Merge pull request 'finish ColumnStack #65' (#624) from ZZZJ/GPUCodeForces:ColumnStack into main

S1/ZZZJ_#65/column_stack_cuda.py

@@ -0,0 +1,125 @@
+import torch
+from torch.utils.cpp_extension import load_inline
+
+stack_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+
+
+__global__ void column_stack_vec4_kernel(
+    const float* __restrict__ a, 
+    const float* __restrict__ b, 
+    float* __restrict__ output, 
+    int n_vec) 
+{
+
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (idx < n_vec) {
+        // 1. Vectorized Load
+        // Load 4 floats from A: [a0, a1, a2, a3]
+        const float4* a_ptr = reinterpret_cast<const float4*>(a);
+        float4 va = a_ptr[idx];
+        
+        // Load 4 floats from B: [b0, b1, b2, b3]
+        const float4* b_ptr = reinterpret_cast<const float4*>(b);
+        float4 vb = b_ptr[idx];
+        
+        // 2. Register Shuffle (Interleave)
+        // Target: [a0, b0, a1, b1], [a2, b2, a3, b3]
+        float4 out1;
+        out1.x = va.x; // a0
+        out1.y = vb.x; // b0
+        out1.z = va.y; // a1
+        out1.w = vb.y; // b1
+        
+        float4 out2;
+        out2.x = va.z; // a2
+        out2.y = vb.z; // b2
+        out2.z = va.w; // a3
+        out2.w = vb.w; // b3
+        
+        // 3. Vectorized Store
+        // Output index: idx * 2 (because we produce 2 float4s per idx)
+        float4* out_ptr = reinterpret_cast<float4*>(output);
+        out_ptr[idx * 2 + 0] = out1;
+        out_ptr[idx * 2 + 1] = out2;
+    }
+}
+
+
+__global__ void column_stack_scalar_kernel(
+    const float* __restrict__ a, 
+    const float* __restrict__ b, 
+    float* __restrict__ output, 
+    int n,
+    int offset) 
+{
+    int idx = blockIdx.x * blockDim.x + threadIdx.x + offset;
+    
+    if (idx < n) {
+        float val_a = a[idx];
+        float val_b = b[idx];
+        
+        // Output layout: [a0, b0, a1, b1 ...]
+        // idx -> 2*idx, 2*idx+1
+        output[idx * 2]     = val_a;
+        output[idx * 2 + 1] = val_b;
+    }
+}
+
+torch::Tensor column_stack_cuda(torch::Tensor a, torch::Tensor b) {
+    int N = a.size(0);
+    
+    // Output: [N, 2]
+    auto output = torch::empty({N, 2}, a.options());
+    
+    // 1. Vectorized Path (Process 4 elements at a time)
+    int vec_N = N / 4;
+    if (vec_N > 0) {
+        const int block = 256;
+        const int grid = (vec_N + block - 1) / block;
+        
+        column_stack_vec4_kernel<<<grid, block>>>(
+            a.data_ptr<float>(),
+            b.data_ptr<float>(),
+            output.data_ptr<float>(),
+            vec_N
+        );
+    }
+    
+    // 2. Scalar Tail
+    int remainder = N % 4;
+    if (remainder > 0) {
+        int offset = vec_N * 4;
+        column_stack_scalar_kernel<<<1, 32>>>(
+            a.data_ptr<float>(),
+            b.data_ptr<float>(),
+            output.data_ptr<float>(),
+            N,
+            offset
+        );
+    }
+    
+    return output;
+}
+"""
+
+cpp_source = "torch::Tensor column_stack_cuda(torch::Tensor a, torch::Tensor b);"
+
+stack_module = load_inline(
+    name="column_stack_extension",
+    cpp_sources=cpp_source,
+    cuda_sources=stack_source,
+    functions=["column_stack_cuda"],
+    verbose=True,
+    with_cuda=True
+)
+
+class ModelNew(torch.nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.cuda_op = stack_module
+
+    def forward(self, a, b):
+        return self.cuda_op.column_stack_cuda(a.contiguous(), b.contiguous())

S1/ZZZJ_#65/column_stack_torch.py

@@ -0,0 +1,22 @@
+import torch
+import torch.nn as nn
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
+  
+        return torch.column_stack((a, b))
+
+N = 1024 * 1024 * 16
+
+def get_inputs():
+    a = torch.randn(N, device='cuda', dtype=torch.float32)
+    b = torch.randn(N, device='cuda', dtype=torch.float32)
+    return [a, b]
+
+def get_init_inputs():
+    return []

S1/ZZZJ_#65/prompt.txt

@@ -0,0 +1,30 @@
+You write custom CUDA kernels to replace the pytorch operators in the given 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 matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.  
+  
+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
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
+  
+        return torch.column_stack((a, b))
+
+N = 1024 * 1024 * 16
+
+def get_inputs():
+    a = torch.randn(N, device='cuda', dtype=torch.float32)
+    b = torch.randn(N, device='cuda', dtype=torch.float32)
+    return [a, b]
+
+def get_init_inputs():
+    return []
+```

S1/ZZZJ_#65/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 column_stack_torch import Model,get_inputs,get_init_inputs
+from column_stack_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()