finish exp-square #42

S1/Ljy123_#42/cudacode.py

@@ -0,0 +1,95 @@
+import torch
+from torch.utils.cpp_extension import load_inline
+
+source = """
+#include <torch/extension.h>
+
+__global__ void exp_square_affine_gate_kernel(const float* x, const float* scale, const float* bias, float* y, int B, int D, float alpha, float beta){
+    int b = blockIdx.x;
+    int tid = threadIdx.x;
+    int stride = blockDim.x;
+    int row_start = b * D;
+    int aligned = ((row_start & 3) == 0);
+    if(aligned){
+        int D4 = (D / 4) * 4;
+        #pragma unroll 4
+        for(int i = tid * 4; i < D4; i += stride * 4){
+            int base = row_start + i;
+            float4 xv = reinterpret_cast<const float4*>(x)[base / 4];
+            float4 sv = reinterpret_cast<const float4*>(scale)[i / 4];
+            float4 bv = reinterpret_cast<const float4*>(bias)[i / 4];
+            float z0 = fmaf(xv.x, sv.x, bv.x);
+            float z1 = fmaf(xv.y, sv.y, bv.y);
+            float z2 = fmaf(xv.z, sv.z, bv.z);
+            float z3 = fmaf(xv.w, sv.w, bv.w);
+            float s0 = expf(z0 * z0);
+            float s1 = expf(z1 * z1);
+            float s2 = expf(z2 * z2);
+            float s3 = expf(z3 * z3);
+            float4 yv;
+            yv.x = xv.x * (1.0f / (1.0f + expf(-(alpha * s0 + beta))));
+            yv.y = xv.y * (1.0f / (1.0f + expf(-(alpha * s1 + beta))));
+            yv.z = xv.z * (1.0f / (1.0f + expf(-(alpha * s2 + beta))));
+            yv.w = xv.w * (1.0f / (1.0f + expf(-(alpha * s3 + beta))));
+            reinterpret_cast<float4*>(y)[base / 4] = yv;
+        }
+        #pragma unroll 4
+        for(int i = D4 + tid; i < D; i += stride){
+            int base = row_start + i;
+            float z = fmaf(x[base], scale[i], bias[i]);
+            float s = expf(z * z);
+            float g = 1.0f / (1.0f + expf(-(alpha * s + beta)));
+            y[base] = x[base] * g;
+        }
+    } else {
+        #pragma unroll 4
+        for(int i = tid; i < D; i += stride){
+            int base = row_start + i;
+            float z = fmaf(x[base], scale[i], bias[i]);
+            float s = expf(z * z);
+            float g = 1.0f / (1.0f + expf(-(alpha * s + beta)));
+            y[base] = x[base] * g;
+        }
+    }
+}
+
+torch::Tensor exp_square_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias, torch::Tensor alpha, torch::Tensor beta){
+    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);
+    float a = alpha.item<float>();
+    float be = beta.item<float>();
+    int block = 1024;
+    int grid = B;
+    exp_square_affine_gate_kernel<<<grid, block>>>(xc.data_ptr<float>(), sc.data_ptr<float>(), bc.data_ptr<float>(), y.data_ptr<float>(), B, D, a, be);
+    return y;
+}
+"""
+
+cpp_source = """
+torch::Tensor exp_square_affine_gate_cuda(torch::Tensor x, torch::Tensor scale, torch::Tensor bias, torch::Tensor alpha, torch::Tensor beta);
+"""
+
+ops = load_inline(
+    name="exp_square_affine_gate",
+    cpp_sources=cpp_source,
+    cuda_sources=source,
+    functions=["exp_square_affine_gate_cuda"],
+    extra_cuda_cflags=["-O3","--use_fast_math"],
+    verbose=True
+)
+
+class ModelNew(torch.nn.Module):
+    def __init__(self, scale: torch.Tensor, bias: torch.Tensor, alpha: float, beta: float):
+        super(ModelNew, self).__init__()
+        self.ops = ops
+        self.register_buffer("scale", scale)
+        self.register_buffer("bias", bias)
+        self.register_buffer("alpha", torch.tensor(float(alpha), dtype=torch.float32))
+        self.register_buffer("beta", torch.tensor(float(beta), dtype=torch.float32))
+
+    def forward(self, x: torch.Tensor):
+        return self.ops.exp_square_affine_gate_cuda(x, self.scale, self.bias, self.alpha, self.beta)

S1/Ljy123_#42/prompt.txt

@@ -0,0 +1,2 @@
+You write custom CUDA kernels to replace PyTorch operators for speedups.
+Implement Exp-Square Affine Gate on x[B,D]: Compute z = x*scale + bias, s = exp(z^2), gate g = sigmoid(alpha*s + beta), output y = x * g. Fuse affine, square, exp, sigmoid, and multiplication in a single grid-stride kernel. Provide a PyTorch reference with nn.Parameters for scale, bias, alpha, beta. Accuracy rtol=1e-3.

S1/Ljy123_#42/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 Exp-Square-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_#42/torchcode.py

@@ -0,0 +1,29 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self, scale: torch.Tensor, bias: torch.Tensor, alpha: float, beta: float):
+        super(Model, self).__init__()
+        self.scale = nn.Parameter(scale)
+        self.bias = nn.Parameter(bias)
+        self.alpha = nn.Parameter(torch.tensor(float(alpha), dtype=torch.float32))
+        self.beta = nn.Parameter(torch.tensor(float(beta), dtype=torch.float32))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        z = x * self.scale.view(1,-1) + self.bias.view(1,-1)
+        s = torch.exp(z * z)
+        g = torch.sigmoid(self.alpha * s + self.beta)
+        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)
+    alpha = 1.0
+    beta = 0.0
+    return [scale, bias, alpha, beta]