diff --git a/S1/hli28146_#126/PFLU_cuda.py b/S1/hli28146_#126/PFLU_cuda.py new file mode 100644 index 00000000..ddf8bf9d --- /dev/null +++ b/S1/hli28146_#126/PFLU_cuda.py @@ -0,0 +1,101 @@ +import torch +import torch.nn as nn +from torch.utils.cpp_extension import load_inline + +cpp_source = """ +#include + +torch::Tensor pflu_cuda_forward(const torch::Tensor& input); +""" + +cuda_source = """ +#include +#include +#include + +#define BLOCK_SIZE 256 + +struct __align__(16) Float4 { + float x, y, z, w; +}; + +// PFLU Logic +__device__ __forceinline__ float compute_pflu(float x) { + // Use fast inverse square root for performance + float inv_sqrt = rsqrtf(1.0f + x * x); + return x * (1.0f + x * inv_sqrt); +} + +__global__ void pflu_kernel( + float* __restrict__ output, + const float* __restrict__ input, + const int n) +{ + 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(input)[i]; + Float4 out_vec; + + out_vec.x = compute_pflu(in_vec.x); + out_vec.y = compute_pflu(in_vec.y); + out_vec.z = compute_pflu(in_vec.z); + out_vec.w = compute_pflu(in_vec.w); + + reinterpret_cast(output)[i] = out_vec; + } + + int start_scalar = vec_n * 4; + int global_tid = blockIdx.x * blockDim.x + threadIdx.x; + int total_threads = gridDim.x * gridDim.x; + + int current_idx = start_scalar + global_tid; + while (current_idx < n) { + output[current_idx] = compute_pflu(input[current_idx]); + current_idx += total_threads; + } +} + +torch::Tensor pflu_cuda_forward(const torch::Tensor& input) { + 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; + 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; + + pflu_kernel<<>>( + output.data_ptr(), + input.data_ptr(), + n + ); + + return output; +} +""" + +pflu_op_module = load_inline( + name='pflu_op', + cpp_sources=cpp_source, + cuda_sources=cuda_source, + functions=['pflu_cuda_forward'], + verbose=False, + extra_cuda_cflags=['-O3'] +) + +class ModelNew(nn.Module): + def __init__(self): + super(ModelNew, self).__init__() + self.op = pflu_op_module + + def forward(self, input_tensor: torch.Tensor) -> torch.Tensor: + return self.op.pflu_cuda_forward(input_tensor.contiguous()) \ No newline at end of file diff --git a/S1/hli28146_#126/PFLU_torch.py b/S1/hli28146_#126/PFLU_torch.py new file mode 100644 index 00000000..e018089e --- /dev/null +++ b/S1/hli28146_#126/PFLU_torch.py @@ -0,0 +1,35 @@ +import torch +import torch.nn as nn + +BATCH_SIZE = 4096 +HIDDEN_DIM = 4096 +SHAPE = (BATCH_SIZE, HIDDEN_DIM) + +class PFLU(nn.Module): + """ + Power Function Linear Unit (PFLU). + PFLU and FPFLU: Two novel non-monotonic activation functions in convolutional neural network + https://doi.org/10.1016/j.neucom.2020.11.068 + + Formula: f(x) = x * (1 + x / sqrt(1 + x^2)) + """ + def __init__(self): + super(PFLU, self).__init__() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x * (1.0 + x / torch.sqrt(1.0 + x.pow(2))) + +class Model(nn.Module): + def __init__(self): + super(Model, self).__init__() + self.act = PFLU() + + def forward(self, x): + return self.act(x) + +def get_inputs(): + input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0 + return [input_tensor.contiguous()] + +def get_init_inputs(): + return [] \ No newline at end of file diff --git a/S1/hli28146_#126/prompt.txt b/S1/hli28146_#126/prompt.txt new file mode 100644 index 00000000..f7b999ca --- /dev/null +++ b/S1/hli28146_#126/prompt.txt @@ -0,0 +1,61 @@ +Write a custom CUDA kernel to optimize `PFLU` (Power Function Linear Unit). + +Formula: f(x) = x * (1 + x / sqrt(1 + x^2)) + +Problem Analysis: +1. Memory Bound: This is an element-wise activation with moderate arithmetic intensity (sqrt, div, mul, add). +2. Operator Chaining: A PyTorch implementation creates intermediate tensors for pow, sqrt, etc. + +Optimization Strategy: Fused Element-wise Kernel with Vectorization + +1. One-Thread-per-Element: Map each element to a CUDA thread. + +2. Vectorized Loads (float4): Use `float4` to process 128 bits per memory transaction. + +3. Fused Fast Math: + - For each element `x`: + `inv_sqrt = rsqrtf(1.0f + x * x)` (fast inverse square root) + `term = x * inv_sqrt` + `result = x * (1.0f + term)` + - All computations are fused in registers. + +4. One-Pass: Fuse all steps into a single read-compute-write kernel. + +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_SIZE = 4096 +HIDDEN_DIM = 4096 +SHAPE = (BATCH_SIZE, HIDDEN_DIM) + +class PFLU(nn.Module): + """ + Power Function Linear Unit (PFLU). + PFLU and FPFLU: Two novel non-monotonic activation functions in convolutional neural network + https://doi.org/10.1016/j.neucom.2020.11.068 + + Formula: f(x) = x * (1 + x / sqrt(1 + x^2)) + """ + def __init__(self): + super(PFLU, self).__init__() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x * (1.0 + x / torch.sqrt(1.0 + x.pow(2))) + +class Model(nn.Module): + def __init__(self): + super(Model, self).__init__() + self.act = PFLU() + + def forward(self, x): + return self.act(x) + +def get_inputs(): + input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0 + return [input_tensor.contiguous()] + +def get_init_inputs(): + return [] \ No newline at end of file diff --git a/S1/hli28146_#126/run_code.py b/S1/hli28146_#126/run_code.py new file mode 100644 index 00000000..04f112c0 --- /dev/null +++ b/S1/hli28146_#126/run_code.py @@ -0,0 +1,74 @@ +########################################################### +# 性能和精度验证程序 +########################################################### +import torch +import torch.nn as nn +import time +from PFLU_torch import Model,get_inputs,get_init_inputs +from PFLU_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() \ No newline at end of file