Merge pull request 'finish wasserstein_layernorm #166' (#906) from gsd123/GPUCodeForces:gsd166 into main

S1/gsd123_#166/prompt.txt

@@ -0,0 +1,46 @@
+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.
+
+CUDA C++ kernel for 1‑D Wasserstein‑1 distance computation via sorting
+
+Block‑parallel per‑sample processing: each block handles one batch element
+
+Bitonic sort implemented in shared memory for both input vectors
+
+Shared memory allocation for two vectors of size dim
+
+Element‑wise absolute difference after sorting (matching order statistics)
+
+Post‑processing with PyTorch LayerNorm on the difference map
+
+PyTorch inline C++/CUDA extension via load_inline
+
+
+
+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, input_dim):
+        super(Model, self).__init__()
+        self.ln = nn.LayerNorm(input_dim)
+
+    def forward(self, x, y):
+        x_sorted, _ = torch.sort(x, dim=1)
+        y_sorted, _ = torch.sort(y, dim=1)
+        diff = torch.abs(x_sorted - y_sorted)
+        norm_diff = self.ln(diff)
+        return norm_diff.mean()
+
+batch_size = 16
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    y = torch.randn(batch_size, input_dim)
+    return [x, y]
+
+def get_init_inputs():
+    return [input_dim]

S1/gsd123_#166/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 wasserstein_layernorm_torch import Model, get_inputs, get_init_inputs
+from wasserstein_layernorm_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/gsd123_#166/wasserstein_layernorm_cuda.py

@@ -0,0 +1,94 @@
+import os
+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>
+
+__device__ void bitonic_sort(float* data, int n, int tid) {
+    for (int k = 2; k <= n; k <<= 1) {
+        for (int j = k >> 1; j > 0; j >>= 1) {
+            int ixj = tid ^ j;
+            if (ixj > tid) {
+                if ((tid & k) == 0) {
+                    if (data[tid] > data[ixj]) {
+                        float temp = data[tid];
+                        data[tid] = data[ixj];
+                        data[ixj] = temp;
+                    }
+                } else {
+                    if (data[tid] < data[ixj]) {
+                        float temp = data[tid];
+                        data[tid] = data[ixj];
+                        data[ixj] = temp;
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+}
+
+__global__ void wasserstein_diff_kernel(const float* __restrict__ x, 
+                                        const float* __restrict__ y, 
+                                        float* __restrict__ out, 
+                                        int dim) {
+    int bid = blockIdx.x;
+    int tid = threadIdx.x;
+
+    extern __shared__ float s_mem[];
+    float* s_x = s_mem;
+    float* s_y = s_mem + dim;
+
+    s_x[tid] = x[bid * dim + tid];
+    s_y[tid] = y[bid * dim + tid];
+    __syncthreads();
+
+    bitonic_sort(s_x, dim, tid);
+    bitonic_sort(s_y, dim, tid);
+
+    out[bid * dim + tid] = fabsf(s_x[tid] - s_y[tid]);
+}
+
+torch::Tensor wasserstein_diff_cuda(torch::Tensor x, torch::Tensor y) {
+    int batch_size = x.size(0);
+    int dim = x.size(1);
+
+    auto out = torch::empty_like(x);
+    int shared_mem = 2 * dim * sizeof(float);
+
+    wasserstein_diff_kernel<<<batch_size, dim, shared_mem>>>(
+        x.data_ptr<float>(),
+        y.data_ptr<float>(),
+        out.data_ptr<float>(),
+        dim
+    );
+
+    return out;
+}
+"""
+
+cpp_source = "torch::Tensor wasserstein_diff_cuda(torch::Tensor x, torch::Tensor y);"
+
+module = load_inline(
+    name="wasserstein_layernorm_ext",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["wasserstein_diff_cuda"],
+    verbose=False,
+    with_cuda=True
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self, input_dim):
+        super(ModelNew, self).__init__()
+        self.ln = nn.LayerNorm(input_dim)
+        self.op = module
+
+    def forward(self, x, y):
+        diff = self.op.wasserstein_diff_cuda(x.contiguous(), y.contiguous())
+        norm_diff = self.ln(diff)
+        return norm_diff.mean()

S1/gsd123_#166/wasserstein_layernorm_torch.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self, input_dim):
+        super(Model, self).__init__()
+        self.ln = nn.LayerNorm(input_dim)
+
+    def forward(self, x, y):
+        x_sorted, _ = torch.sort(x, dim=1)
+        y_sorted, _ = torch.sort(y, dim=1)
+        diff = torch.abs(x_sorted - y_sorted)
+        norm_diff = self.ln(diff)
+        return norm_diff.mean()
+
+batch_size = 16
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    y = torch.randn(batch_size, input_dim)
+    return [x, y]
+
+def get_init_inputs():
+    return [input_dim]