diff --git a/S1/28/mseloss_cuda.py b/S1/28/mseloss_cuda.py
new file mode 100644
index 0000000..8a54816
--- /dev/null
+++ b/S1/28/mseloss_cuda.py
@@ -0,0 +1,178 @@
+# mseloss_cuda.py
+import torch
+from torch.utils.cpp_extension import load_inline
+
+class ModelNew(torch.nn.Module):
+    
+    def __init__(self):
+        super().__init__()
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        #include <torch/extension.h>
+
+        torch::Tensor mse_forward_cuda(torch::Tensor pred, torch::Tensor target);
+        """
+
+        cuda_source = """
+        #include <cuda_runtime.h>
+        #include <device_launch_parameters.h>
+
+        #define BLOCK_SIZE 256
+        #define VEC_SIZE 4
+        #define WARP_SIZE 32
+
+        // Warp-level reduction using shuffle instructions (faster than shared memory)
+        __device__ __forceinline__ float warp_reduce_sum(float val) {
+            #pragma unroll
+            for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {
+                val += __shfl_down_sync(0xffffffff, val, offset);
+            }
+            return val;
+        }
+
+        // Optimized kernel: use float accumulation, warp primitives, and better memory access
+        __global__ void mse_optimized_kernel(
+            const float* __restrict__ pred,
+            const float* __restrict__ target,
+            float* __restrict__ output_sum,
+            int N_elements
+        ) {
+            // Use float instead of double for accumulation (much faster on GPU)
+            float thread_sum = 0.0f;
+            
+            int N_vec = N_elements / VEC_SIZE;
+            int grid_stride_vec = gridDim.x * blockDim.x;
+
+            const float4* __restrict__ pred4 = reinterpret_cast<const float4*>(pred);
+            const float4* __restrict__ target4 = reinterpret_cast<const float4*>(target);
+
+            // Process vectorized data with grid-stride loop
+            for (int idx_vec = blockIdx.x * blockDim.x + threadIdx.x;
+                 idx_vec < N_vec; 
+                 idx_vec += grid_stride_vec)
+            {
+                float4 p4 = __ldg(&pred4[idx_vec]);    // Use read-only cache
+                float4 t4 = __ldg(&target4[idx_vec]);
+
+                // Compute differences and accumulate squared errors
+                float diff1 = p4.x - t4.x;
+                float diff2 = p4.y - t4.y;
+                float diff3 = p4.z - t4.z;
+                float diff4 = p4.w - t4.w;
+                
+                // Use fused multiply-add for better performance
+                thread_sum = fmaf(diff1, diff1, thread_sum);
+                thread_sum = fmaf(diff2, diff2, thread_sum);
+                thread_sum = fmaf(diff3, diff3, thread_sum);
+                thread_sum = fmaf(diff4, diff4, thread_sum);
+            }
+
+            // Warp-level reduction (no shared memory needed for this step)
+            thread_sum = warp_reduce_sum(thread_sum);
+
+            // Use shared memory only for inter-warp reduction
+            __shared__ float warp_sums[BLOCK_SIZE / WARP_SIZE];
+            
+            int lane = threadIdx.x % WARP_SIZE;
+            int warp_id = threadIdx.x / WARP_SIZE;
+
+            if (lane == 0) {
+                warp_sums[warp_id] = thread_sum;
+            }
+            __syncthreads();
+
+            // Final reduction by first warp only
+            if (warp_id == 0) {
+                thread_sum = (threadIdx.x < BLOCK_SIZE / WARP_SIZE) ? warp_sums[lane] : 0.0f;
+                thread_sum = warp_reduce_sum(thread_sum);
+                
+                if (threadIdx.x == 0) {
+                    output_sum[blockIdx.x] = thread_sum;
+                }
+            }
+        }
+
+        // Two-phase reduction kernel for final sum
+        __global__ void final_reduction_kernel(
+            const float* __restrict__ partial_sums,
+            float* __restrict__ output,
+            int n
+        ) {
+            __shared__ float sh_sum[BLOCK_SIZE];
+            
+            float sum = 0.0f;
+            for (int i = threadIdx.x; i < n; i += blockDim.x) {
+                sum += partial_sums[i];
+            }
+            
+            sh_sum[threadIdx.x] = sum;
+            __syncthreads();
+
+            // Tree reduction in shared memory
+            #pragma unroll
+            for (int s = BLOCK_SIZE / 2; s > 0; s /= 2) {
+                if (threadIdx.x < s) {
+                    sh_sum[threadIdx.x] += sh_sum[threadIdx.x + s];
+                }
+                __syncthreads();
+            }
+
+            if (threadIdx.x == 0) {
+                output[0] = sh_sum[0];
+            }
+        }
+
+        torch::Tensor mse_forward_cuda(torch::Tensor pred, torch::Tensor target) {
+            TORCH_CHECK(pred.is_cuda() && target.is_cuda(), "Inputs must be CUDA tensors");
+            
+            pred = pred.contiguous();
+            target = target.contiguous();
+
+            int N_elements = pred.numel();
+            TORCH_CHECK(N_elements % VEC_SIZE == 0, 
+                       "Total elements must be divisible by VEC_SIZE (4)");
+
+            // Adaptive grid size based on data size and GPU occupancy
+            const int block_size = BLOCK_SIZE;
+            int num_sms;
+            cudaDeviceGetAttribute(&num_sms, cudaDevAttrMultiProcessorCount, 0);
+            const int grid_size = min(num_sms * 4, (N_elements / VEC_SIZE + block_size - 1) / block_size);
+
+            auto partial_sum = torch::empty({grid_size}, pred.options());
+            auto final_result = torch::empty({1}, pred.options());
+
+            // Launch main kernel
+            mse_optimized_kernel<<<grid_size, block_size>>>(
+                pred.data_ptr<float>(),
+                target.data_ptr<float>(),
+                partial_sum.data_ptr<float>(),
+                N_elements
+            );
+
+            // Launch final reduction kernel (all on GPU, no CPU sync)
+            final_reduction_kernel<<<1, block_size>>>(
+                partial_sum.data_ptr<float>(),
+                final_result.data_ptr<float>(),
+                grid_size
+            );
+
+            // Divide by N to get mean (done on GPU)
+            final_result.div_(N_elements);
+
+            return final_result;
+        }
+        """
+        
+        self.mse_op = load_inline(
+            name="mse_optimized_op",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["mse_forward_cuda"],
+            extra_cuda_cflags=["-O3", "--use_fast_math", "-lineinfo"],
+            verbose=True
+        )
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        return self.mse_op.mse_forward_cuda(pred, target)
\ No newline at end of file
diff --git a/S1/28/mseloss_torch.py b/S1/28/mseloss_torch.py
new file mode 100644
index 0000000..2e15f4f
--- /dev/null
+++ b/S1/28/mseloss_torch.py
@@ -0,0 +1,23 @@
+# mseloss_torch.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+BATCH_SIZE = 16
+DIM = 16384 * 16 
+
+class Model(nn.Module):
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        # F.mse_loss 默认返回 mean reduction
+        return F.mse_loss(pred, target, reduction='mean') 
+
+def get_inputs():
+
+    pred = torch.randn(BATCH_SIZE, DIM, dtype=torch.float32)
+    target = pred + torch.rand_like(pred) * 0.1 
+    return [pred, target]
+
+def get_init_inputs():
+    return []
\ No newline at end of file
diff --git a/S1/28/prompt.txt b/S1/28/prompt.txt
new file mode 100644
index 0000000..73ff4bd
--- /dev/null
+++ b/S1/28/prompt.txt
@@ -0,0 +1,29 @@
+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
+import torch.nn.functional as F
+
+
+BATCH_SIZE = 16
+DIM = 16384 * 16 
+
+class Model(nn.Module):
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        # F.mse_loss 默认返回 mean reduction
+        return F.mse_loss(pred, target, reduction='mean') 
+
+def get_inputs():
+
+    pred = torch.randn(BATCH_SIZE, DIM, dtype=torch.float32)
+    target = pred + torch.rand_like(pred) * 0.1 
+    return [pred, target]
+
+def get_init_inputs():
+    return []
\ No newline at end of file
diff --git a/S1/28/run_code.py b/S1/28/run_code.py
new file mode 100644
index 0000000..0f531a7
--- /dev/null
+++ b/S1/28/run_code.py
@@ -0,0 +1,88 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from mseloss_torch import Model, get_inputs, get_init_inputs
+from mseloss_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)
+
+    # 更严格的精度检查
+    abs_diff = (output_torch - output_cuda).abs()
+    max_diff = abs_diff.max().item()
+    mean_diff = abs_diff.mean().item()
+    
+    print(f"最大差异: {max_diff:.6f}")
+    print(f"平均差异: {mean_diff:.6f}")
+    
+    precision_flag = torch.allclose(output_torch, output_cuda, rtol=1e-05, atol=1e-05)
+    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 = 1000  # 增加迭代次数以获得更准确的时间测量
+    
+    # Warm up
+    for _ in range(100):
+        _ = torch_model(*inputs)
+        _ = cuda_model(*inputs)
+    
+    # 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 (matmul + relu) 平均执行时间: {torch_time:.6f} 秒")
+    print(f"自定义 CUDA ReLU 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()
\ No newline at end of file