N, C, H, W)");
|
||||
|
||||
// Ensure contiguous memory for correct pointer arithmetic
|
||||
input = input.contiguous();
|
||||
|
||||
int N = input.size(0);
|
||||
int C = input.size(1);
|
||||
int H = input.size(2);
|
||||
int W = input.size(3);
|
||||
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
dim3 block(BLOCK_W, BLOCK_H);
|
||||
dim3 grid((W + BLOCK_W - 1) / BLOCK_W, (H + BLOCK_H - 1) / BLOCK_H, N * C);
|
||||
|
||||
median_filter_kernel<<<grid, block>>>(
|
||||
input.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
H, W
|
||||
);
|
||||
|
||||
return output;
|
||||
}}
|
||||
"""
|
||||
|
||||
self.op = load_inline(
|
||||
name='median_filter_opt_fixed_sort_v3',
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=['median_filter_cuda'],
|
||||
extra_cuda_cflags=['-O3', '--use_fast_math'],
|
||||
verbose=False
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
if not x.is_cuda: x = x.cuda()
|
||||
return self.op.median_filter_cuda(x)
|
||||
|
|
@ -0,0 +1,43 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
N, C, H, W = 4, 3, 512, 512
|
||||
KERNEL_SIZE = 3
|
||||
|
||||
class MedianFilter2d(nn.Module):
|
||||
def __init__(self, kernel_size=3):
|
||||
super().__init__()
|
||||
self.k = kernel_size
|
||||
self.pad = kernel_size // 2
|
||||
|
||||
def forward(self, x):
|
||||
|
||||
N, C, H, W = x.shape
|
||||
|
||||
x_pad = F.pad(x, (self.pad, self.pad, self.pad, self.pad), mode='replicate')
|
||||
|
||||
|
||||
patches = F.unfold(x_pad, kernel_size=self.k)
|
||||
|
||||
|
||||
patches = patches.view(N, C, self.k * self.k, -1)
|
||||
|
||||
median_val, _ = torch.median(patches, dim=2)
|
||||
|
||||
return median_val.view(N, C, H, W)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.op = MedianFilter2d(kernel_size=KERNEL_SIZE)
|
||||
|
||||
def forward(self, x):
|
||||
return self.op(x)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,51 @@
|
|||
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
|
||||
|
||||
N, C, H, W = 4, 3, 512, 512
|
||||
KERNEL_SIZE = 3
|
||||
|
||||
class MedianFilter2d(nn.Module):
|
||||
def __init__(self, kernel_size=3):
|
||||
super().__init__()
|
||||
self.k = kernel_size
|
||||
self.pad = kernel_size // 2
|
||||
|
||||
def forward(self, x):
|
||||
|
||||
N, C, H, W = x.shape
|
||||
|
||||
x_pad = F.pad(x, (self.pad, self.pad, self.pad, self.pad), mode='replicate')
|
||||
|
||||
|
||||
patches = F.unfold(x_pad, kernel_size=self.k)
|
||||
|
||||
|
||||
patches = patches.view(N, C, self.k * self.k, -1)
|
||||
|
||||
median_val, _ = torch.median(patches, dim=2)
|
||||
|
||||
return median_val.view(N, C, H, W)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.op = MedianFilter2d(kernel_size=KERNEL_SIZE)
|
||||
|
||||
def forward(self, x):
|
||||
return self.op(x)
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(N, C, H, W, dtype=torch.float32)
|
||||
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 MedianFilter2d_torch import Model,get_inputs,get_init_inputs
|
||||
from MedianFilter2d_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