forked from ccf-ai-infra/GPUCodeForces
finish hardwish based on afgate #48
This commit is contained in:
parent
10eed82956
commit
425fdde5cf
|
|
@ -0,0 +1,94 @@
|
|||
import torch
|
||||
from torch.utils.cpp_extension import load_inline
|
||||
|
||||
source = """
|
||||
#include <torch/extension.h>
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
__global__ void hardswish_affine_gate_kernel(const float* x, const float* scale, const float* bias, float* y, int B, int D, float alpha, float beta){
|
||||
int b = blockIdx.x;
|
||||
int tid = threadIdx.x;
|
||||
int stride = blockDim.x;
|
||||
int row_start = b * D;
|
||||
const float* xr = x + row_start;
|
||||
float* yr = y + row_start;
|
||||
int aligned = ((((long long)xr & 15LL) == 0) && (((long long)yr & 15LL) == 0) && (((long long)scale & 15LL) == 0) && (((long long)bias & 15LL) == 0) && ((D & 3) == 0));
|
||||
if(aligned){
|
||||
int D4 = (D / 4) * 4;
|
||||
#pragma unroll 4
|
||||
for(int i = tid * 4; i < D4; i += stride * 4){
|
||||
float4 xv = reinterpret_cast<const float4*>(xr)[i / 4];
|
||||
float4 sv = reinterpret_cast<const float4*>(scale)[i / 4];
|
||||
float4 bv = reinterpret_cast<const float4*>(bias)[i / 4];
|
||||
float4 yv;
|
||||
float z0 = fmaf(xv.x, sv.x, bv.x);
|
||||
float z1 = fmaf(xv.y, sv.y, bv.y);
|
||||
float z2 = fmaf(xv.z, sv.z, bv.z);
|
||||
float z3 = fmaf(xv.w, sv.w, bv.w);
|
||||
float t0 = fminf(fmaxf(alpha * z0 + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
float t1 = fminf(fmaxf(alpha * z1 + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
float t2 = fminf(fmaxf(alpha * z2 + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
float t3 = fminf(fmaxf(alpha * z3 + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
yv.x = xv.x * t0;
|
||||
yv.y = xv.y * t1;
|
||||
yv.z = xv.z * t2;
|
||||
yv.w = xv.w * t3;
|
||||
reinterpret_cast<float4*>(yr)[i / 4] = yv;
|
||||
}
|
||||
#pragma unroll 4
|
||||
for(int i = D4 + tid; i < D; i += stride){
|
||||
float z = fmaf(xr[i], scale[i], bias[i]);
|
||||
float t = fminf(fmaxf(alpha * z + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
yr[i] = xr[i] * t;
|
||||
}
|
||||
} else {
|
||||
#pragma unroll 4
|
||||
for(int i = tid; i < D; i += stride){
|
||||
float z = fmaf(xr[i], scale[i], bias[i]);
|
||||
float t = fminf(fmaxf(alpha * z + beta + 3.0f, 0.0f), 6.0f) * (1.0f / 6.0f);
|
||||
yr[i] = xr[i] * t;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
torch::Tensor hardswish_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias, torch::Tensor alpha, torch::Tensor beta){
|
||||
auto xc = x.contiguous();
|
||||
auto sc = scale.contiguous();
|
||||
auto bc = bias.contiguous();
|
||||
auto y = torch::empty_like(xc);
|
||||
int B = (int)xc.size(0);
|
||||
int D = (int)xc.size(1);
|
||||
float a = alpha.item<float>();
|
||||
float be = beta.item<float>();
|
||||
int block = 1024;
|
||||
int grid = B;
|
||||
hardswish_affine_gate_kernel<<<grid, block>>>(xc.data_ptr<float>(), sc.data_ptr<float>(), bc.data_ptr<float>(), y.data_ptr<float>(), B, D, a, be);
|
||||
return y;
|
||||
}
|
||||
"""
|
||||
|
||||
cpp_source = """
|
||||
torch::Tensor hardswish_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias, torch::Tensor alpha, torch::Tensor beta);
|
||||
"""
|
||||
|
||||
ops = load_inline(
|
||||
name="hardswish_affine_gate",
|
||||
cpp_sources=cpp_source,
|
||||
cuda_sources=source,
|
||||
functions=["hardswish_affine_gate_cuda"],
|
||||
extra_cuda_cflags=["-O3","--use_fast_math"],
|
||||
verbose=True
|
||||
)
|
||||
|
||||
class ModelNew(torch.nn.Module):
|
||||
def __init__(self, scale: torch.Tensor, bias: torch.Tensor, alpha: float, beta: float):
|
||||
super(ModelNew, self).__init__()
|
||||
self.ops = ops
|
||||
self.register_buffer("scale", scale)
|
||||
self.register_buffer("bias", bias)
|
||||
self.register_buffer("alpha", torch.tensor(float(alpha), dtype=torch.float32))
|
||||
self.register_buffer("beta", torch.tensor(float(beta), dtype=torch.float32))
|
||||
|
||||
def forward(self, x):
|
||||
return self.ops.hardswish_affine_gate_cuda(x, self.scale, self.bias, self.alpha, self.beta)
|
||||
|
||||
|
|
@ -0,0 +1,27 @@
|
|||
融合算子：Hardswish-Affine-Gate（一次核内完成仿射与 Hardswish 风格门控，返回 y = x * clamp(α*z + β + 3, 0, 6) / 6，其中 z = x*scale + bias）。该门控具备近似线性区域与饱和特性，适合稳健的逐维缩放。
|
||||
|
||||
目标与定义
|
||||
- 输入张量：`x[B, D]`
|
||||
- 逐维参数：`scale[D]`、`bias[D]`
|
||||
- 标量超参：`alpha`、`beta`
|
||||
- 计算流程：`z = x*scale + bias`，`g = clamp(alpha*z + beta + 3, 0, 6) / 6`，`y = x * g`
|
||||
|
||||
参考实现（文件要求）
|
||||
- `torchcode.py`：PyTorch 参考 `Model`；统一输入与初始化接口
|
||||
- `cudacode.py`：单核融合（仿射+clamp 线段门控+乘法）；编译 `-O3 --use_fast_math`
|
||||
- `run_code.py`：迭代 100 次；精度 `rtol=1e-03, atol=1e-06`；打印加速比
|
||||
|
||||
CUDA 实现要点
|
||||
- 并行布局：`grid = B`；块内沿 D 连续访存并一次写回
|
||||
- 对齐向量化：满足 16 字节对齐且 `D%4==0` 时走 `float4`；否则标量路径
|
||||
- 指令优化：仿射用 `fmaf`；clamp 分支尽量保持分支收敛；循环 `#pragma unroll 4`
|
||||
- 线程配置：参考 `block=1024`，按设备与规模微调至 `256/512/1024`
|
||||
|
||||
评估与目标
|
||||
- 精度：参考与 CUDA 在 `rtol=1e-03, atol=1e-06` 内一致
|
||||
- 性能：≥1.0x 加速；在对齐触发向量化时更佳
|
||||
|
||||
加分项（可选）
|
||||
- 将 clamp 写成逐元素最小/最大组合减少分支（`fminf/fmaxf`）
|
||||
- 优化尾部处理与非对齐路径的开销
|
||||
|
||||
|
|
@ -0,0 +1,57 @@
|
|||
import torch
|
||||
import time
|
||||
from torchcode import Model, get_inputs, get_init_inputs
|
||||
from cudacode import ModelNew
|
||||
|
||||
def run_benchmark():
|
||||
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
|
||||
device = torch.device("cuda")
|
||||
|
||||
init_inputs = [x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in get_init_inputs()]
|
||||
inputs = [x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in get_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, atol=1e-06)
|
||||
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！")
|
||||
diff = (output_torch - output_cuda).abs().max().item()
|
||||
print(f"最大绝对误差: {diff}")
|
||||
print(f"输出张量形状: torch={tuple(output_torch.shape)}, cuda={tuple(output_cuda.shape)}")
|
||||
print(f"数据类型: torch={output_torch.dtype}, cuda={output_cuda.dtype}")
|
||||
print(f"设备: torch={output_torch.device}, cuda={output_cuda.device}")
|
||||
|
||||
print("\n-------------------- 性能加速比测试 --------------------")
|
||||
num_iterations = 100
|
||||
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
|
||||
|
||||
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 Hardswish-Affine-Gate 平均执行时间: {torch_time:.6f} 秒")
|
||||
print(f"自定义 CUDA 融合内核 平均执行时间: {cuda_time:.6f} 秒")
|
||||
speedup = torch_time / cuda_time if cuda_time > 0 else 0
|
||||
if cuda_time > 0:
|
||||
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__":
|
||||
run_benchmark()
|
||||
|
||||
|
|
@ -0,0 +1,28 @@
|
|||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
class Model(nn.Module):
|
||||
def __init__(self, scale: torch.Tensor, bias: torch.Tensor, alpha: float, beta: float):
|
||||
super(Model, self).__init__()
|
||||
self.register_buffer("scale", scale)
|
||||
self.register_buffer("bias", bias)
|
||||
self.register_buffer("alpha", torch.tensor(float(alpha), dtype=torch.float32))
|
||||
self.register_buffer("beta", torch.tensor(float(beta), dtype=torch.float32))
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
z = x * self.scale + self.bias
|
||||
t = torch.clamp(self.alpha * z + self.beta + 3.0, min=0.0, max=6.0) * (1.0 / 6.0)
|
||||
return x * t
|
||||
|
||||
batch_size = 16
|
||||
dim = 16384
|
||||
|
||||
def get_inputs():
|
||||
x = torch.randn(batch_size, dim)
|
||||
return [x]
|
||||
|
||||
def get_init_inputs():
|
||||
scale = torch.randn(dim)
|
||||
bias = torch.randn(dim)
|
||||
return [scale, bias, 1.0, 0.0]
|
||||
|
||||
Loading…
Reference in New Issue