Merge pull request 'finish P-GELU #99' (#718) from hli28146/GPUCodeForces:h99 into main

S1/hli28146_#99/PGELU_cuda.py

@@ -0,0 +1,117 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+import math
+
+cpp_source = """
+#include <torch/extension.h>
+
+torch::Tensor pgelu_cuda_forward(
+    const torch::Tensor& input, 
+    const torch::Tensor& alpha, 
+    const torch::Tensor& beta);
+"""
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+#include <math.h>
+
+#define BLOCK_SIZE 256
+#define PI_HALF 1.5707963267948966
+#define EPS 1e-7
+
+// double2 for 128-bit vectorization
+struct __align__(16) Double2 {
+    double x, y;
+};
+
+// P-GELU for double
+__device__ __forceinline__ double compute_pgelu_double(double x, double alpha, double beta) {
+    double inner = alpha * x + beta * x * x * x;
+    double clamped_inner = fmin(fmax(inner, -PI_HALF + EPS), PI_HALF - EPS);
+    return x * (1.0 + tan(clamped_inner));
+}
+
+template <typename T>
+__global__ void pgelu_kernel(
+    T* __restrict__ output,
+    const T* __restrict__ input,
+    const int n,
+    T alpha,
+    T beta) 
+{
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int vec_n = n / 2; // double2
+    
+    int i = idx;
+    const int stride = blockDim.x * gridDim.x;
+
+    for (; i < vec_n; i += stride) {
+        Double2 in_vec = reinterpret_cast<const Double2*>(input)[i];
+        Double2 out_vec;
+
+        out_vec.x = compute_pgelu_double(in_vec.x, alpha, beta);
+        out_vec.y = compute_pgelu_double(in_vec.y, alpha, beta);
+
+        reinterpret_cast<Double2*>(output)[i] = out_vec;
+    }
+
+    int tail_idx = vec_n * 2 + idx;
+    if (idx == 0 && tail_idx < n) {
+        output[tail_idx] = compute_pgelu_double(input[tail_idx], alpha, beta);
+    }
+}
+
+torch::Tensor pgelu_cuda_forward(
+    const torch::Tensor& input, 
+    const torch::Tensor& alpha_t, 
+    const torch::Tensor& beta_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 double alpha = alpha_t.item<double>();
+    const double beta = beta_t.item<double>();
+
+    const int vec_n = n / 2;
+    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;
+    
+    AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "pgelu_kernel", ([&]{
+        pgelu_kernel<scalar_t><<<final_grid, BLOCK_SIZE>>>(
+            output.data_ptr<scalar_t>(),
+            input.data_ptr<scalar_t>(),
+            n,
+            static_cast<scalar_t>(alpha),
+            static_cast<scalar_t>(beta)
+        );
+    }));
+
+    return output;
+}
+"""
+
+pgelu_op_module = load_inline(
+    name='pgelu_tan_op_double',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['pgelu_cuda_forward'],
+    verbose=False,
+    extra_cuda_cflags=['-O3']
+)
+
+class ModelNew(nn.Module):
+    def __init__(self, alpha_init=1.0, beta_init=1.0):
+        super(ModelNew, self).__init__()
+        self.alpha = nn.Parameter(torch.tensor(alpha_init, dtype=torch.float64))
+        self.beta = nn.Parameter(torch.tensor(beta_init, dtype=torch.float64))
+        self.op = pgelu_op_module
+
+    def forward(self, input_tensor: torch.Tensor) -> torch.Tensor:
+        return self.op.pgelu_cuda_forward(input_tensor.contiguous(), self.alpha, self.beta)

S1/hli28146_#99/PGELU_torch.py

@@ -0,0 +1,45 @@
+import torch
+import torch.nn as nn
+import math
+
+BATCH_SIZE = 4096
+HIDDEN_DIM = 4096
+SHAPE = (BATCH_SIZE, HIDDEN_DIM)
+
+ALPHA_INIT = 1.0
+BETA_INIT = 1.0
+
+DTYPE = torch.float64
+
+class PGELU(nn.Module):
+    '''
+    P-GELU: A Novel Activation Function to Optimize Whisper for Darija Speech Translation
+    https://ieeexplore.ieee.org/document/11016691
+    Formula: f(x) = x * (1 + tan(alpha * x + beta * x^3))
+    '''
+    def __init__(self, alpha_init=1.0, beta_init=1.0):
+        super(PGELU, self).__init__()
+        self.alpha = nn.Parameter(torch.tensor(alpha_init, dtype=DTYPE))
+        self.beta = nn.Parameter(torch.tensor(beta_init, dtype=DTYPE))
+        self.pi_half = math.pi / 2.0
+        self.eps = 1e-7 # Epsilon for double
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        inner = self.alpha * x + self.beta * x.pow(3)
+        inner_clamped = torch.clamp(inner, -self.pi_half + self.eps, self.pi_half - self.eps)
+        return x * (1.0 + torch.tan(inner_clamped))
+
+class Model(nn.Module):
+    def __init__(self, alpha_init=1.0, beta_init=1.0):
+        super(Model, self).__init__()
+        self.act = PGELU(alpha_init, beta_init)
+    
+    def forward(self, x):
+        return self.act(x)
+
+def get_inputs():
+    input_tensor = torch.randn(SHAPE, dtype=DTYPE)
+    return [input_tensor.contiguous()]
+
+def get_init_inputs():
+    return [ALPHA_INIT, BETA_INIT]

S1/hli28146_#99/prompt.txt

@@ -0,0 +1,71 @@
+Write a custom CUDA kernel to optimize `P-GELU` using `float64` (double) precision.
+
+Formula: f(x) = x * (1 + tan(alpha * x + beta * x^3))
+
+Problem Analysis:
+1. Precision Issues with float32: The combination of a cubic polynomial and the `tan` function amplifies floating-point rounding errors.
+2. Memory Bottleneck: The operation is memory-bound, now with 8 bytes per element.
+3. Numerical Stability: The input to `tan` must be clamped to avoid asymptotes.
+
+Optimization Strategy: Fused Element-wise Kernel with Double Precision
+
+1. Data Type: All computations are performed in `double`.
+
+2. Vectorized Loads (double2): Use `double2` to load 128 bits (2 double elements) per memory transaction.
+
+3. Fused Stable Math (in double):
+   - For `x`, compute `inner = alpha * x + beta * x*x*x`.
+   - Clamp `inner` to stay away from `pi/2`.
+   - Compute `result = x * (1.0 + tan(inner))`.
+   - Use standard `double` precision math functions (`tan`).
+
+4. One-Pass: Fuse all logic 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 math
+
+BATCH_SIZE = 4096
+HIDDEN_DIM = 4096
+SHAPE = (BATCH_SIZE, HIDDEN_DIM)
+
+ALPHA_INIT = 1.0
+BETA_INIT = 1.0
+
+DTYPE = torch.float64
+
+class PGELU(nn.Module):
+    '''
+    P-GELU: A Novel Activation Function to Optimize Whisper for Darija Speech Translation
+    https://ieeexplore.ieee.org/document/11016691
+    Formula: f(x) = x * (1 + tan(alpha * x + beta * x^3))
+    '''
+    def __init__(self, alpha_init=1.0, beta_init=1.0):
+        super(PGELU, self).__init__()
+        self.alpha = nn.Parameter(torch.tensor(alpha_init, dtype=DTYPE))
+        self.beta = nn.Parameter(torch.tensor(beta_init, dtype=DTYPE))
+        self.pi_half = math.pi / 2.0
+        self.eps = 1e-7 # Epsilon for double
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        inner = self.alpha * x + self.beta * x.pow(3)
+        inner_clamped = torch.clamp(inner, -self.pi_half + self.eps, self.pi_half - self.eps)
+        return x * (1.0 + torch.tan(inner_clamped))
+
+class Model(nn.Module):
+    def __init__(self, alpha_init=1.0, beta_init=1.0):
+        super(Model, self).__init__()
+        self.act = PGELU(alpha_init, beta_init)
+    
+    def forward(self, x):
+        return self.act(x)
+
+def get_inputs():
+    input_tensor = torch.randn(SHAPE, dtype=DTYPE)
+    return [input_tensor.contiguous()]
+
+def get_init_inputs():
+    return [ALPHA_INIT, BETA_INIT]

S1/hli28146_#99/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 PGELU_torch import Model,get_inputs,get_init_inputs
+from PGELU_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()