finish tanhexp #20

S1/hli28146_#20/prompt.txt

@@ -0,0 +1,56 @@
+Write a custom CUDA kernel to optimize the TanhExp activation function.
+
+The mathematical definition is:
+f(x) = x * tanh(exp(x))
+
+Problem Analysis:
+The standard PyTorch implementation involves a chain of element-wise operations: exponential, hyperbolic tangent, and multiplication.
+1. exp(x) creates an intermediate tensor.
+2. tanh(intermediate) creates another intermediate tensor.
+3. x * result creates the final output.
+This chain results in excessive global memory read/write traffic, making the operation memory-bound. Additionally, computing two transcendental functions (exp, tanh) per element creates high arithmetic pressure.
+
+Optimization Strategy: Fused Element-wise Kernel with Vectorized Access and Fast Math
+
+1. Operator Fusion: Create a single CUDA kernel that computes `x * tanh(exp(x))` in one pass. Each thread reads `x` once into a register, computes the entire mathematical expression, and writes the result back. This minimizes global memory accesses.
+
+2. Vectorized Memory Access: Use `float4` types to load and store 128 bits (4 floats) per instruction. This drastically improves memory bandwidth utilization and reduces instruction overhead.
+
+3. Grid-Stride Loop: Implement the kernel using a grid-stride loop pattern. This ensures the kernel works correctly and efficiently for input tensors of any size, decoupling the grid configuration from the specific data size.
+
+4. Fast Math Intrinsics: Since the kernel involves `exp` and `tanh`, utilizing fast math intrinsics (like `__expf` or compiling with `--use_fast_math`) is crucial to reduce the latency of the ALU operations, allowing them to be effectively hidden by the optimized memory access.
+
+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)
+
+class TanhExp(nn.Module):
+    """
+    公式: f(x) = x * tanh(e^x)
+    """
+    def __init__(self):
+        super(TanhExp, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.tanh(torch.exp(x))
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = TanhExp()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=torch.float32)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []

S1/hli28146_#20/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 tanhexp_torch import Model,get_inputs,get_init_inputs
+from tanhexp_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()

S1/hli28146_#20/tanhexp_cuda.py

@@ -0,0 +1,103 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cpp_source = """
+#include <torch/extension.h>
+
+torch::Tensor tanhexp_cuda_forward(const torch::Tensor& input);
+"""
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+#include <math.h>
+
+struct __align__(16) Float4 {
+    float x, y, z, w;
+};
+
+// 计算 TanhExp: x * tanh(exp(x))
+__device__ __forceinline__ float tanhexp_op(float x) {
+    // 标准 expf/tanhf 能妥善处理 inf
+    return x * tanhf(expf(x));
+}
+
+__global__ void tanhexp_kernel(
+    const float* __restrict__ input,
+    float* __restrict__ output,
+    const int n_elements) 
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    int stride = blockDim.x * gridDim.x;
+
+    // 1. 向量化处理循环 (每次处理 4 个 float)
+    int vec_loops = n_elements / 4;
+    const Float4* vec_input = reinterpret_cast<const Float4*>(input);
+    Float4* vec_output = reinterpret_cast<Float4*>(output);
+
+    for (int i = idx; i < vec_loops; i += stride) {
+        Float4 in_val = vec_input[i];
+        Float4 out_val;
+        
+        out_val.x = tanhexp_op(in_val.x);
+        out_val.y = tanhexp_op(in_val.y);
+        out_val.z = tanhexp_op(in_val.z);
+        out_val.w = tanhexp_op(in_val.w);
+        
+        vec_output[i] = out_val;
+    }
+
+    // 2. 处理尾部剩余元素 (非 4 对齐的部分)
+    int tail_start = vec_loops * 4;
+    for (int i = tail_start + idx; i < n_elements; i += stride) {
+        output[i] = tanhexp_op(input[i]);
+    }
+}
+
+torch::Tensor tanhexp_cuda_forward(const torch::Tensor& input) {
+    TORCH_CHECK(input.is_cuda(), "Input tensor must be a CUDA tensor");
+    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;
+    
+    int grid_size = (n_elements + block_size * 4 - 1) / (block_size * 4);
+    
+    if (grid_size > 65535) grid_size = 65535;
+
+    tanhexp_kernel<<<grid_size, block_size>>>(
+        input.data_ptr<float>(),
+        output.data_ptr<float>(),
+        n_elements
+    );
+
+    return output;
+}
+"""
+
+tanhexp_op = load_inline(
+    name='tanhexp_op',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['tanhexp_cuda_forward'],
+    verbose=False,
+    extra_cuda_cflags=['-O3', '--use_fast_math'] 
+)
+
+class TanhExpNew(nn.Module):
+    def __init__(self):
+        super(TanhExpNew, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return tanhexp_op.tanhexp_cuda_forward(x)
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.act = TanhExpNew()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)

S1/hli28146_#20/tanhexp_torch.py

@@ -0,0 +1,31 @@
+import torch
+import torch.nn as nn
+
+BATCH_SIZE = 4096
+DIM = 4096
+SHAPE = (BATCH_SIZE, DIM)
+
+class TanhExp(nn.Module):
+    """
+    公式: f(x) = x * tanh(e^x)
+    """
+    def __init__(self):
+        super(TanhExp, self).__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.tanh(torch.exp(x))
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.act = TanhExp()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.act(x)
+
+def get_inputs():
+    x = torch.randn(SHAPE, dtype=torch.float32)
+    return [x.contiguous()]
+
+def get_init_inputs():
+    return []