Compare commits
1 Commits
| Author | SHA1 | Date |
|---|---|---|
|
|
f382eed1b4 |
|
|
@ -0,0 +1,130 @@
|
||||||
|
import torch
|
||||||
|
from torch.utils.cpp_extension import load_inline
|
||||||
|
|
||||||
|
source = """
|
||||||
|
#include <torch/extension.h>
|
||||||
|
#include <cuda_runtime.h>
|
||||||
|
#include <algorithm>
|
||||||
|
|
||||||
|
__device__ __forceinline__ float warp_reduce_sum(float v){
|
||||||
|
for(int offset=16; offset>0; offset/=2){
|
||||||
|
v += __shfl_down_sync(0xffffffff, v, offset);
|
||||||
|
}
|
||||||
|
return v;
|
||||||
|
}
|
||||||
|
|
||||||
|
__global__ __launch_bounds__(256) void layernorm_silu_kernel(const float* __restrict__ x, const float* __restrict__ gamma, const float* __restrict__ beta, float* __restrict__ y, int B, int D, float eps){
|
||||||
|
int b = blockIdx.x;
|
||||||
|
int tid = threadIdx.x;
|
||||||
|
int lane = tid & 31;
|
||||||
|
int warp_id = tid >> 5;
|
||||||
|
int warps = blockDim.x >> 5;
|
||||||
|
const float* xr = x + b * D;
|
||||||
|
float* yr = y + b * D;
|
||||||
|
__shared__ float s_mean;
|
||||||
|
__shared__ float s_invstd;
|
||||||
|
// Welford per-thread
|
||||||
|
float count = 0.0f;
|
||||||
|
float mean = 0.0f;
|
||||||
|
float M2 = 0.0f;
|
||||||
|
for(int i = tid; i < D; i += blockDim.x){
|
||||||
|
float xk = xr[i];
|
||||||
|
count += 1.0f;
|
||||||
|
float delta = xk - mean;
|
||||||
|
mean += delta / count;
|
||||||
|
float delta2 = xk - mean;
|
||||||
|
M2 += delta * delta2;
|
||||||
|
}
|
||||||
|
// Warp reduce Welford
|
||||||
|
for(int offset=16; offset>0; offset/=2){
|
||||||
|
float mean_other = __shfl_down_sync(0xffffffff, mean, offset);
|
||||||
|
float M2_other = __shfl_down_sync(0xffffffff, M2, offset);
|
||||||
|
float cnt_other = __shfl_down_sync(0xffffffff, count, offset);
|
||||||
|
if(cnt_other > 0.0f){
|
||||||
|
float delta = mean_other - mean;
|
||||||
|
float tot = count + cnt_other;
|
||||||
|
M2 += M2_other + delta * delta * (count * cnt_other) / tot;
|
||||||
|
mean += (mean_other - mean) * (cnt_other / tot);
|
||||||
|
count = tot;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
__shared__ float s_mean_warp[32];
|
||||||
|
__shared__ float s_M2_warp[32];
|
||||||
|
__shared__ float s_cnt_warp[32];
|
||||||
|
if(lane == 0){
|
||||||
|
s_mean_warp[warp_id] = mean;
|
||||||
|
s_M2_warp[warp_id] = M2;
|
||||||
|
s_cnt_warp[warp_id] = count;
|
||||||
|
}
|
||||||
|
__syncthreads();
|
||||||
|
if(warp_id == 0){
|
||||||
|
float mean0 = (lane < warps) ? s_mean_warp[lane] : 0.0f;
|
||||||
|
float M20 = (lane < warps) ? s_M2_warp[lane] : 0.0f;
|
||||||
|
float cnt0 = (lane < warps) ? s_cnt_warp[lane] : 0.0f;
|
||||||
|
// reduce across warps
|
||||||
|
for(int offset=16; offset>0; offset/=2){
|
||||||
|
float mean_other = __shfl_down_sync(0xffffffff, mean0, offset);
|
||||||
|
float M2_other = __shfl_down_sync(0xffffffff, M20, offset);
|
||||||
|
float cnt_other = __shfl_down_sync(0xffffffff, cnt0, offset);
|
||||||
|
if(cnt_other > 0.0f){
|
||||||
|
float delta = mean_other - mean0;
|
||||||
|
float tot = cnt0 + cnt_other;
|
||||||
|
M20 += M2_other + delta * delta * (cnt0 * cnt_other) / tot;
|
||||||
|
mean0 += (mean_other - mean0) * (cnt_other / tot);
|
||||||
|
cnt0 = tot;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
if(lane == 0){
|
||||||
|
float var = M20 / fmaxf(cnt0, 1.0f);
|
||||||
|
s_mean = mean0;
|
||||||
|
s_invstd = rsqrtf(fmaxf(var, 0.0f) + eps);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
__syncthreads();
|
||||||
|
for(int i = tid; i < D; i += blockDim.x){
|
||||||
|
float v = (__ldg(xr + i) - s_mean) * s_invstd;
|
||||||
|
float u = v * __ldg(gamma + i) + __ldg(beta + i);
|
||||||
|
float g = __fdividef(1.0f, 1.0f + __expf(-u));
|
||||||
|
yr[i] = u * g;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
torch::Tensor layernorm_silu_cuda(torch::Tensor x, torch::Tensor gamma, torch::Tensor beta, torch::Tensor eps){
|
||||||
|
auto xc = x.contiguous();
|
||||||
|
auto gc = gamma.contiguous();
|
||||||
|
auto bc = beta.contiguous();
|
||||||
|
auto y = torch::empty_like(xc);
|
||||||
|
int B = (int)xc.size(0);
|
||||||
|
int D = (int)xc.size(1);
|
||||||
|
float e = eps.item<float>();
|
||||||
|
int block = 256;
|
||||||
|
dim3 grid(B);
|
||||||
|
layernorm_silu_kernel<<<grid, block>>>(xc.data_ptr<float>(), gc.data_ptr<float>(), bc.data_ptr<float>(), y.data_ptr<float>(), B, D, e);
|
||||||
|
return y;
|
||||||
|
}
|
||||||
|
"""
|
||||||
|
|
||||||
|
cpp_source = """
|
||||||
|
torch::Tensor layernorm_silu_cuda(torch::Tensor x, torch::Tensor gamma, torch::Tensor beta, torch::Tensor eps);
|
||||||
|
"""
|
||||||
|
|
||||||
|
ops = load_inline(
|
||||||
|
name="layernorm_silu",
|
||||||
|
cpp_sources=cpp_source,
|
||||||
|
cuda_sources=source,
|
||||||
|
functions=["layernorm_silu_cuda"],
|
||||||
|
extra_cflags=["-O3","-std=c++17"],
|
||||||
|
extra_cuda_cflags=["-O3","--use_fast_math","-std=c++17"],
|
||||||
|
verbose=True
|
||||||
|
)
|
||||||
|
|
||||||
|
class ModelNew(torch.nn.Module):
|
||||||
|
def __init__(self, gamma: torch.Tensor, beta: torch.Tensor, eps: float):
|
||||||
|
super(ModelNew, self).__init__()
|
||||||
|
self.ops = ops
|
||||||
|
self.register_buffer("gamma", gamma)
|
||||||
|
self.register_buffer("beta", beta)
|
||||||
|
self.register_buffer("eps", torch.tensor(float(eps), dtype=torch.float32))
|
||||||
|
|
||||||
|
def forward(self, x):
|
||||||
|
return self.ops.layernorm_silu_cuda(x, self.gamma, self.beta, self.eps)
|
||||||
|
|
@ -0,0 +1,15 @@
|
||||||
|
Objective: Fused LayerNorm + SiLU CUDA kernel with rtol=1e-3 accuracy and ≥1.3x speedup.
|
||||||
|
|
||||||
|
Computation:
|
||||||
|
- Row layer normalization (mean/var over dim) with epsilon, then affine transform, then SiLU: `y = u * sigmoid(u)`.
|
||||||
|
|
||||||
|
Numerical Method:
|
||||||
|
- Use Welford’s online algorithm for mean/variance to match PyTorch precision and avoid catastrophic cancellation.
|
||||||
|
- Perform warp-level reductions and shared-memory aggregation across warps.
|
||||||
|
|
||||||
|
Performance:
|
||||||
|
- 512 threads per block; coalesced loads and parallel write-back.
|
||||||
|
- Fuse normalization, affine, and activation to minimize memory traffic.
|
||||||
|
|
||||||
|
Benchmark:
|
||||||
|
- Batch size: 16; Dim: 16384; 100 iterations; target speedup ≥1.3x.
|
||||||
|
|
@ -0,0 +1,51 @@
|
||||||
|
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().eval()
|
||||||
|
cuda_model = ModelNew(*init_inputs).cuda().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
|
||||||
|
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 LayerNorm-SiLU 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 +1,30 @@
|
||||||
|
import torch
|
||||||
|
import torch.nn as nn
|
||||||
|
|
||||||
|
class Model(nn.Module):
|
||||||
|
def __init__(self, gamma: torch.Tensor, beta: torch.Tensor, eps: float):
|
||||||
|
super(Model, self).__init__()
|
||||||
|
self.register_buffer("gamma", gamma)
|
||||||
|
self.register_buffer("beta", beta)
|
||||||
|
self.register_buffer("eps", torch.tensor(float(eps), dtype=torch.float32))
|
||||||
|
|
||||||
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||||
|
mean = x.mean(dim=1, keepdim=True)
|
||||||
|
var = x.var(dim=1, unbiased=False, keepdim=True)
|
||||||
|
xn = (x - mean) / torch.sqrt(var + self.eps)
|
||||||
|
u = xn * self.gamma + self.beta
|
||||||
|
g = torch.sigmoid(u)
|
||||||
|
return u * g
|
||||||
|
|
||||||
|
batch_size = 16
|
||||||
|
dim = 16384
|
||||||
|
|
||||||
|
def get_inputs():
|
||||||
|
x = torch.randn(batch_size, dim)
|
||||||
|
return [x]
|
||||||
|
|
||||||
|
def get_init_inputs():
|
||||||
|
gamma = torch.randn(dim)
|
||||||
|
beta = torch.randn(dim)
|
||||||
|
return [gamma, beta, 1e-5]
|
||||||
|
|
||||||
Loading…
Reference in New Issue