finish rank_normalize_scale #122

S1/gsd123_#122/prompt.txt

@@ -0,0 +1,51 @@
+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 C++ kernel for rank‑based normalization with scaling
+
+Bitonic sort in shared memory with value‑index pairs (s_val, s_idx)
+
+Stable tie‑breaking using original index when values are equal
+
+Rank calculation after sorting: rank = tid (0‑based position)
+
+Normalization: rank / (width‑1) scaled by scale parameter
+
+In‑place reordering to restore original element positions
+
+Fixed block size of 1024 threads; width must be ≤ 1024
+
+Block‑per‑sample processing (one block per batch row)
+
+PyTorch inline C++/CUDA extension via load_inline
+
+
+
+
+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):
+        super(Model, self).__init__()
+        self.scale = 10.0
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        ranks = x.argsort(dim=-1).argsort(dim=-1).float()
+        n = x.size(-1)
+        if n > 1:
+            norm = ranks / (n - 1)
+        else:
+            norm = torch.zeros_like(ranks)
+        return norm * self.scale
+
+batch_size = 128
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    return [x]
+
+def get_init_inputs():
+    return

S1/gsd123_#122/rank_normalize_scale_cuda.py

@@ -0,0 +1,95 @@
+import torch
+import torch.nn as nn
+from torch.utils.cpp_extension import load_inline
+
+cuda_source = """
+#include <torch/extension.h>
+#include <cuda_runtime.h>
+
+__global__ void rank_normalize_scale_kernel(const float* __restrict__ x, float* __restrict__ y, int batch_size, int width, float scale) {
+    int tid = threadIdx.x;
+    int row = blockIdx.x;
+
+    if (row >= batch_size) return;
+
+    __shared__ float s_val[1024];
+    __shared__ int s_idx[1024];
+
+    int idx = row * width + tid;
+
+    if (tid < width) {
+        s_val[tid] = x[idx];
+        s_idx[tid] = tid;
+    } else {
+        s_val[tid] = 3.402823466e+38F; 
+        s_idx[tid] = -1;
+    }
+    __syncthreads();
+
+    for (int k = 2; k <= 1024; k <<= 1) {
+        for (int j = k >> 1; j > 0; j >>= 1) {
+            int ixj = tid ^ j;
+            if (tid < ixj) {
+                bool up = ((tid & k) == 0);
+                float a_val = s_val[tid];
+                float b_val = s_val[ixj];
+                int a_idx = s_idx[tid];
+                int b_idx = s_idx[ixj];
+
+                bool greater = (a_val > b_val) || (a_val == b_val && a_idx > b_idx);
+
+                if (greater == up) {
+                    s_val[tid] = b_val;
+                    s_val[ixj] = a_val;
+                    s_idx[tid] = b_idx;
+                    s_idx[ixj] = a_idx;
+                }
+            }
+            __syncthreads();
+        }
+    }
+
+    if (tid < width) {
+        int original_idx = s_idx[tid];
+        float rank = (float)tid;
+        float denom = (width > 1) ? (float)(width - 1) : 1.0f;
+        float val = (rank / denom) * scale;
+
+        y[row * width + original_idx] = val;
+    }
+}
+
+torch::Tensor launch_rank_norm_scale(torch::Tensor x, float scale) {
+    auto batch_size = x.size(0);
+    auto width = x.size(1);
+    auto y = torch::empty_like(x);
+
+    const int threads = 1024;
+    const int blocks = batch_size;
+
+    rank_normalize_scale_kernel<<<blocks, threads>>>(x.data_ptr<float>(), y.data_ptr<float>(), batch_size, width, scale);
+    return y;
+}
+"""
+
+cpp_source = """
+torch::Tensor launch_rank_norm_scale(torch::Tensor x, float scale);
+"""
+
+rank_norm_scale_module = load_inline(
+    name='rank_norm_scale_op',
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=['launch_rank_norm_scale'],
+    verbose=False
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.scale = 10.0
+        self.op = rank_norm_scale_module
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.op.launch_rank_norm_scale(x.contiguous(), self.scale)

S1/gsd123_#122/rank_normalize_scale_torch.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+        self.scale = 10.0
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        ranks = x.argsort(dim=-1).argsort(dim=-1).float()
+        n = x.size(-1)
+        if n > 1:
+            norm = ranks / (n - 1)
+        else:
+            norm = torch.zeros_like(ranks)
+        return norm * self.scale
+
+batch_size = 128
+input_dim = 1024
+
+def get_inputs():
+    x = torch.randn(batch_size, input_dim)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/gsd123_#122/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 rank_normalize_scale_torch import Model, get_inputs, get_init_inputs
+from rank_normalize_scale_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()