finish Channel-var-gate #86

S1/gsd123_#86/channelvargate_cuda.py

@@ -0,0 +1,124 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        torch::Tensor channel_var_gate_cuda(torch::Tensor x);
+        """
+
+        cuda_source = """
+        #include <torch/extension.h>
+        #include <cuda_runtime.h>
+        #include <math.h>
+
+        __device__ __forceinline__ float sigmoid_op(float x) {
+            if (x >= 0.0f) {
+                return 1.0f / (1.0f + expf(-x));
+            } else {
+                float z = expf(x);
+                return z / (1.0f + z);
+            }
+        }
+
+        __global__ void channel_var_gate_kernel(
+            const float* __restrict__ x,
+            float* __restrict__ output,
+            const int batch_size,
+            const int channels)
+        {
+            extern __shared__ float shared_mem[];
+            float* s_sum = shared_mem;
+            float* s_sum_sq = &shared_mem[blockDim.x];
+
+            const int b = blockIdx.x;
+            const int tid = threadIdx.x;
+
+            if (b >= batch_size) return;
+
+            const float* x_batch = x + b * channels;
+            float* output_batch = output + b * channels;
+
+            float local_sum = 0.0f;
+            float local_sum_sq = 0.0f;
+
+            for (int c = tid; c < channels; c += blockDim.x) {
+                float val = x_batch[c];
+                local_sum += val;
+            }
+
+            s_sum[tid] = local_sum;
+            __syncthreads();
+
+            for (int stride = blockDim.x / 2; stride > 0; stride >>= 1) {
+                if (tid < stride) {
+                    s_sum[tid] += s_sum[tid + stride];
+                }
+                __syncthreads();
+            }
+
+            float mean = s_sum[0] / channels;
+            __syncthreads();
+
+            for (int c = tid; c < channels; c += blockDim.x) {
+                float diff = x_batch[c] - mean;
+                local_sum_sq += diff * diff;
+            }
+
+            s_sum_sq[tid] = local_sum_sq;
+            __syncthreads();
+
+            for (int stride = blockDim.x / 2; stride > 0; stride >>= 1) {
+                if (tid < stride) {
+                    s_sum_sq[tid] += s_sum_sq[tid + stride];
+                }
+                __syncthreads();
+            }
+
+            float var = s_sum_sq[0] / channels;
+            float gate = sigmoid_op(var);
+
+            for (int c = tid; c < channels; c += blockDim.x) {
+                output_batch[c] = x_batch[c] * gate;
+            }
+        }
+
+        torch::Tensor channel_var_gate_cuda(torch::Tensor x) {
+            auto x_c = x.contiguous();
+            const int batch_size = x_c.size(0);
+            const int channels = x_c.size(1);
+
+            auto output = torch::empty_like(x_c);
+
+            const int threads = 256;
+            const int blocks = batch_size;
+            const int shared_mem_size = threads * 2 * sizeof(float);
+
+            channel_var_gate_kernel<<<blocks, threads, shared_mem_size>>>(
+                x_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                batch_size,
+                channels
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="channel_var_gate_op",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["channel_var_gate_cuda"],
+            extra_cuda_cflags=["-O3"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.channel_var_gate_cuda(x)

S1/gsd123_#86/channelvargate_torch.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        var = torch.var(x, dim=1, keepdim=True)
+        gate = torch.sigmoid(var)
+        return x * gate
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return []

S1/gsd123_#86/prompt.txt

@@ -0,0 +1,48 @@
+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.
+Custom CUDA kernel extension via torch.utils.cpp_extension.load_inline
+
+Channel variance computation with two-pass reduction (mean → variance)
+
+Shared memory optimization for intermediate sum and squared sum storage
+
+Parallel reduction in shared memory using tree-based approach
+
+Per-channel gate application based on channel variance
+
+One CUDA block per batch element for batch-parallel processing
+
+Contiguous tensor handling for memory coalescing
+
+Numerically stable sigmoid implementation (separated positive/negative cases)
+
+Efficient memory reuse with in-place-like output allocation
+
+
+
+
+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):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        var = torch.var(x, dim=1, keepdim=True)
+        gate = torch.sigmoid(var)
+        return x * gate
+
+
+batch_size = 128
+feature_dim = 512
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return []

S1/gsd123_#86/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 channelvargate_torch import Model, get_inputs, get_init_inputs
+from channelvargate_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()