Merge pull request 'finish Mish #31' (#282) from gsd123/GPUCodeForces:gsd31 into main

S1/gsd123_#31/Mish_cuda.py

@@ -0,0 +1,116 @@
+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 = """
+        #include <torch/extension.h>
+        torch::Tensor mish_cuda(torch::Tensor x);
+        """
+
+        cuda_source = """
+        #include <cuda_runtime.h>
+        __device__ __forceinline__ float mish_op_fast(float x) {
+            // 阈值保护：x > 20 时，Softplus(x) ≈ x, Tanh(x) ≈ 1, Mish ≈ x
+            // 同时也防止 e^(2x) 在 float32 下溢出 (e^88 溢出)
+            if (x > 20.0f) return x;
+
+            float e = expf(x);
+            float n = e * (2.0f + e);
+            float d = 2.0f + 2.0f * e + e * e;
+
+            return x * (n / d);
+        }
+
+        __global__ void mish_kernel_alg(
+            const float* __restrict__ x,
+            float* __restrict__ y,
+            int n) 
+        {
+            int tid = blockIdx.x * blockDim.x + threadIdx.x;
+            int stride = blockDim.x * gridDim.x;
+
+            int vec_n = n / 4;
+            const float4* x_vec = reinterpret_cast<const float4*>(x);
+            float4* y_vec = reinterpret_cast<float4*>(y);
+
+            int i = tid;
+
+            for (; i < vec_n - 1; i += stride) {
+                float4 v1 = x_vec[i];
+                float4 v2 = x_vec[i + 1]; 
+
+                float4 o1, o2;
+
+                o1.x = mish_op_fast(v1.x);
+                o1.y = mish_op_fast(v1.y);
+                o1.z = mish_op_fast(v1.z);
+                o1.w = mish_op_fast(v1.w);
+
+                o2.x = mish_op_fast(v2.x);
+                o2.y = mish_op_fast(v2.y);
+                o2.z = mish_op_fast(v2.z);
+                o2.w = mish_op_fast(v2.w);
+
+                y_vec[i] = o1;
+                y_vec[i + 1] = o2;
+
+                i++; // Skip next
+            }
+
+            for (; i < vec_n; i += stride) {
+                float4 v = x_vec[i];
+                float4 o;
+                o.x = mish_op_fast(v.x);
+                o.y = mish_op_fast(v.y);
+                o.z = mish_op_fast(v.z);
+                o.w = mish_op_fast(v.w);
+                y_vec[i] = o;
+            }
+
+            int tail_start = vec_n * 4;
+            for (int j = tail_start + tid; j < n; j += stride) {
+                y[j] = mish_op_fast(x[j]);
+            }
+        }
+
+        torch::Tensor mish_cuda(torch::Tensor x) {
+            auto x_c = x.contiguous();
+            auto output = torch::empty_like(x_c);
+
+            int total_elements = x_c.numel();
+            int threads = 256;
+
+            int vec_elements = total_elements / 4;
+            int blocks = (vec_elements + threads - 1) / threads;
+
+            if (blocks > 65535) blocks = 65535;
+            if (blocks == 0) blocks = 1;
+
+            mish_kernel_alg<<<blocks, threads>>>(
+                x_c.data_ptr<float>(),
+                output.data_ptr<float>(),
+                total_elements
+            );
+
+            return output;
+        }
+        """
+
+        self.op = load_inline(
+            name="mish_opt_alg",
+            cpp_sources=cpp_source,
+            cuda_sources=cuda_source,
+            functions=["mish_cuda"],
+            extra_cuda_cflags=["-O3", "--use_fast_math"],
+            verbose=False
+        )
+
+    def forward(self, x):
+        return self.op.mish_cuda(x)

S1/gsd123_#31/Mish_torch.py

@@ -0,0 +1,20 @@
+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:
+        return x * torch.tanh(F.softplus(x))
+
+batch_size = 1024
+feature_dim = 4096
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/gsd123_#31/prompt.txt

@@ -0,0 +1,81 @@
+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:
+
+Numerical Stability Optimization
+
+Threshold protection: Returns x directly when x > 20.0f
+
+Prevents expf(2*x) overflow in float32 (avoids e^88 overflow)
+
+Algebraic reformulation for better numerical behavior
+
+Vectorized Memory Access + ILP
+
+Uses float4 for 4-element vector loads/stores
+
+Instruction-Level Parallelism (ILP): Processes 2 vectors (8 elements) per loop iteration
+
+Increases computational density and hides memory latency
+
+Algebraic Reformulation
+
+Optimized Mish computation: x * (e*(2+e)) / (2 + 2*e + e*e)
+
+Avoids separate tanh and softplus computations
+
+Reduces mathematical operations
+
+Grid-Stride Loop
+
+Processes elements with grid-stride pattern
+
+Handles arbitrary tensor sizes efficiently
+
+Better GPU utilization
+
+Memory Access
+
+contiguous() tensors for coalescing
+
+__restrict__ pointers
+
+Coalesced memory access patterns
+
+Performance Tuning
+
+Fixed 256 threads per block
+
+Block count capped at 65535
+
+Compiler flags: -O3, --use_fast_math
+
+Key Innovation: Algebraic reformulation with numerical stability protection prevents overflow while maintaining mathematical equivalence, combined with ILP for performance.
+
+
+
+
+
+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:
+        return x * torch.tanh(F.softplus(x))
+
+batch_size = 1024
+feature_dim = 4096
+
+def get_inputs():
+    x = torch.randn(batch_size, feature_dim, dtype=torch.float32)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/gsd123_#31/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 Mish_torch import Model, get_inputs, get_init_inputs
+from Mish_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()