finish Schur-affine #98

S1/gsd123_#98/prompt.txt

@@ -0,0 +1,44 @@
+You write custom CUDA kernels to replace the pytorch operators in the given GeGLU 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 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
+
+Element-wise affine transformation: y = scale * x + bias
+
+Element-wise parallelization using CUDA grid-stride loops
+
+Per-feature learnable parameters (scale, bias)
+
+Contiguous tensor handling for all input tensors
+
+Memory-efficient in-place-like computation with torch.empty_like
+
+Fused multiply-add operations with __fmul_rn and __fadd_rn for precision
+
+Auto-tuning block/grid size based on tensor size (up to 65535 blocks)
+
+
+
+
+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, num_features=512):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(1, num_features))
+        self.bias = nn.Parameter(torch.zeros(1, num_features))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * self.scale + self.bias
+
+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_#98/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 schuraffine_torch import Model, get_inputs, get_init_inputs
+from schuraffine_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/gsd123_#98/schuraffine_cuda.py

@@ -0,0 +1,93 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+
+class ModelNew(nn.Module):
+    def __init__(self, num_features=512):
+        super().__init__()
+        self.num_features = num_features
+        self.scale = nn.Parameter(torch.ones(num_features))
+        self.bias = nn.Parameter(torch.zeros(num_features))
+        self._compile_cuda_kernel()
+
+    def _compile_cuda_kernel(self):
+        cpp_source = """
+        torch::Tensor schuraffine_cuda(
+            torch::Tensor x,
+            torch::Tensor scale,
+            torch::Tensor bias);
+        """
+
+        cuda_source = """
+        #include <torch/extension.h>
+        #include <cuda_runtime.h>
+        #include <math.h>
+
+        __global__ void schuraffine_kernel(
+            const float* __restrict__ x,
+            const float* __restrict__ scale,
+            const float* __restrict__ bias,
+            float* __restrict__ output,
+            const int rows,
+            const int cols)
+        {
+            const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+            const int stride = blockDim.x * gridDim.x;
+            const int n_elements = rows * cols;
+
+            for (int i = tid; i < n_elements; i += stride) {
+                const int c = i % cols;
+                float val = x[i];
+                float s = scale[c];
+                float b = bias[c];
+
+                float res = __fadd_rn(__fmul_rn(val, s), b);
+                output[i] = res;
+            }
+        }
+
+        torch::Tensor schuraffine_cuda(
+            torch::Tensor x,
+            torch::Tensor scale,
+            torch::Tensor bias) 
+        {
+            auto x_c = x.contiguous();
+            auto s_c = scale.contiguous();
+            auto b_c = bias.contiguous();
+
+            const int rows = x_c.size(0);
+            const int cols = x_c.size(1);
+
+            auto output = torch::empty_like(x_c);
+
+            const int n_elements = rows * cols;
+            const int threads = 256;
+            const int blocks = min((n_elements + threads - 1) / threads, 65535);
+
+            schuraffine_kernel<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                s_c.data_ptr<float>(),
+                b_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                rows,
+                cols
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="schuraffine_op",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["schuraffine_cuda"],
+            extra_cuda_cflags=["-O3"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.schuraffine_cuda(
+            x, self.scale, self.bias
+        )

S1/gsd123_#98/schuraffine_torch.py

@@ -0,0 +1,21 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self, num_features=512):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(1, num_features))
+        self.bias = nn.Parameter(torch.zeros(1, num_features))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * self.scale + self.bias
+
+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 []