forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish GlobalResponseNormalization #119' (#767) from ZZZJ/GPUCodeForces:GlobalResponseNormalization into main
This commit is contained in:
commit
d54e7dec01
|
|
@ -0,0 +1,184 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.gamma = nn.Parameter(torch.zeros(384, device='cuda'))
|
||||
self.beta = nn.Parameter(torch.zeros(384, device='cuda'))
|
||||
self._compile_cuda_kernel()
|
||||
|
||||
def _compile_cuda_kernel(self):
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
torch::Tensor grn_cuda(torch::Tensor input, torch::Tensor gamma, torch::Tensor beta, float eps);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
|
||||
__device__ __forceinline__ double warpReduceSum(double val) {
|
||||
#pragma unroll
|
||||
for (int offset = 16; offset > 0; offset /= 2)
|
||||
val += __shfl_down_sync(0xffffffff, val, offset);
|
||||
return val;
|
||||
}
|
||||
|
||||
|
||||
__device__ __forceinline__ double blockReduceSum(double val) {
|
||||
static __shared__ double shared[32];
|
||||
int lane = threadIdx.x % 32;
|
||||
int wid = threadIdx.x / 32;
|
||||
val = warpReduceSum(val);
|
||||
if (lane == 0) shared[wid] = val;
|
||||
__syncthreads();
|
||||
val = (threadIdx.x < (BLOCK_SIZE / 32)) ? shared[lane] : 0.0;
|
||||
if (wid == 0) val = warpReduceSum(val);
|
||||
return val;
|
||||
}
|
||||
|
||||
|
||||
__global__ void grn_stats_kernel(
|
||||
const float* __restrict__ input,
|
||||
double* __restrict__ norms, // Output: [N, C] stored in double
|
||||
int spatial_size, // H * W
|
||||
int n_vec_spatial // H * W / 4
|
||||
) {
|
||||
int nc = blockIdx.x; // Batch * Channel index
|
||||
int tid = threadIdx.x;
|
||||
|
||||
// Input pointer for this channel
|
||||
const float* in_ptr = input + (long long)nc * spatial_size;
|
||||
|
||||
double sum_sq = 0.0;
|
||||
|
||||
// Grid-Stride Loop over Spatial (Float4)
|
||||
for (int i = tid; i < n_vec_spatial; i += BLOCK_SIZE) {
|
||||
float4 v = reinterpret_cast<const float4*>(in_ptr)[i];
|
||||
sum_sq += (double)v.x * v.x + (double)v.y * v.y +
|
||||
(double)v.z * v.z + (double)v.w * v.w;
|
||||
}
|
||||
|
||||
// Reduction
|
||||
sum_sq = blockReduceSum(sum_sq);
|
||||
|
||||
if (tid == 0) {
|
||||
norms[nc] = sqrt(sum_sq);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
__global__ void grn_apply_kernel(
|
||||
const float* __restrict__ input,
|
||||
const double* __restrict__ norms, // [N, C]
|
||||
const float* __restrict__ gamma, // [C]
|
||||
const float* __restrict__ beta, // [C]
|
||||
float* __restrict__ output,
|
||||
int channels,
|
||||
int spatial_size,
|
||||
int n_vec_spatial,
|
||||
float eps
|
||||
) {
|
||||
int n = blockIdx.x;
|
||||
int c = blockIdx.y;
|
||||
int tid = threadIdx.x;
|
||||
|
||||
const double* norms_n = norms + n * channels;
|
||||
double local_norm_sum = 0.0;
|
||||
for (int i = tid; i < channels; i += BLOCK_SIZE) {
|
||||
local_norm_sum += norms_n[i];
|
||||
}
|
||||
|
||||
|
||||
double total_norm_sum = blockReduceSum(local_norm_sum);
|
||||
|
||||
__shared__ float s_nx; // Scaling factor for this channel
|
||||
|
||||
if (tid == 0) {
|
||||
double mean_norm = total_norm_sum / channels;
|
||||
double my_norm = norms_n[c];
|
||||
// nx = gx / (mean + eps)
|
||||
s_nx = (float)(my_norm / (mean_norm + (double)eps));
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
float nx = s_nx;
|
||||
|
||||
float g = gamma[c];
|
||||
float b = beta[c];
|
||||
|
||||
long long offset = (long long)n * (channels * spatial_size) + c * spatial_size;
|
||||
const float* in_ptr = input + offset;
|
||||
float* out_ptr = output + offset;
|
||||
|
||||
for (int i = tid; i < n_vec_spatial; i += BLOCK_SIZE) {
|
||||
float4 v = reinterpret_cast<const float4*>(in_ptr)[i];
|
||||
float4 out_v;
|
||||
|
||||
// Fused computation
|
||||
out_v.x = g * (v.x * nx) + b + v.x;
|
||||
out_v.y = g * (v.y * nx) + b + v.y;
|
||||
out_v.z = g * (v.z * nx) + b + v.z;
|
||||
out_v.w = g * (v.w * nx) + b + v.w;
|
||||
|
||||
reinterpret_cast<float4*>(out_ptr)[i] = out_v;
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor grn_cuda(torch::Tensor input, torch::Tensor gamma, torch::Tensor beta, float eps) {
|
||||
int batch = input.size(0);
|
||||
int channels = input.size(1);
|
||||
int height = input.size(2);
|
||||
int width = input.size(3);
|
||||
|
||||
int spatial = height * width;
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
if (spatial % 4 != 0) return output;
|
||||
int n_vec_spatial = spatial / 4;
|
||||
|
||||
auto norms = torch::empty({batch, channels}, input.options().dtype(torch::kFloat64));
|
||||
|
||||
dim3 grid_stats(batch * channels);
|
||||
dim3 block(BLOCK_SIZE);
|
||||
|
||||
grn_stats_kernel<<<grid_stats, block>>>(
|
||||
input.data_ptr<float>(),
|
||||
norms.data_ptr<double>(),
|
||||
spatial,
|
||||
n_vec_spatial
|
||||
);
|
||||
|
||||
dim3 grid_apply(batch, channels);
|
||||
|
||||
grn_apply_kernel<<<grid_apply, block>>>(
|
||||
input.data_ptr<float>(),
|
||||
norms.data_ptr<double>(),
|
||||
gamma.data_ptr<float>(),
|
||||
beta.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
channels,
|
||||
spatial,
|
||||
n_vec_spatial,
|
||||
eps
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
self.op = load_inline(
|
||||
name="grn_cuda_opt_v2",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=["grn_cuda"],
|
||||
extra_cuda_cflags=["-O3"],
|
||||
verbose=False
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
if not x.is_contiguous(): x = x.contiguous()
|
||||
return self.op.grn_cuda(x, self.gamma, self.beta, 1e-6)
|
||||
|
|
@ -0,0 +1,31 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
BATCH = 32
|
||||
CHANNELS = 384
|
||||
HEIGHT = 56
|
||||
WIDTH = 56
|
||||
EPS = 1e-6
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.gamma = nn.Parameter(torch.zeros(CHANNELS))
|
||||
self.beta = nn.Parameter(torch.zeros(CHANNELS))
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
gx = torch.norm(x, p=2, dim=(2, 3), keepdim=True)
|
||||
|
||||
|
||||
nx = gx / (gx.mean(dim=1, keepdim=True) + EPS)
|
||||
|
||||
return self.gamma.view(1, -1, 1, 1) * (x * nx) + self.beta.view(1, -1, 1, 1) + x
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randint(low=-2, high=3, size=(BATCH, CHANNELS, HEIGHT, WIDTH), device='cuda').float()
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
return []
|
||||
|
|
@ -0,0 +1,39 @@
|
|||
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
|
||||
|
||||
|
||||
BATCH = 32
|
||||
CHANNELS = 384
|
||||
HEIGHT = 56
|
||||
WIDTH = 56
|
||||
EPS = 1e-6
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.gamma = nn.Parameter(torch.zeros(CHANNELS))
|
||||
self.beta = nn.Parameter(torch.zeros(CHANNELS))
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
gx = torch.norm(x, p=2, dim=(2, 3), keepdim=True)
|
||||
|
||||
|
||||
nx = gx / (gx.mean(dim=1, keepdim=True) + EPS)
|
||||
|
||||
return self.gamma.view(1, -1, 1, 1) * (x * nx) + self.beta.view(1, -1, 1, 1) + x
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randint(low=-2, high=3, size=(BATCH, CHANNELS, HEIGHT, WIDTH), 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 global_response_normalization_torch import Model,get_inputs,get_init_inputs
|
||||
from global_response_normalization_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