finish zscore_sigmoid_denormalize #116

S1/uucoco_#116/prompt.txt

@@ -0,0 +1,88 @@
+You write custom CUDA kernels to replace the pytorch operators in the given GeGLU architecture to get speedups.
+
+You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining chunk+gelu+elementwise_mul), or algorithmic changes (such as optimized memory access patterns). You are only limited by your imagination.
+
+# Technologies Used in This Code
+
+## Core Libraries
+- **PyTorch**: Deep learning framework
+- **CUDA**: NVIDIA GPU parallel computing
+- **C++**: Kernel implementation
+
+## CUDA Components
+- **CUDA kernel**: `zscore_sigmoid_denormalize_kernel`
+- **CUDA math functions**: `expf()`, `fmaf()` (fused multiply-add)
+- **Element-wise parallelism**: One thread per element
+- **FMA optimization**: Using fused multiply-add for better precision
+
+## Mathematical Operations Pipeline
+1. **Z-score normalization**: `(x - mean) / std`
+2. **Sigmoid activation**: `1 / (1 + exp(-y))`
+3. **Inverse Z-score**: `z * std + mean`
+- **Fused operations**: Three-step transformation in single kernel
+
+## Architecture
+- **Standard 1D grid**: Simple block/grid configuration
+- **Element-wise computation**: Independent processing per element
+- **Memory pattern**: Coalesced memory access
+
+## CUDA Math Optimizations
+- **expf()**: Single-precision exponential
+- **fmaf()**: Fused multiply-add for denormalization (z*std + mean)
+- **Efficient computation**: Minimizes rounding errors
+
+## Mathematical Properties
+- **Reversible transformation**: Sigmoid on normalized data
+- **Range preservation**: Output remains in similar range as input
+- **Non-linear transformation**: Sigmoid introduces non-linearity
+- **Parameterized**: User-provided mean and std parameters
+
+## Performance Features
+- **GPU acceleration**: Parallel computation across all elements
+- **Fused kernel**: Three operations combined for efficiency
+- **Numerical stability**: FMA reduces rounding errors
+- **Simple operations**: Moderate computational cost
+
+## Numerical Considerations
+- **Std requirement**: std ≠ 0 (no protection in code)
+- **Numerical range**: Sigmoid output in (0,1)
+- **Denormalization**: Maps sigmoid output back to original scale
+- **Potential overflow**: exp(-y) could underflow for large y
+
+## Use Case Applications
+- **Normalized activation**: Sigmoid on standardized data
+- **Range transformation**: Map data through normalized sigmoid
+- **Custom scaling**: User-defined mean/std for specific ranges
+- **Element-wise processing**: Independent transformation per element
+
+
+
+Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self, mean, std):
+        super(Model, self).__init__()
+        self.mean = mean
+        self.std = std
+
+    def forward(self, x):
+        y = (x - self.mean) / self.std
+        z = torch.sigmoid(y)
+
+        return z * self.std + self.mean
+
+
+batch_size = 1024
+dim = 1024
+
+
+def get_inputs():
+    x = torch.randn(batch_size, dim) * 5.0 + 10.0
+    return [x]
+
+
+def get_init_inputs():
+    return [10.0, 5.0]

S1/uucoco_#116/run_code.py

@@ -0,0 +1,77 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from zscoresigmoiddenormalize_torch import Model, get_inputs, get_init_inputs
+from zscoresigmoiddenormalize_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/uucoco_#116/zscoresigmoiddenormalize_cuda.py

@@ -0,0 +1,72 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+
+__global__ void zscore_sigmoid_denormalize_kernel(
+    const float* __restrict__ input,
+    float* __restrict__ output,
+    float mean,
+    float std,
+    int size
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < size) {
+        float x = input[idx];
+
+        // 1. Z-Score Normalization
+        float y = (x - mean) / std;
+
+        // 2. Sigmoid
+        float z = 1.0f / (1.0f + expf(-y));
+
+        // 3. Denormalization (Inverse Z-Score)
+        // Use fmaf for better precision/speed: z * std + mean
+        output[idx] = fmaf(z, std, mean);
+    }
+}
+
+torch::Tensor zscore_sigmoid_denormalize_cuda(torch::Tensor input, float mean, float std) {
+    auto output = torch::empty_like(input);
+    int size = input.numel();
+
+    const int block_size = 256;
+    int num_blocks = (size + block_size - 1) / block_size;
+
+    zscore_sigmoid_denormalize_kernel<<<num_blocks, block_size>>>(
+        input.data_ptr<float>(),
+        output.data_ptr<float>(),
+        mean,
+        std,
+        size
+    );
+
+    return output;
+}
+"""
+
+cpp_source = """
+torch::Tensor zscore_sigmoid_denormalize_cuda(torch::Tensor input, float mean, float std);
+"""
+
+module = load_inline(
+    name="zscore_sigmoid_denormalize",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["zscore_sigmoid_denormalize_cuda"],
+    verbose=True
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self, mean, std):
+        super(ModelNew, self).__init__()
+        self.mean = mean
+        self.std = std
+        self.module = module
+
+    def forward(self, x):
+        return self.module.zscore_sigmoid_denormalize_cuda(x, self.mean, self.std)

S1/uucoco_#116/zscoresigmoiddenormalize_torch.py

@@ -0,0 +1,28 @@
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self, mean, std):
+        super(Model, self).__init__()
+        self.mean = mean
+        self.std = std
+
+    def forward(self, x):
+        y = (x - self.mean) / self.std
+        z = torch.sigmoid(y)
+
+        return z * self.std + self.mean
+
+
+batch_size = 1024
+dim = 1024
+
+
+def get_inputs():
+    x = torch.randn(batch_size, dim) * 5.0 + 10.0
+    return [x]
+
+
+def get_init_inputs():
+    return [10.0, 5.0]