Merge pull request 'finish Elish #35' (#286) from gsd123/GPUCodeForces:gsd35 into main

S1/gsd123_#35/Elish_cuda.py

@@ -0,0 +1,91 @@
+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 elish_cuda(torch::Tensor x);
+        """
+
+        cuda_source = """
+        #include <torch/extension.h>
+        #include <cuda_runtime.h>
+        #include <math.h>
+
+        __device__ __forceinline__ float elish_op(float x) {
+            // Sigmoid: 1 / (1 + exp(-x))
+            float sigmoid_val = 1.0f / (1.0f + expf(-x));
+
+            // ELU: x if x >= 0 else (exp(x) - 1)
+            // 使用 expm1f(x) 计算 exp(x) - 1 可以获得更高的精度
+            float elu_val = (x >= 0.0f) ? x : expm1f(x);
+
+            return elu_val * sigmoid_val;
+        }
+
+        __global__ void elish_kernel(
+            const float* __restrict__ x,
+            float* __restrict__ output,
+            const int n_elements) 
+        {
+            const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+            const int stride = blockDim.x * gridDim.x;
+
+            const int vec_loops = n_elements >> 2;
+            const float4* x_vec = reinterpret_cast<const float4*>(x);
+            float4* out_vec = reinterpret_cast<float4*>(output);
+
+            for (int i = tid; i < vec_loops; i += stride) {
+                float4 v = __ldg(&x_vec[i]);
+                float4 r;
+
+                r.x = elish_op(v.x);
+                r.y = elish_op(v.y);
+                r.z = elish_op(v.z);
+                r.w = elish_op(v.w);
+
+                out_vec[i] = r;
+            }
+
+            const int tail_start = vec_loops << 2;
+            for (int i = tail_start + tid; i < n_elements; i += stride) {
+                output[i] = elish_op(x[i]);
+            }
+        }
+
+        torch::Tensor elish_cuda(torch::Tensor x) {
+            auto x_c = x.contiguous();
+            const int n_elements = x_c.numel();
+            auto output = torch::empty_like(x_c);
+
+            const int threads = 256;
+            const int max_blocks = 65535;
+            const int blocks = std::min((n_elements + threads * 4 - 1) / (threads * 4), max_blocks);
+
+            elish_kernel<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                n_elements
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="elish_v1",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["elish_cuda"],
+            extra_cuda_cflags=["-O3", "--use_fast_math"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.elish_cuda(x)

S1/gsd123_#35/Elish_torch.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # Elish = ELU(x) * Sigmoid(x)
+        return F.elu(x) * torch.sigmoid(x)
+
+
+batch_size = 1024
+feature_dim = 1024
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return []

S1/gsd123_#35/prompt.txt

@@ -0,0 +1,84 @@
+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.
+
+CUDA Optimization Strategies:
+
+Vectorized Memory Access
+
+Uses float4 for 4-element vector loads/stores
+
+__ldg() for read-only caching through texture memory
+
+Bit shifts for division (>> 2, << 2) for efficiency
+
+Elish Activation Function
+
+Computes Elish(x) = ELU(x) * Sigmoid(x)
+
+Combination of ELU and Sigmoid activations
+
+Requires careful numerical handling
+
+Numerical Precision
+
+Uses expm1f(x) for exp(x) - 1 in negative region
+
+Higher accuracy for small x values
+
+Standard expf(-x) for sigmoid
+
+Memory Access
+
+contiguous() tensors for coalescing
+
+__restrict__ pointers
+
+Grid-stride loop for arbitrary sizes
+
+Performance Optimization
+
+Compiler flags: -O3, --use_fast_math
+
+Efficient kernel launch configuration
+
+Block count limited to 65535
+
+Branch for ELU (x >= 0) condition
+
+Mathematical Efficiency
+
+Inline ELU and Sigmoid computations
+
+Vectorized operations for 4 elements simultaneously
+
+Minimal conditional branching
+
+Key Innovation: Vectorized Elish activation function combining ELU and Sigmoid, optimized with high-precision expm1f for numerical stability in the negative region.
+
+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
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # Elish = ELU(x) * Sigmoid(x)
+        return F.elu(x) * torch.sigmoid(x)
+
+
+batch_size = 1024
+feature_dim = 1024
+
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+
+def get_init_inputs():
+    return [

S1/gsd123_#35/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 Elish_torch import Model, get_inputs, get_init_inputs
+from Elish_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()