finish logsigmoid (affine-gate) #41

S1/Ljy123_#41/cudacode.py

@@ -0,0 +1,55 @@
+import torch
+from torch.utils.cpp_extension import load_inline
+
+source = """
+#include <torch/extension.h>
+
+__global__ void logsigmoid_affine_gate_kernel(const float* x, const float* scale, const float* bias, float* y, int B, int D){
+    long long tid = blockIdx.x * blockDim.x + threadIdx.x;
+    long long stride = blockDim.x * gridDim.x;
+    long long total = (long long)B * D;
+    for(long long i = tid; i < total; i += stride){
+        int b = (int)(i / D);
+        int d = (int)(i % D);
+        float z = x[i] * scale[d] + bias[d];
+        float g = -log1pf(expf(-z));
+        y[i] = x[i] * g;
+    }
+}
+
+torch::Tensor logsigmoid_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias){
+    auto xc = x.contiguous();
+    auto sc = scale.contiguous();
+    auto bc = bias.contiguous();
+    auto y = torch::empty_like(xc);
+    int B = (int)xc.size(0);
+    int D = (int)xc.size(1);
+    int block = 1024;
+    int grid = (int)((long long)B * D / block);
+    if(grid < 1) grid = 1; if(grid > 65535) grid = 65535;
+    logsigmoid_affine_gate_kernel<<<grid, block>>>(xc.data_ptr<float>(), sc.data_ptr<float>(), bc.data_ptr<float>(), y.data_ptr<float>(), B, D);
+    return y;
+}
+"""
+
+cpp_source = """
+torch::Tensor logsigmoid_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias);
+"""
+
+ops = load_inline(
+    name="logsigmoid_affine_gate",
+    cpp_sources=cpp_source,
+    cuda_sources=source,
+    functions=["logsigmoid_affine_gate_cuda"],
+    verbose=True
+)
+
+class ModelNew(torch.nn.Module):
+    def __init__(self, scale: torch.Tensor, bias: torch.Tensor):
+        super(ModelNew, self).__init__()
+        self.ops = ops
+        self.register_buffer("scale", scale)
+        self.register_buffer("bias", bias)
+
+    def forward(self, x: torch.Tensor):
+        return self.ops.logsigmoid_affine_gate_cuda(x, self.scale, self.bias)

S1/Ljy123_#41/prompt.txt

@@ -0,0 +1,2 @@
+You write custom CUDA kernels to replace PyTorch operators for speedups.
+Implement LogSigmoid Affine Gate: For x[B,D], per-dim scale[D] and bias[D], compute z = x*scale + bias, gate g = logsigmoid(z) = -log(1+exp(-z)), and output y = x * g. Use a single grid-stride kernel to fuse affine, logsigmoid, and gating into one pass. Provide a PyTorch reference using nn.Parameters. Accuracy rtol=1e-3.

S1/Ljy123_#41/run_code.py

@@ -0,0 +1,52 @@
+import torch
+import time
+from torchcode import Model, get_inputs, get_init_inputs
+from cudacode import ModelNew
+
+def run_benchmark():
+    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
+    device = torch.device("cuda")
+
+    init_inputs = [x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in get_init_inputs()]
+    inputs = [x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in get_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():
+        out_torch = torch_model(*inputs)
+        out_cuda = cuda_model(*inputs)
+    flag = torch.allclose(out_torch, out_cuda, rtol=1e-03)
+    if 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(f"最大绝对误差: {(out_torch - out_cuda).abs().max().item()}" )
+
+    print("\n-------------------- 性能加速比测试 --------------------")
+    iters = 100
+    torch.cuda.synchronize(); t0 = time.time()
+    for _ in range(iters):
+        _ = torch_model(*inputs)
+    torch.cuda.synchronize(); t_torch = (time.time() - t0) / iters
+
+    torch.cuda.synchronize(); t0 = time.time()
+    for _ in range(iters):
+        _ = cuda_model(*inputs)
+    torch.cuda.synchronize(); t_cuda = (time.time() - t0) / iters
+
+    print(f"PyTorch LogSigmoid-Affine-Gate 平均执行时间: {t_torch:.6f} 秒")
+    print(f"自定义 CUDA 融合内核 平均执行时间: {t_cuda:.6f} 秒")
+    sp = t_torch / t_cuda if t_cuda > 0 else 0
+    if t_cuda > 0:
+        print(f"加速比 (Speedup): {sp:.2f}x")
+    else:
+        print("CUDA 内核执行时间为0，无法计算加速比。")
+    return flag, sp
+
+if __name__ == "__main__":
+    run_benchmark()

S1/Ljy123_#41/torchcode.py

@@ -0,0 +1,24 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self, scale: torch.Tensor, bias: torch.Tensor):
+        super(Model, self).__init__()
+        self.scale = nn.Parameter(scale)
+        self.bias = nn.Parameter(bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        z = x * self.scale.view(1,-1) + self.bias.view(1,-1)
+        g = torch.nn.functional.logsigmoid(z)
+        return x * g
+
+B, D = 64, 8192
+
+def get_inputs():
+    x = torch.randn(B, D)
+    return [x]
+
+def get_init_inputs():
+    scale = torch.randn(D)
+    bias = torch.randn(D)
+    return [scale, bias]