Merge pull request 'finish SegmentReduce #69' (#628) from ZZZJ/GPUCodeForces:SegmentReduce into main

S1/ZZZJ_#69/prompt.txt

@@ -0,0 +1,45 @@
+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
+
+
+NUM_SEGMENTS = 2048
+AVG_LEN = 128 
+FEATURE_DIM = 32 
+
+class SegmentReduce(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.reduce_op = 'sum'
+
+    def forward(self, values, lengths):
+   
+        return torch.segment_reduce(values, self.reduce_op, lengths=lengths)
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.op = SegmentReduce()
+
+    def forward(self, v, l):
+        return self.op(v, l)
+
+def get_inputs():
+
+    lengths = torch.randint(1, 2 * AVG_LEN, (NUM_SEGMENTS,), dtype=torch.int32).cuda()
+
+    total_len = lengths.sum().item()
+
+    values = torch.randint(0, 5, (total_len, FEATURE_DIM), dtype=torch.float32).cuda()
+    
+    return [values, lengths]
+
+def get_init_inputs():
+    return []
+```

S1/ZZZJ_#69/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 segment_reduce_torch import Model,get_inputs,get_init_inputs
+from segment_reduce_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/ZZZJ_#69/segment_reduce_cuda.py

@@ -0,0 +1,107 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+from segment_reduce_torch import NUM_SEGMENTS, FEATURE_DIM
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        
+        macros = f"""
+        #define WARP_SIZE 32
+        #define FEATURE_DIM {FEATURE_DIM}
+        """
+        
+        cpp_source = """
+        #include <torch/extension.h>
+        torch::Tensor segment_reduce_cuda(torch::Tensor values, torch::Tensor offsets, int num_segments);
+        """
+
+        cuda_source = f"""
+        #include <cuda_runtime.h>
+        
+        {macros}
+
+        __device__ __forceinline__ float warpReduceSum(float val) {{
+            for (int offset = WARP_SIZE / 2; offset > 0; offset /= 2) {{
+                val += __shfl_down_sync(0xffffffff, val, offset);
+            }}
+            return val;
+        }}
+
+       
+        __global__ void segment_reduce_kernel_features(
+            const float* __restrict__ values,
+            const int* __restrict__ offsets,
+            float* __restrict__ output,
+            int num_segments
+        ) {{
+            int warp_id = threadIdx.x / WARP_SIZE;
+            int lane_id = threadIdx.x % WARP_SIZE;
+            
+            int warps_per_block = blockDim.x / WARP_SIZE;
+            int segment_idx = blockIdx.x * warps_per_block + warp_id;
+
+            if (segment_idx >= num_segments) return;
+            
+            // Boundary Check for Feature Dim
+            if (lane_id >= FEATURE_DIM) return;
+
+            int start = offsets[segment_idx];
+            int end = offsets[segment_idx + 1];
+            
+            float sum = 0.0f;
+            
+            // Coalesced Read: Each thread reads column `lane_id`
+            for (int r = start; r < end; ++r) {{
+                sum += values[r * FEATURE_DIM + lane_id];
+            }}
+
+            // Write Result
+            // [Fix] Removed f-string conflict in other places just in case, though usually output access is array style []
+            output[segment_idx * FEATURE_DIM + lane_id] = sum;
+        }}
+
+        torch::Tensor segment_reduce_cuda(torch::Tensor values, torch::Tensor offsets, int num_segments) {{
+            TORCH_CHECK(values.is_cuda(), "values must be CUDA");
+            TORCH_CHECK(offsets.is_cuda(), "offsets must be CUDA");
+            
+            // [Fix] Use double curly braces {{ }} for C++ initializer list in f-string
+            auto output = torch::empty({{num_segments, FEATURE_DIM}}, values.options());
+            
+            int threads = 256;
+            int warps_per_block = threads / 32;
+            int blocks = (num_segments + warps_per_block - 1) / warps_per_block;
+            
+            segment_reduce_kernel_features<<<blocks, threads>>>(
+                values.data_ptr<float>(),
+                offsets.data_ptr<int>(),
+                output.data_ptr<float>(),
+                num_segments
+            );
+            
+            return output;
+        }}
+        """
+
+        self.op = load_inline(
+            name='segment_reduce_feat_opt_fix',
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=['segment_reduce_cuda'],
+            extra_cuda_cflags=['-O3'],
+            verbose=False
+        )
+
+    def forward(self, values, lengths):
+        # Precompute offsets for CUDA kernel
+        offsets = torch.zeros(lengths.size(0) + 1, dtype=torch.int32, device=values.device)
+        offsets[1:] = torch.cumsum(lengths, dim=0)
+        
+        num_segments = lengths.size(0)
+        return self.op.segment_reduce_cuda(values, offsets, num_segments)

S1/ZZZJ_#69/segment_reduce_torch.py

@@ -0,0 +1,37 @@
+import torch
+import torch.nn as nn
+
+
+NUM_SEGMENTS = 2048
+AVG_LEN = 128 
+FEATURE_DIM = 32 
+
+class SegmentReduce(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.reduce_op = 'sum'
+
+    def forward(self, values, lengths):
+   
+        return torch.segment_reduce(values, self.reduce_op, lengths=lengths)
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.op = SegmentReduce()
+
+    def forward(self, v, l):
+        return self.op(v, l)
+
+def get_inputs():
+
+    lengths = torch.randint(1, 2 * AVG_LEN, (NUM_SEGMENTS,), dtype=torch.int32).cuda()
+
+    total_len = lengths.sum().item()
+
+    values = torch.randint(0, 5, (total_len, FEATURE_DIM), dtype=torch.float32).cuda()
+    
+    return [values, lengths]
+
+def get_init_inputs():
+    return []