finish gateblendnormalize #89

S1/uucoco_#89/gateblendnormalize_cuda.py

@@ -0,0 +1,119 @@
+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>
+
+__inline__ __device__ float warpReduceSum(float val) {
+    for (int offset = 16; offset > 0; offset /= 2)
+        val += __shfl_down_sync(0xffffffff, val, offset);
+    return val;
+}
+
+__inline__ __device__ float blockReduceSum(float val) {
+    static __shared__ float shared[32];
+    int lane = threadIdx.x % 32;
+    int wid = threadIdx.x / 32;
+
+    val = warpReduceSum(val);
+
+    if (lane == 0) shared[wid] = val;
+    __syncthreads();
+
+    val = (threadIdx.x < blockDim.x / 32) ? shared[lane] : 0.0f;
+
+    if (wid == 0) val = warpReduceSum(val);
+
+    return val;
+}
+
+__global__ void gate_blend_normalize_kernel(
+    const float* __restrict__ a,
+    const float* __restrict__ b,
+    const float* __restrict__ gate,
+    float* __restrict__ output,
+    int rows,
+    int cols,
+    float eps
+) {
+    int bid = blockIdx.x;
+    int tid = threadIdx.x;
+
+    if (bid >= rows) return;
+
+    const float* a_row = a + bid * cols;
+    const float* b_row = b + bid * cols;
+    const float* gate_row = gate + bid * cols;
+    float* out_row = output + bid * cols;
+
+    float sum_sq = 0.0f;
+    for (int i = tid; i < cols; i += blockDim.x) {
+        float g = gate_row[i];
+        float blended = g * a_row[i] + (1.0f - g) * b_row[i];
+        sum_sq += blended * blended;
+    }
+
+    sum_sq = blockReduceSum(sum_sq);
+
+    __shared__ float inv_norm;
+    if (tid == 0) {
+        inv_norm = rsqrtf(sum_sq + eps);
+    }
+    __syncthreads();
+
+    float norm_factor = inv_norm;
+
+    for (int i = tid; i < cols; i += blockDim.x) {
+        float g = gate_row[i];
+        float blended = g * a_row[i] + (1.0f - g) * b_row[i];
+        out_row[i] = blended * norm_factor;
+    }
+}
+
+torch::Tensor gate_blend_normalize_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor gate) {
+    auto output = torch::empty_like(a);
+
+    int cols = a.size(a.dim() - 1);
+    int rows = a.numel() / cols;
+
+    int block_size = 256;
+    while (block_size < cols && block_size < 1024) {
+        block_size *= 2;
+    }
+
+    gate_blend_normalize_kernel<<<rows, block_size>>>(
+        a.data_ptr<float>(),
+        b.data_ptr<float>(),
+        gate.data_ptr<float>(),
+        output.data_ptr<float>(),
+        rows,
+        cols,
+        1e-12f
+    );
+
+    return output;
+}
+"""
+
+cpp_source = """
+torch::Tensor gate_blend_normalize_cuda(torch::Tensor a, torch::Tensor b, torch::Tensor gate);
+"""
+
+module = load_inline(
+    name="gate_blend_normalize",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["gate_blend_normalize_cuda"],
+    verbose=True
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+        self.module = module
+
+    def forward(self, a, b, gate):
+        return self.module.gate_blend_normalize_cuda(a, b, gate)

S1/uucoco_#89/gateblendnormalize_torch.py

@@ -0,0 +1,23 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, a, b, gate):
+        blended = gate * a + (1 - gate) * b
+        return F.normalize(blended, p=2.0, dim=-1, eps=1e-12)
+
+batch_size = 1024
+dim = 1024
+
+def get_inputs():
+    a = torch.randn(batch_size, dim)
+    b = torch.randn(batch_size, dim)
+    gate = torch.rand(batch_size, dim)
+    return [a, b, gate]
+
+def get_init_inputs():
+    return []

S1/uucoco_#89/prompt.txt

@@ -0,0 +1,94 @@
+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.
+Technologies Used in This Code
+Core Libraries & Frameworks
+PyTorch: Deep learning framework
+
+CUDA: NVIDIA's parallel computing platform
+
+C++: For high-performance kernel implementation
+
+CUDA/C++ Advanced Features
+Warp Reduction: __shfl_down_sync() for warp-level operations
+
+Block Reduction: Two-level reduction (warp + shared memory)
+
+CUDA Intrinsics: rsqrtf() for reciprocal square root
+
+Dynamic Block Sizing: Adaptive thread block size based on columns
+
+Row-Level Parallelism: One CUDA block per row
+
+Mathematical Operations
+Gated Blending: g*a + (1-g)*b (element-wise gating)
+
+L2 Normalization: Compute and apply vector norms
+
+Reciprocal Square Root: Efficient 1/sqrt(x) computation
+
+Sum of Squares: Compute squared L2 norm
+
+Parallel Patterns
+Row-Based Processing: Each block processes one row
+
+Two-Pass Algorithm: First compute norm, then normalize
+
+Efficient Reduction: Warp shuffles + shared memory
+
+Grid-Stride Loops: Within each row for column processing
+
+Optimization Techniques
+Fused Operations: Blend and normalize in single kernel
+
+Numerical Stability: Epsilon (1e-12) for division safety
+
+Memory Coalescing: Row-major access patterns
+
+Adaptive Block Size: Dynamically adjusted for column count
+
+Performance Features
+Massive Parallelism: Row-level and column-level parallelism
+
+Low Synchronization: Minimal __syncthreads() usage
+
+Efficient Math: Use of rsqrtf() intrinsic
+
+Memory Efficiency: Shared memory for reduction results
+
+Unique Aspects
+Three-Input Gating: Uses gate tensor to blend two inputs
+
+Per-Row Normalization: Each output row has unit L2 norm
+
+Advanced Reduction: Custom two-level reduction functions
+
+Auto-tuning: Block size adapts to input dimensions
+
+
+
+
+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(Model, self).__init__()
+
+    def forward(self, a, b, gate):
+        blended = gate * a + (1 - gate) * b
+        return F.normalize(blended, p=2.0, dim=-1, eps=1e-12)
+
+batch_size = 1024
+dim = 1024
+
+def get_inputs():
+    a = torch.randn(batch_size, dim)
+    b = torch.randn(batch_size, dim)
+    gate = torch.rand(batch_size, dim)
+    return [a, b, gate]
+
+def get_init_inputs():
+    return []

S1/uucoco_#89/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 gateblendnormalize_torch import Model, get_inputs, get_init_inputs
+from gateblendnormalize_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()