finish Smelu #48

S1/uucoco_#48/SmeLU_cuda.py

@@ -0,0 +1,97 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+class ModelNew(nn.Module):
+    def __init__(self, beta=2.0):
+        super().__init__()
+        self.beta = beta
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        #include <torch/extension.h>
+        torch::Tensor smelu_cuda(torch::Tensor x, float beta);
+        """
+
+        cuda_source = """
+        #include <cuda_runtime.h>
+
+        __device__ __forceinline__ float smelu_op(float x, float beta, float inv_4beta) {
+            float abs_x = fabsf(x);
+            if (abs_x < beta) {
+                float tmp = x + beta;
+                return tmp * tmp * inv_4beta;
+            } else {
+                return (x > 0.0f) ? x : 0.0f;
+            }
+        }
+
+        __global__ void smelu_kernel_vec4(
+            const float* __restrict__ x,
+            float* __restrict__ y,
+            int n,
+            float beta,
+            float inv_4beta) 
+        {
+            int idx = blockIdx.x * blockDim.x + threadIdx.x;
+            int stride = blockDim.x * gridDim.x;
+
+            int vec_n = n / 4;
+            const float4* x_vec = reinterpret_cast<const float4*>(x);
+            float4* y_vec = reinterpret_cast<float4*>(y);
+
+            for (int i = idx; i < vec_n; i += stride) {
+                float4 v = x_vec[i];
+                float4 out;
+
+                out.x = smelu_op(v.x, beta, inv_4beta);
+                out.y = smelu_op(v.y, beta, inv_4beta);
+                out.z = smelu_op(v.z, beta, inv_4beta);
+                out.w = smelu_op(v.w, beta, inv_4beta);
+
+                y_vec[i] = out;
+            }
+
+            int tail = vec_n * 4;
+            for (int i = tail + idx; i < n; i += stride) {
+                y[i] = smelu_op(x[i], beta, inv_4beta);
+            }
+        }
+
+        torch::Tensor smelu_cuda(torch::Tensor x, float beta) {
+            auto x_c = x.contiguous();
+            auto output = torch::empty_like(x_c);
+
+            int n = x_c.numel();
+            int threads = 256;
+            int blocks = (n / 4 + threads - 1) / threads;
+            if (blocks > 65535) blocks = 65535;
+            if (blocks == 0) blocks = 1;
+
+            float inv_4beta = 1.0f / (4.0f * beta);
+
+            smelu_kernel_vec4<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                n,
+                beta,
+                inv_4beta
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="smelu_opt_vec4",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["smelu_cuda"],
+            extra_cuda_cflags=["-O3", "--use_fast_math"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.smelu_cuda(x, self.beta)

S1/uucoco_#48/SmeLU_torch.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self, beta=2.0):
+        super().__init__()
+        self.beta = beta
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        abs_x = torch.abs(x)
+        return torch.where(
+            abs_x < self.beta,
+            torch.pow(x + self.beta, 2) / (4.0 * self.beta),
+            F.relu(x)
+        )
+
+batch_size = 128
+feature_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return [2.0]

S1/uucoco_#48/prompt.txt

@@ -0,0 +1,77 @@
+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 optimized Smooth ReLU (SmeLU) activation with:
+
+Memory Optimization:
+
+Vectorized memory access using float4 for 4x bandwidth
+
+Contiguous tensor inputs for coalesced memory access
+
+Separate handling for vectorized main loop and scalar tail
+
+Parallelization Strategy:
+
+Grid-stride loop for efficient workload distribution
+
+256 threads per block optimal configuration
+
+Automatic grid size calculation with 65535 block limit
+
+Computational Optimization:
+
+SmeLU activation with configurable beta parameter
+
+Precomputed reciprocal: inv_4beta = 1.0f / (4.0f * beta)
+
+Fast math compilation flags for optimized arithmetic
+
+Branching implementation:
+
+For |x| < beta: (x + beta)² / (4 * beta)
+
+For x ≥ beta: x
+
+For x ≤ -beta: 0
+
+Work Distribution:
+
+Vectorized main loop processes 4 elements per thread via float4
+
+Scalar tail handles remaining elements (n % 4)
+
+Each thread computes independent SmeLU operations
+
+The implementation provides maximum throughput through vectorization while maintaining the smooth transition characteristic of SmeLU activation with configurable beta parameter.
+
+
+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, beta=2.0):
+        super().__init__()
+        self.beta = beta
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        abs_x = torch.abs(x)
+        return torch.where(
+            abs_x < self.beta,
+            torch.pow(x + self.beta, 2) / (4.0 * self.beta),
+            F.relu(x)
+        )
+
+batch_size = 128
+feature_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return [2

S1/uucoco_#48/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 SmeLU_torch import Model, get_inputs, get_init_inputs
+from SmeLU_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()