finish ASU #30

S1/hli28146_#30/asu_cuda.py

@@ -0,0 +1,102 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cpp_source = """
+#include <torch/extension.h>
+
+torch::Tensor asu_cuda_forward(const torch::Tensor& input);
+"""
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+#include <math.h>
+
+// Vectorized type for 128-bit access with doubles
+struct __align__(16) Double2 {
+    double x, y;
+};
+
+// Core computation
+// Formula: x * sin(x)
+__device__ __forceinline__ double asu_op(double x) {
+    return x * sin(x);
+}
+
+__global__ void asu_kernel_double(
+    const double* __restrict__ input,
+    double* __restrict__ output,
+    const int n_elements) 
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int stride = blockDim.x * gridDim.x;
+
+    // 1. Vectorized Loop
+    int vec_loops = n_elements / 2;
+    const Double2* vec_input = reinterpret_cast<const Double2*>(input);
+    Double2* vec_output = reinterpret_cast<Double2*>(output);
+
+    for (int i = idx; i < vec_loops; i += stride) {
+        Double2 in_val = vec_input[i];
+        Double2 out_val;
+        
+        out_val.x = asu_op(in_val.x);
+        out_val.y = asu_op(in_val.y);
+        
+        vec_output[i] = out_val;
+    }
+
+    // 2. Scalar Loop
+    int tail_start = vec_loops * 2;
+    for (int i = tail_start + idx; i < n_elements; i += stride) {
+        output[i] = asu_op(input[i]);
+    }
+}
+
+torch::Tensor asu_cuda_forward(const torch::Tensor& input) {
+    TORCH_CHECK(input.is_cuda(), "Input tensor must be a CUDA tensor");
+    TORCH_CHECK(input.scalar_type() == torch::kDouble, "Input tensor must be float64");
+    TORCH_CHECK(input.is_contiguous(), "Input tensor must be contiguous");
+
+    auto output = torch::empty_like(input);
+    const int n_elements = input.numel();
+
+    const int block_size = 256;
+    // Grid size for Double2 (2 elements per thread)
+    int grid_size = (n_elements + block_size * 2 - 1) / (block_size * 2);
+    if (grid_size > 65535) grid_size = 65535;
+
+    asu_kernel_double<<<grid_size, block_size>>>(
+        input.data_ptr<double>(),
+        output.data_ptr<double>(),
+        n_elements
+    );
+
+    return output;
+}
+"""
+
+asu_op = load_inline(
+    name='asu_op',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['asu_cuda_forward'],
+    verbose=False,
+    extra_cuda_cflags=['-O3']
+)
+
+class ASUNew(nn.Module):
+    def __init__(self):
+        super(ASUNew, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return asu_op.asu_cuda_forward(x)
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.act = ASUNew()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)

S1/hli28146_#30/asu_torch.py

@@ -0,0 +1,35 @@
+import torch
+import torch.nn as nn
+
+BATCH_SIZE = 4096
+DIM = 4096
+SHAPE = (BATCH_SIZE, DIM)
+
+DTYPE = torch.float64
+
+class ASU(nn.Module):
+    """
+    Amplifying Sine Unit: An Oscillatory Activation Function for Deep Neural Networks to Recover Nonlinear Oscillations Efficiently 
+    https://arxiv.org/pdf/2304.09759
+    Formula: f(x) = x * sin(x)
+    """
+    def __init__(self):
+        super(ASU, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.sin(x)
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = ASU()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=DTYPE)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []

S1/hli28146_#30/prompt.txt

@@ -0,0 +1,57 @@
+Write a custom CUDA kernel to optimize the ASU activation function as defined in the provided table.
+
+The mathematical definition is:
+f(x) = x * sin(x)
+
+Problem Analysis:
+1. Memory Bandwidth: The operation is element-wise and strictly memory-bound. The arithmetic intensity is low (one sin, one mul). Standard PyTorch implementation executes `sin(x)` followed by `x * result`, involving intermediate memory traffic.
+2. Precision: Trigonometric functions are sensitive to precision. Double precision (float64) is required for strict accuracy alignment with the reference.
+
+Optimization Strategy: Fused Vectorized Kernel in Double Precision
+
+1. Data Type: Use `double` for all computations to guarantee numerical stability and accuracy.
+
+2. Vectorized Memory Access: Use `double2` types to load/store 128 bits (2 doubles) per instruction. This is the optimal transaction size for float64 data on GPUs, significantly reducing instruction overhead and maximizing bandwidth.
+
+3. Fused Computation: Compute `val * sin(val)` entirely in registers. This fuses the two element-wise operations into a single kernel pass (1 read, 1 write).
+
+4. Grid-Stride Loop: Implement a robust grid-stride loop to handle arbitrary input tensor sizes efficiently.
+  
+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
+DIM = 4096
+SHAPE = (BATCH_SIZE, DIM)
+
+DTYPE = torch.float64
+
+class ASU(nn.Module):
+    """
+    Amplifying Sine Unit: An Oscillatory Activation Function for Deep Neural Networks to Recover Nonlinear Oscillations Efficiently 
+    https://arxiv.org/pdf/2304.09759
+    Formula: f(x) = x * sin(x)
+    """
+    def __init__(self):
+        super(ASU, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.sin(x)
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = ASU()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=DTYPE)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []

S1/hli28146_#30/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 asu_torch import Model,get_inputs,get_init_inputs
+from asu_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()