forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish ISRLU #49' (#300) from gsd123/GPUCodeForces:gsd49 into main
This commit is contained in:
commit
84da765424
|
|
@ -0,0 +1,91 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, a=1.0):
|
||||
super().__init__()
|
||||
self.a = a
|
||||
self._compile_cuda_kernel()
|
||||
|
||||
def _compile_cuda_kernel(self):
|
||||
cpp_source = """
|
||||
torch::Tensor isrlu_cuda(torch::Tensor x, float a);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <math.h>
|
||||
|
||||
__device__ __forceinline__ float isrlu_op(float x, float a) {
|
||||
// ISRLU(x) = x if x >= 0, else x / sqrt(1 + a * x^2)
|
||||
if (x >= 0.0f) {
|
||||
return x;
|
||||
}
|
||||
float denom = sqrtf(1.0f + a * x * x);
|
||||
return x / denom;
|
||||
}
|
||||
|
||||
__global__ void isrlu_kernel(
|
||||
const float* __restrict__ x,
|
||||
float* __restrict__ output,
|
||||
const int n_elements,
|
||||
const float a)
|
||||
{
|
||||
const int tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
const int stride = blockDim.x * gridDim.x;
|
||||
|
||||
const int vec_loops = n_elements >> 2;
|
||||
const float4* x_vec = reinterpret_cast<const float4*>(x);
|
||||
float4* out_vec = reinterpret_cast<float4*>(output);
|
||||
|
||||
for (int i = tid; i < vec_loops; i += stride) {
|
||||
float4 v = __ldg(&x_vec[i]);
|
||||
float4 r;
|
||||
|
||||
r.x = isrlu_op(v.x, a);
|
||||
r.y = isrlu_op(v.y, a);
|
||||
r.z = isrlu_op(v.z, a);
|
||||
r.w = isrlu_op(v.w, a);
|
||||
|
||||
out_vec[i] = r;
|
||||
}
|
||||
|
||||
const int tail_start = vec_loops << 2;
|
||||
for (int i = tail_start + tid; i < n_elements; i += stride) {
|
||||
output[i] = isrlu_op(x[i], a);
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor isrlu_cuda(torch::Tensor x, float a) {
|
||||
auto x_c = x.contiguous();
|
||||
const int n_elements = x_c.numel();
|
||||
auto output = torch::empty_like(x_c);
|
||||
|
||||
const int threads = 256;
|
||||
const int max_blocks = 65535;
|
||||
const int blocks = std::min((n_elements + threads * 4 - 1) / (threads * 4), max_blocks);
|
||||
|
||||
isrlu_kernel<<<blocks, threads>>>(
|
||||
x_c.data_ptr<float>(),
|
||||
output.data_ptr<float>(),
|
||||
n_elements,
|
||||
a
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
self.op = load_inline(
|
||||
name="isrlu_v1",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=["isrlu_cuda"],
|
||||
extra_cuda_cflags=["-O3", "--use_fast_math"],
|
||||
verbose=False
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
return self.op.isrlu_cuda(x, self.a)
|
||||
|
|
@ -0,0 +1,27 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, a=1.0):
|
||||
super().__init__()
|
||||
self.a = a
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
# ISRLU Formula: x if x >= 0, else x / sqrt(1 + a * x^2)
|
||||
y_neg = x / torch.sqrt(1.0 + self.a * x.pow(2))
|
||||
|
||||
return torch.where(x >= 0, x, y_neg)
|
||||
|
||||
|
||||
batch_size = 128
|
||||
feature_dim = 512
|
||||
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
|
||||
return [x]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return [1.0]
|
||||
|
|
@ -0,0 +1,90 @@
|
|||
You write custom CUDA kernels to replace the pytorch operators in the given GeGLU 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 chunk+gelu+elementwise_mul), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.
|
||||
|
||||
CUDA Optimization Strategies:
|
||||
|
||||
Vectorized Memory Access
|
||||
|
||||
Uses float4 for 4-element vector loads/stores
|
||||
|
||||
__ldg() for read-only caching through texture memory
|
||||
|
||||
Bit shifts for division (>> 2, << 2) for efficiency
|
||||
|
||||
ISRLU (Inverse Square Root Linear Unit) Function
|
||||
|
||||
Piecewise definition:
|
||||
|
||||
x ≥ 0: x (ReLU-like)
|
||||
|
||||
x < 0: x / √(1 + a * x²) (ISRU-like)
|
||||
|
||||
Combines ReLU for positives with ISRU for negatives
|
||||
|
||||
Parameter a controls negative saturation
|
||||
|
||||
Optimized Branching
|
||||
|
||||
Simple if (x >= 0) condition
|
||||
|
||||
Early return for positive branch
|
||||
|
||||
Negative branch uses ISRU computation
|
||||
|
||||
Memory Access
|
||||
|
||||
contiguous() tensors for coalescing
|
||||
|
||||
__restrict__ pointers
|
||||
|
||||
Grid-stride loop for arbitrary sizes
|
||||
|
||||
Performance Optimization
|
||||
|
||||
Compiler flags: -O3, --use_fast_math
|
||||
|
||||
Efficient kernel launch configuration
|
||||
|
||||
Block count limited to 65535
|
||||
|
||||
Inline function for ISRLU computation
|
||||
|
||||
Numerical Properties
|
||||
|
||||
Non-negative for x ≥ 0
|
||||
|
||||
Smoothly saturates for negative inputs
|
||||
|
||||
Differentiable everywhere except origin
|
||||
|
||||
Key Innovation: Vectorized Inverse Square Root Linear Unit activation combining ReLU's simplicity for positives with ISRU's smooth saturation for negatives, optimized with efficient branching and mathematical computation.
|
||||
|
||||
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, a=1.0):
|
||||
super().__init__()
|
||||
self.a = a
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
# ISRLU Formula: x if x >= 0, else x / sqrt(1 + a * x^2)
|
||||
y_neg = x / torch.sqrt(1.0 + self.a * x.pow(2))
|
||||
|
||||
return torch.where(x >= 0, x, y_neg)
|
||||
|
||||
|
||||
batch_size = 128
|
||||
feature_dim = 512
|
||||
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
|
||||
return [x]
|
||||
|
||||
|
||||
def get_init_inputs():
|
||||
return [1.0
|
||||
|
|
@ -0,0 +1,77 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from ISRLU_torch import Model, get_inputs, get_init_inputs
|
||||
from ISRLU_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