finish PSmish#42

S1/uucoco_#42/PSMish_cuda.py

@@ -0,0 +1,79 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+class ModelNew(nn.Module):
+    def __init__(self, alpha: float = 1.0, beta: float = 1.0):
+        super().__init__()
+        self.alpha = alpha
+        self.beta = beta
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        torch::Tensor psmish_cuda(torch::Tensor x, float alpha, float beta);
+        """
+
+        cuda_source = """
+        #include <torch/extension.h>
+        #include <cuda_runtime.h>
+        #include <math.h>
+
+        __device__ __forceinline__ double psmish_op_double(double x, double alpha, double beta) {
+            double bx = beta * x;
+            double gate = tanh(log(1.0 + exp(bx)));
+            return alpha * x * gate;
+        }
+
+        __global__ void psmish_kernel_double_intermediate(
+            const float* __restrict__ x,
+            float* __restrict__ output,
+            const int n_elements,
+            const float alpha,
+            const float beta) 
+        {
+            const int block_start = blockIdx.x * blockDim.x;
+            const int block_end = min(block_start + blockDim.x, n_elements);
+
+            double alpha_d = (double)alpha;
+            double beta_d = (double)beta;
+
+            for (int i = block_start + threadIdx.x; i < block_end; i += blockDim.x) {
+                double val_d = (double)x[i];
+                double result_d = psmish_op_double(val_d, alpha_d, beta_d);
+                output[i] = (float)result_d;
+            }
+        }
+
+        torch::Tensor psmish_cuda(torch::Tensor x, float alpha, float beta) {
+            auto x_c = x.contiguous();
+            const int n_elements = x_c.numel();
+            auto output = torch::empty_like(x_c);
+
+            const int threads = 256;
+            const int max_blocks = 65535;
+            const int blocks = std::min((n_elements + threads - 1) / threads, max_blocks);
+
+            psmish_kernel_double_intermediate<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                n_elements,
+                alpha,
+                beta
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="psmish_v3",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["psmish_cuda"],
+            extra_cuda_cflags=["-O3"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.psmish_cuda(x, self.alpha, self.beta)

S1/uucoco_#42/PSMish_torch.py

@@ -0,0 +1,27 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self, alpha: float = 1.0, beta: float = 1.0):
+        super().__init__()
+        self.alpha = alpha
+        self.beta = beta
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate = torch.tanh(torch.log(1.0 + torch.exp(self.beta * x)))
+        return self.alpha * x * gate
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return [1.0, 1.0]

S1/uucoco_#42/prompt.txt

@@ -0,0 +1,44 @@
+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.
+
+
+This CUDA kernel implements a custom PSMish activation function with the following optimizations:
+Double Precision Intermediate Calculation: Uses doubleprecision for the core mathematical operations (tanh, log, exp) to maintain numerical stability and precision, then converts back to floatfor storage, balancing accuracy with memory efficiency.
+Tiled Kernel Design: Employs a block-based tiling approach where each thread processes elements within its assigned block, improving memory locality and cache efficiency.
+Mathematical Function: Implements the PSMish activation: α * x * tanh(ln(1 + exp(β * x))), combining scaling factors with smooth gating behavior.
+Memory Access Optimization: Uses __restrict__qualifiers and contiguous memory tensors to enable better compiler optimizations and reduce memory bank conflicts.
+Compiler Optimizations: Enabled with -O3flag for aggressive performance optimization of the generated code.
+Occupancy Optimization: Configures 256 threads per block and dynamically calculates grid size (up to 65535 blocks) to maximize GPU occupancy.
+Parameterized Activation: Supports learnable or configurable alphaand betaparameters passed directly to the CUDA kernel, allowing flexible activation behavior.
+Inlined Device Function: The core mathematical operation is marked with __forceinline__to eliminate function call overhead within the kernel.
+Numerical Stability: The use of double precision for intermediate calculations prevents precision loss in the complex exponential and logarithmic operations.
+
+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
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self, alpha: float = 1.0, beta: float = 1.0):
+        super().__init__()
+        self.alpha = alpha
+        self.beta = beta
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate = torch.tanh(torch.log(1.0 + torch.exp(self.beta * x)))
+        return self.alpha * x * gate
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return [1.0, 1.0]

S1/uucoco_#42/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 PSMish_torch import Model, get_inputs, get_init_inputs
+from PSMish_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()