forked from ccf-ai-infra/GPUCodeForces
Merge pull request 'finish penalizedtanh #64' (#682) from hli28146/GPUCodeForces:h64 into main
This commit is contained in:
commit
6006bf32ec
|
|
@ -0,0 +1,108 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
cpp_source = """
|
||||
#include <torch/extension.h>
|
||||
|
||||
torch::Tensor ptanh_cuda_forward(const torch::Tensor& input, float alpha);
|
||||
"""
|
||||
|
||||
cuda_source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
#include <math.h>
|
||||
|
||||
#define BLOCK_SIZE 256
|
||||
|
||||
struct __align__(16) Float4 {
|
||||
float x, y, z, w;
|
||||
};
|
||||
|
||||
// Penalized Tanh Logic
|
||||
// t = tanh(x)
|
||||
// res = (x > 0) ? t : t * alpha
|
||||
__device__ __forceinline__ float compute_ptanh(float x, float alpha) {
|
||||
float t = tanhf(x);
|
||||
if (x > 0.0f) {
|
||||
return t;
|
||||
} else {
|
||||
return t * alpha;
|
||||
}
|
||||
}
|
||||
|
||||
__global__ void ptanh_kernel(
|
||||
float* __restrict__ output,
|
||||
const float* __restrict__ input,
|
||||
const int n,
|
||||
const float alpha)
|
||||
{
|
||||
const int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
const int vec_n = n / 4;
|
||||
|
||||
int i = idx;
|
||||
const int stride = blockDim.x * gridDim.x;
|
||||
|
||||
for (; i < vec_n; i += stride) {
|
||||
Float4 in_vec = reinterpret_cast<const Float4*>(input)[i];
|
||||
Float4 out_vec;
|
||||
|
||||
out_vec.x = compute_ptanh(in_vec.x, alpha);
|
||||
out_vec.y = compute_ptanh(in_vec.y, alpha);
|
||||
out_vec.z = compute_ptanh(in_vec.z, alpha);
|
||||
out_vec.w = compute_ptanh(in_vec.w, alpha);
|
||||
|
||||
reinterpret_cast<Float4*>(output)[i] = out_vec;
|
||||
}
|
||||
|
||||
int tail_start = vec_n * 4;
|
||||
int global_tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||||
|
||||
if (global_tid < (n - tail_start)) {
|
||||
int real_idx = tail_start + global_tid;
|
||||
output[real_idx] = compute_ptanh(input[real_idx], alpha);
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor ptanh_cuda_forward(const torch::Tensor& input, float alpha) {
|
||||
TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
|
||||
TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
|
||||
|
||||
const int n = input.numel();
|
||||
auto output = torch::empty_like(input);
|
||||
|
||||
const int vec_n = n / 4;
|
||||
// Launch enough blocks
|
||||
const int grid_size = (vec_n + BLOCK_SIZE - 1) / BLOCK_SIZE;
|
||||
|
||||
int final_grid = (grid_size < 1) ? 1 : grid_size;
|
||||
if (final_grid > 65535) final_grid = 65535;
|
||||
|
||||
ptanh_kernel<<<final_grid, BLOCK_SIZE>>>(
|
||||
output.data_ptr<float>(),
|
||||
input.data_ptr<float>(),
|
||||
n,
|
||||
alpha
|
||||
);
|
||||
|
||||
return output;
|
||||
}
|
||||
"""
|
||||
|
||||
ptanh_op_module = load_inline(
|
||||
name='ptanh_op',
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=cuda_source,
|
||||
functions=['ptanh_cuda_forward'],
|
||||
verbose=False,
|
||||
extra_cuda_cflags=['-O3', '--use_fast_math']
|
||||
)
|
||||
|
||||
class ModelNew(nn.Module):
|
||||
def __init__(self, alpha=0.25):
|
||||
super(ModelNew, self).__init__()
|
||||
self.alpha = alpha
|
||||
self.op = ptanh_op_module
|
||||
|
||||
def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
|
||||
return self.op.ptanh_cuda_forward(input_tensor.contiguous(), self.alpha)
|
||||
|
|
@ -0,0 +1,39 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
BATCH_SIZE = 4096
|
||||
HIDDEN_DIM = 4096
|
||||
SHAPE = (BATCH_SIZE, HIDDEN_DIM)
|
||||
|
||||
ALPHA_VAL = 0.25
|
||||
|
||||
class PenalizedTanh(nn.Module):
|
||||
"""
|
||||
Penalized Tanh
|
||||
https://arxiv.org/pdf/1602.05980
|
||||
f(x) = tanh(x) if x > 0
|
||||
alpha * tanh(x) if x <= 0
|
||||
"""
|
||||
def __init__(self, alpha=0.25):
|
||||
super(PenalizedTanh, self).__init__()
|
||||
self.alpha = alpha
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
t = torch.tanh(x)
|
||||
return torch.where(x > 0, t, self.alpha * t)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, alpha=0.25):
|
||||
super(Model, self).__init__()
|
||||
self.act = PenalizedTanh(alpha=alpha)
|
||||
|
||||
def forward(self, x):
|
||||
return self.act(x)
|
||||
|
||||
def get_inputs():
|
||||
input_tensor = torch.randn(SHAPE, dtype=torch.float32)
|
||||
return [input_tensor.contiguous()]
|
||||
|
||||
def get_init_inputs():
|
||||
return [ALPHA_VAL]
|
||||
|
|
@ -0,0 +1,62 @@
|
|||
Write a custom CUDA kernel to optimize `Penalized Tanh` activation.
|
||||
|
||||
Formula: f(x) = tanh(x) if x > 0 else alpha * tanh(x)
|
||||
This is equivalent to: tanh(x) * (x > 0 ? 1 : alpha)
|
||||
|
||||
Problem Analysis:
|
||||
1. Memory Bandwidth: As an element-wise activation function, the arithmetic intensity is low. The performance is dominated by the speed of reading input and writing output (Memory Bound).
|
||||
2. Tanh Cost: calculating `tanh` involves expensive exponential operations. However, modern GPUs have Special Function Units (SFUs), and the memory latency usually dominates.
|
||||
3. Multiple Passes: A naive PyTorch implementation might compute `tanh(x)`, create a mask `x>0`, and then combine, resulting in redundant reads/writes.
|
||||
|
||||
Optimization Strategy: Fused Element-wise Kernel with Vectorization
|
||||
|
||||
1. One-Pass Fused Kernel: Perform the tanh computation and the conditional scaling in a single pass. Load `x`, compute `t = tanh(x)`, apply scaling based on the sign of `x`, and store.
|
||||
|
||||
2. Vectorized Loads (float4): Use `float4` to load 4 float elements (128 bits) per thread instruction. This is the most effective optimization for memory-bound kernels on Nvidia GPUs.
|
||||
|
||||
3. Instruction Optimization: Calculate `tanh(x)` once per element. The branching logic `x > 0` is cheap compared to memory access.
|
||||
|
||||
4. Kernel Configuration: Launch a 1D grid with enough blocks to cover the entire tensor size.
|
||||
|
||||
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
|
||||
|
||||
BATCH_SIZE = 4096
|
||||
HIDDEN_DIM = 4096
|
||||
SHAPE = (BATCH_SIZE, HIDDEN_DIM)
|
||||
|
||||
ALPHA_VAL = 0.25
|
||||
|
||||
class PenalizedTanh(nn.Module):
|
||||
"""
|
||||
Penalized Tanh
|
||||
https://arxiv.org/pdf/1602.05980
|
||||
f(x) = tanh(x) if x > 0
|
||||
alpha * tanh(x) if x <= 0
|
||||
"""
|
||||
def __init__(self, alpha=0.25):
|
||||
super(PenalizedTanh, self).__init__()
|
||||
self.alpha = alpha
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
t = torch.tanh(x)
|
||||
return torch.where(x > 0, t, self.alpha * t)
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, alpha=0.25):
|
||||
super(Model, self).__init__()
|
||||
self.act = PenalizedTanh(alpha=alpha)
|
||||
|
||||
def forward(self, x):
|
||||
return self.act(x)
|
||||
|
||||
def get_inputs():
|
||||
input_tensor = torch.randn(SHAPE, dtype=torch.float32)
|
||||
return [input_tensor.contiguous()]
|
||||
|
||||
def get_init_inputs():
|
||||
return [ALPHA_VAL]
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
###########################################################
|
||||
# 性能和精度验证程序
|
||||
###########################################################
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import time
|
||||
from penalizedtanh_torch import Model,get_inputs,get_init_inputs
|
||||
from penalizedtanh_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