finish PointNetLoss #120

S1/gsd123_#120/PointNetLoss_cuda.py

@@ -0,0 +1,93 @@
+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 pointnet_loss_kernel(const float* __restrict__ mat, float* __restrict__ out, int batch_size, int n) {
+    extern __shared__ float sdata[];
+
+    int bid = blockIdx.x;
+    if (bid >= batch_size) return;
+
+    const float* A = mat + bid * n * n;
+    int tid = threadIdx.x;
+    int num_elements = n * n;
+
+    for (int i = tid; i < num_elements; i += blockDim.x) {
+        sdata[i] = A[i];
+    }
+    __syncthreads();
+
+    float local_sum_sq = 0.0f;
+    for (int i = tid; i < num_elements; i += blockDim.x) {
+        int r = i / n;
+        int c = i % n;
+        float dot = 0.0f;
+        for (int k = 0; k < n; ++k) {
+            dot += sdata[r * n + k] * sdata[c * n + k];
+        }
+        float target = (r == c) ? 1.0f : 0.0f;
+        float diff = dot - target;
+        local_sum_sq += diff * diff;
+    }
+
+    __syncthreads();
+
+    sdata[tid] = local_sum_sq;
+    __syncthreads();
+
+    for (int s = blockDim.x / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sdata[tid] += sdata[tid + s];
+        }
+        __syncthreads();
+    }
+
+    if (tid == 0) {
+        out[bid] = sdata[0];
+    }
+}
+
+torch::Tensor pointnet_loss_cuda(torch::Tensor x) {
+    int batch_size = x.size(0);
+    int n = x.size(1);
+
+    auto out = at::empty({(long)batch_size}, x.options());
+
+    int threads = 256;
+    int blocks = batch_size;
+    int shared_mem_size = max(n * n, threads) * 4;
+
+    pointnet_loss_kernel<<<blocks, threads, shared_mem_size>>>(
+        x.data_ptr<float>(),
+        out.data_ptr<float>(),
+        batch_size,
+        n
+    );
+
+    return out.mean();
+}
+"""
+
+cpp_source = """
+torch::Tensor pointnet_loss_cuda(torch::Tensor x);
+"""
+
+pointnet_loss = load_inline(
+    name="pointnet_loss",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["pointnet_loss_cuda"],
+    verbose=False
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+
+    def forward(self, x):
+        return pointnet_loss.pointnet_loss_cuda(x)

S1/gsd123_#120/PointNetLoss_torch.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x):
+        batch_size = x.size(0)
+        num_channels = x.size(1)
+        identity = torch.eye(num_channels, device=x.device).unsqueeze(0).repeat(batch_size, 1, 1)
+        x_t = x.transpose(1, 2)
+        product = torch.bmm(x, x_t)
+        diff = identity - product
+        loss = torch.sum(diff ** 2, dim=(1, 2))
+        return torch.mean(loss)
+
+batch_size = 32
+num_channels = 64
+
+def get_inputs():
+    x = torch.randn(batch_size, num_channels, num_channels, requires_grad=True)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/gsd123_#120/prompt.txt

@@ -0,0 +1,53 @@
+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.
+Custom CUDA kernel extension via torch.utils.cpp_extension.load_inline
+
+PointNet-style transformation regularization (orthogonality loss)
+
+Matrix multiplication within kernel for AᵀA computation
+
+Identity target matrix with 1s on diagonal, 0s elsewhere
+
+Frobenius norm squared error between AᵀA and identity matrix
+
+Shared memory caching for input matrix tiles
+
+Tree-based parallel reduction for loss accumulation
+
+Per-batch parallel processing (one CUDA block per sample)
+
+Dynamic shared memory allocation based on matrix size
+
+Contiguous tensor handling for memory coalescing
+
+
+
+
+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__()
+
+    def forward(self, x):
+        batch_size = x.size(0)
+        num_channels = x.size(1)
+        identity = torch.eye(num_channels, device=x.device).unsqueeze(0).repeat(batch_size, 1, 1)
+        x_t = x.transpose(1, 2)
+        product = torch.bmm(x, x_t)
+        diff = identity - product
+        loss = torch.sum(diff ** 2, dim=(1, 2))
+        return torch.mean(loss)
+
+batch_size = 32
+num_channels = 64
+
+def get_inputs():
+    x = torch.randn(batch_size, num_channels, num_channels, requires_grad=True)
+    return [x]
+
+def get_init_inputs():
+    return []

S1/gsd123_#120/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 PointNetLoss_torch import Model, get_inputs, get_init_inputs
+from PointNetLoss_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()