Merge pull request 'finish SCL Mish #119' (#850) from hli28146/GPUCodeForces:h119 into main

S1/hli28146_#119/SCLMish_cuda.py

@@ -0,0 +1,113 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cpp_source = """
+#include <torch/extension.h>
+
+torch::Tensor scl_mish_cuda_forward(const torch::Tensor& input, const torch::Tensor& alpha);
+"""
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+#include <math.h>
+
+#define BLOCK_SIZE 256
+
+struct __align__(16) Float4 {
+    float x, y, z, w;
+};
+
+// SCL Mish Logic
+__device__ __forceinline__ float compute_scl_mish(float x, float alpha) {
+    float ax = alpha * x;
+    float sp;
+    // Stable Softplus
+    if (ax > 20.0f) {
+        sp = ax;
+    } else {
+        sp = log1pf(expf(ax));
+    }
+    float mish_part = x * tanhf(sp);
+    return fmaxf(mish_part, 0.0f);
+}
+
+__global__ void scl_mish_kernel(
+    float* __restrict__ output,
+    const float* __restrict__ input,
+    const int n,
+    const float alpha) 
+{
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int vec_n = n / 4;
+    
+    int i = idx;
+    const int stride = blockDim.x * gridDim.x;
+
+    for (; i < vec_n; i += stride) {
+        Float4 in_vec = reinterpret_cast<const Float4*>(input)[i];
+        Float4 out_vec;
+
+        out_vec.x = compute_scl_mish(in_vec.x, alpha);
+        out_vec.y = compute_scl_mish(in_vec.y, alpha);
+        out_vec.z = compute_scl_mish(in_vec.z, alpha);
+        out_vec.w = compute_scl_mish(in_vec.w, alpha);
+
+        reinterpret_cast<Float4*>(output)[i] = out_vec;
+    }
+
+    int start_scalar = vec_n * 4;
+    int global_tid = blockIdx.x * blockDim.x + threadIdx.x;
+    int total_threads = gridDim.x * gridDim.x;
+    
+    int current_idx = start_scalar + global_tid;
+    while (current_idx < n) {
+        output[current_idx] = compute_scl_mish(input[current_idx], alpha);
+        current_idx += total_threads;
+    }
+}
+
+torch::Tensor scl_mish_cuda_forward(const torch::Tensor& input, const torch::Tensor& alpha_t) {
+    TORCH_CHECK(input.is_cuda(), "Input must be a CUDA tensor");
+    TORCH_CHECK(input.is_contiguous(), "Input must be contiguous");
+    
+    const int n = input.numel();
+    auto output = torch::empty_like(input);
+
+    const float alpha = alpha_t.item<float>();
+
+    const int vec_n = n / 4;
+    const int grid_size = (vec_n + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    
+    int final_grid = (grid_size < 1) ? 1 : grid_size;
+    if (final_grid > 65535) final_grid = 65535;
+
+    scl_mish_kernel<<<final_grid, BLOCK_SIZE>>>(
+        output.data_ptr<float>(),
+        input.data_ptr<float>(),
+        n,
+        alpha
+    );
+
+    return output;
+}
+"""
+
+scl_mish_op_module = load_inline(
+    name='scl_mish_op',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['scl_mish_cuda_forward'],
+    verbose=False,
+    extra_cuda_cflags=['-O3']
+)
+
+class ModelNew(nn.Module):
+    def __init__(self, alpha_init=0.25):
+        super(ModelNew, self).__init__()
+        self.alpha = nn.Parameter(torch.tensor(alpha_init))
+        self.op = scl_mish_op_module
+
+    def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
+        return self.op.scl_mish_cuda_forward(input_tensor.contiguous(), self.alpha)

S1/hli28146_#119/SCLMish_torch.py

@@ -0,0 +1,42 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+BATCH_SIZE = 4096
+HIDDEN_DIM = 4096
+SHAPE = (BATCH_SIZE, HIDDEN_DIM)
+
+ALPHA_INIT = 0.25
+
+class SCLMish(nn.Module):
+    """
+    Soft Clipping Mish (learnable).
+    Soft Clipping Mish - A Novel Activation Function for Deep Learning
+    DOI:10.1109/ICICT52872.2021.00010
+
+    Formula: f(x) = max(0, x * tanh(softplus(alpha * x)))
+    """
+    def __init__(self, alpha_init=0.25):
+        super(SCLMish, self).__init__()
+        self.alpha = nn.Parameter(torch.tensor(alpha_init))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        mish_part = x * F.mish(self.alpha * x) / (self.alpha * x + 1e-8) # Re-normalize
+        mish_part_correct = x * torch.tanh(F.softplus(self.alpha * x))
+        
+        return F.relu(mish_part_correct)
+
+class Model(nn.Module):
+    def __init__(self, alpha_init=0.25):
+        super(Model, self).__init__()
+        self.act = SCLMish(alpha_init)
+    
+    def forward(self, x):
+        return self.act(x)
+
+def get_inputs():
+    input_tensor = torch.randn(SHAPE, dtype=torch.float32) * 5.0
+    return [input_tensor.contiguous()]
+
+def get_init_inputs():
+    return [ALPHA_INIT]

S1/hli28146_#119/prompt.txt

@@ -0,0 +1,50 @@
+Write a custom CUDA kernel to optimize `SCL Mish` (Soft Clipping Mish learnable).
+
+Formula: f(x) = max(0, x * tanh(softplus(alpha * x)))
+where softplus(z) = log(1 + exp(z)).
+
+Problem Analysis:
+1. Computationally Intensive & Memory Bound: The operation is element-wise but involves a long chain of transcendental functions (exp, log, tanh).
+2. Operator Chaining: A standard PyTorch implementation creates multiple intermediate tensors and kernel launches.
+
+Optimization Strategy: Fused Element-wise Kernel with Vectorization
+
+1. One-Thread-per-Element: Map each element to a CUDA thread.
+
+2. Vectorized Loads (float4): Use `float4` to process 128 bits per memory transaction.
+
+3. Fused Stable Math:
+   - For each element `x`, compute `ax = alpha * x`.
+   - Compute stable softplus: `sp = (ax > 20) ? ax : log1pf(__expf(ax))`.
+   - Compute `mish_part = x * tanhf(sp)`.
+   - Result `fmaxf(mish_part, 0.0f)`.
+   - All steps are fused in registers.
+
+4. One-Pass: Fuse all steps into a single read-compute-write kernel.
+  
+Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:   
+  
+```python
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Model(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+
+    def forward(self, a, b):
+        return a + b
+
+
+def get_inputs():
+    # randomly generate input tensors based on the model architecture
+    a = torch.randn(1, 128).cuda()
+    b = torch.randn(1, 128).cuda()
+    return [a, b]
+
+
+def get_init_inputs():
+    # randomly generate tensors required for initialization based on the model architecture
+    return []

S1/hli28146_#119/run_code.py

@@ -0,0 +1,74 @@
+###########################################################
+# 性能和精度验证程序
+###########################################################
+import torch
+import torch.nn as nn
+import time
+from SCLMish_torch import Model,get_inputs,get_init_inputs
+from SCLMish_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()