finish segment_logsumexp #129

S1/gsd123_#129/prompt.txt

@@ -0,0 +1,137 @@
+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.
+Core Mathematical Algorithm
+Log-Sum-Exp (LSE) Numerical Stabilization
+
+Formula: log(∑ exp(x_i)) = m + log(∑ exp(x_i - m))
+
+Uses segment-wise maximum m for numerical stability
+
+Prevents overflow in exponential computations
+
+CUDA Implementation Details
+Four-Kernel Pipeline Design
+
+init_max_kernel: Initializes maximum values to -1e38
+
+segment_max_kernel: Finds per-segment maximum values
+
+segment_sum_exp_kernel: Computes ∑exp(x_i - m)
+
+finalize_lse_kernel: Finalizes LSE: m + log(sum_exp)
+
+Custom Atomic Float Operations
+
+Reused atomicMaxFloat from previous implementation
+
+Uses atomicCAS for thread-safe float maximum
+
+atomicAdd for accumulating exponential sums
+
+Memory Management Strategy
+
+Three intermediate tensors:
+
+max_val: Per-segment maximum values
+
+sum_exp: Sum of exponentials (offset by max)
+
+out: Final logsumexp results
+
+torch::empty for uninitialized tensors (performance)
+
+torch::zeros for accumulation buffer (safety)
+
+Numerical Computing Techniques
+Stable Exponential Computation
+
+Offset by maximum: expf(val - m)
+
+Uses expf and logf CUDA math intrinsics
+
+Initialization to -1e38 instead of -FLT_MAX
+
+Parallel Reduction Pattern
+
+Segment-wise reduction with atomic operations
+
+Two-pass approach: max reduction → sum reduction
+
+Channel-wise independent computation
+
+Performance Optimization
+Kernel Fusion/Optimization
+
+Shared index computation between kernels
+
+Reused grid/block calculations
+
+Efficient 2D→1D indexing: segment_id * channels + col
+
+CUDA Best Practices
+
+__restrict__ keyword for compiler optimization
+
+Grid-stride loops with 256 threads per block
+
+Coalesced memory access patterns
+
+PyTorch Integration
+Extension Framework
+
+Runtime compilation via load_inline
+
+Automatic device/dtype propagation
+
+Seamless autograd integration
+
+Module Design
+
+nn.Module wrapper for reusability
+
+Configurable dim_size parameter
+
+Clean Python-CUDA interface
+
+
+
+
+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, dim_size):
+        super(Model, self).__init__()
+        self.dim_size = dim_size
+
+    def forward(self, src, index):
+        max_val = torch.full((self.dim_size, src.size(1)), -float('inf'), device=src.device, dtype=src.dtype)
+        index_expanded = index.unsqueeze(1).expand_as(src)
+
+        max_val.scatter_reduce_(0, index_expanded, src, reduce='amax', include_self=False)
+
+        gathered_max = max_val.gather(0, index_expanded)
+        exp_src = torch.exp(src - gathered_max)
+
+        sum_exp = torch.zeros_like(max_val)
+        sum_exp.scatter_add_(0, index_expanded, exp_src)
+
+        return max_val + torch.log(sum_exp)
+
+
+batch_size = 1024
+features = 64
+dim_size = 128
+
+
+def get_inputs():
+    src = torch.randn(batch_size, features)
+    index = torch.randint(0, dim_size, (batch_size,))
+    return [src, index]
+
+
+def get_init_inputs():
+    return [dim_size]

S1/gsd123_#129/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 segmentlogsumexp_torch import Model, get_inputs, get_init_inputs
+from segmentlogsumexp_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()

S1/gsd123_#129/segmentlogsumexp_cuda.py

@@ -0,0 +1,153 @@
+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>
+#include <cfloat>
+
+__device__ __forceinline__ void atomicMaxFloat(float* address, float val) {
+    int* address_as_i = (int*)address;
+    int old = *address_as_i, assumed;
+    do {
+        assumed = old;
+        float old_val = __int_as_float(assumed);
+        float new_val = fmaxf(val, old_val);
+        if (new_val == old_val) break;
+        old = atomicCAS(address_as_i, assumed, __float_as_int(new_val));
+    } while (assumed != old);
+}
+
+__global__ void init_max_kernel(float* out, int size, float val) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < size) {
+        out[idx] = val;
+    }
+}
+
+__global__ void segment_max_kernel(
+    const float* __restrict__ src,
+    const long* __restrict__ index,
+    float* __restrict__ max_val,
+    int num_elements,
+    int channels,
+    int dim_size
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < num_elements) {
+        int row = idx / channels;
+        int col = idx % channels;
+        long segment_id = index[row];
+
+        if (segment_id >= 0 && segment_id < dim_size) {
+            int out_idx = segment_id * channels + col;
+            atomicMaxFloat(&max_val[out_idx], src[idx]);
+        }
+    }
+}
+
+__global__ void segment_sum_exp_kernel(
+    const float* __restrict__ src,
+    const long* __restrict__ index,
+    const float* __restrict__ max_val,
+    float* __restrict__ sum_exp,
+    int num_elements,
+    int channels,
+    int dim_size
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < num_elements) {
+        int row = idx / channels;
+        int col = idx % channels;
+        long segment_id = index[row];
+
+        if (segment_id >= 0 && segment_id < dim_size) {
+            int out_idx = segment_id * channels + col;
+            float m = max_val[out_idx];
+            float val = src[idx];
+            atomicAdd(&sum_exp[out_idx], expf(val - m));
+        }
+    }
+}
+
+__global__ void finalize_lse_kernel(
+    float* __restrict__ out,
+    const float* __restrict__ max_val,
+    const float* __restrict__ sum_exp,
+    int size
+) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < size) {
+        float m = max_val[idx];
+        float s = sum_exp[idx];
+        out[idx] = m + logf(s);
+    }
+}
+
+torch::Tensor segment_logsumexp_cuda(torch::Tensor src, torch::Tensor index, int dim_size) {
+    int N = src.size(0);
+    int C = src.size(1);
+    int num_elements = N * C;
+    int out_elements = dim_size * C;
+
+    auto max_val = torch::empty({dim_size, C}, src.options());
+    auto sum_exp = torch::zeros({dim_size, C}, src.options());
+    auto out = torch::empty({dim_size, C}, src.options());
+
+    const int block_size = 256;
+    int grid_init = (out_elements + block_size - 1) / block_size;
+    init_max_kernel<<<grid_init, block_size>>>(max_val.data_ptr<float>(), out_elements, -1e38f);
+
+    int grid_scatter = (num_elements + block_size - 1) / block_size;
+    segment_max_kernel<<<grid_scatter, block_size>>>(
+        src.data_ptr<float>(),
+        index.data_ptr<long>(),
+        max_val.data_ptr<float>(),
+        num_elements,
+        C,
+        dim_size
+    );
+
+    segment_sum_exp_kernel<<<grid_scatter, block_size>>>(
+        src.data_ptr<float>(),
+        index.data_ptr<long>(),
+        max_val.data_ptr<float>(),
+        sum_exp.data_ptr<float>(),
+        num_elements,
+        C,
+        dim_size
+    );
+
+    finalize_lse_kernel<<<grid_init, block_size>>>(
+        out.data_ptr<float>(),
+        max_val.data_ptr<float>(),
+        sum_exp.data_ptr<float>(),
+        out_elements
+    );
+
+    return out;
+}
+"""
+
+cpp_source = """
+torch::Tensor segment_logsumexp_cuda(torch::Tensor src, torch::Tensor index, int dim_size);
+"""
+
+segment_lse_lib = load_inline(
+    name="segment_logsumexp",
+    cpp_sources=cpp_source,
+    cuda_sources=cuda_source,
+    functions=["segment_logsumexp_cuda"],
+    verbose=True
+)
+
+
+class ModelNew(nn.Module):
+    def __init__(self, dim_size):
+        super(ModelNew, self).__init__()
+        self.dim_size = dim_size
+        self.lib = segment_lse_lib
+
+    def forward(self, src, index):
+        return self.lib.segment_logsumexp_cuda(src, index, self.dim_size)

S1/gsd123_#129/segmentlogsumexp_torch.py

@@ -0,0 +1,37 @@
+import torch
+import torch.nn as nn
+
+
+class Model(nn.Module):
+    def __init__(self, dim_size):
+        super(Model, self).__init__()
+        self.dim_size = dim_size
+
+    def forward(self, src, index):
+        max_val = torch.full((self.dim_size, src.size(1)), -float('inf'), device=src.device, dtype=src.dtype)
+        index_expanded = index.unsqueeze(1).expand_as(src)
+
+        max_val.scatter_reduce_(0, index_expanded, src, reduce='amax', include_self=False)
+
+        gathered_max = max_val.gather(0, index_expanded)
+        exp_src = torch.exp(src - gathered_max)
+
+        sum_exp = torch.zeros_like(max_val)
+        sum_exp.scatter_add_(0, index_expanded, exp_src)
+
+        return max_val + torch.log(sum_exp)
+
+
+batch_size = 1024
+features = 64
+dim_size = 128
+
+
+def get_inputs():
+    src = torch.randn(batch_size, features)
+    index = torch.randint(0, dim_size, (batch_size,))
+    return [src, index]
+
+
+def get_init_inputs():
+    return [dim_size]