nebula vsearch note 1

Signed-off-by: Zhi Yiliu <2584074296@qq.com>

Author: tom-jerr

docs/.obsidian/workspace.json

@@ -7,20 +7,6 @@
         "id": "07150ca20b5d2386",
         "type": "tabs",
         "children": [
-          {
-            "id": "87be62f38607b0ae",
-            "type": "leaf",
-            "state": {
-              "type": "markdown",
-              "state": {
-                "file": "projects/nebula-vsearch/上篇：初识 Nebula Graph —— 向量类型支持.md",
-                "mode": "source",
-                "source": false
-              },
-              "icon": "lucide-file",
-              "title": "上篇：初识 Nebula Graph —— 向量类型支持"
-            }
-          },
           {
             "id": "cfef105082839db4",
             "type": "leaf",
@@ -185,53 +171,54 @@
       "bases:Create new base": false
     }
   },
-  "active": "87be62f38607b0ae",
+  "active": "cfef105082839db4",
   "lastOpenFiles": [
-    "projects/active slam/kr 3d active slam.md",
-    "paperreadings/activeslam/3D Active Metric-Semantic SLAM.md",
+    "notes/cuda/img/vector_add_slo.png",
+    "notes/cuda/img/slo_v1.png",
+    "notes/cuda/img/roofline.png",
+    "notes/cuda/img/memory_chat.png",
+    "notes/cuda/Vector Add Optimization Example.md",
+    "notes/cuda/Roofline & Basic Analysis.md",
+    "notes/cuda/img",
+    "notes/cuda",
     "projects/nebula-vsearch/上篇：初识 Nebula Graph —— 向量类型支持.md",
-    "projects/active slam/img/data_flow.png",
-    "projects/active slam/img",
-    "projects/active slam/新建文件夹",
-    "paperreadings/activeslam/vins基础.md",
-    "projects/active slam",
-    "img.md",
-    "projects/nebula-vsearch/img/vector_value.png",
-    "projects/nebula-vsearch/understanding/img/vector.png",
-    "projects/nebula-vsearch/understanding/8.a kv life.md",
-    "projects/nebula-vsearch/img/metad.png",
-    "projects/nebula-vsearch/understanding/7.how to modify sql.md",
-    "projects/nebula-vsearch/understanding/6.raft-wal.md",
-    "projects/nebula-vsearch/img/exec.png",
-    "projects/nebula-vsearch/img/processor.png",
-    "projects/nebula-vsearch/understanding/img/exec.png",
-    "projects/nebula-vsearch/understanding/img/LRU.png",
-    "projects/nebula-vsearch/understanding/img/folly_async.png",
-    "projects/nebula-vsearch/understanding/img/doPut.png",
-    "projects/nebula-vsearch/understanding/5.nGQL life.md",
-    "projects/nebula-vsearch/understanding/4.folly future promise.md",
-    "projects/nebula-vsearch/understanding/3.concurrent lru cache.md",
-    "projects/nebula-vsearch/understanding/2.memory management.md",
-    "projects/nebula-vsearch/understanding/1.visitor pattern.md",
-    "index.md",
-    "projects/nebula-vsearch/Untitled",
-    "projects/nebula-vsearch/summary.md",
-    "notes/CSAPP/5-优化程序性能.md",
-    "blogs/posts/img",
-    "blogs/posts/rocksdb.md",
-    "blogs/posts/Implement of Concurrent.md",
-    "blogs/posts/CS144.md",
-    "blogs/posts/C++异步方案.md",
-    "blogs/posts/bustub通关指北.md",
-    "blogs/posts/bision debug.md",
-    "blogs/posts",
-    "projects/nebula-vsearch/understanding/img",
-    "projects/nebula-vsearch/understanding",
-    "blogs/CS144.md",
-    "blogs/bision debug.md",
-    "blogs/img",
-    "blogs/bustub通关指北.md",
-    "projects/nebula-vsearch/img",
+    "notes/nebula-vsearch/understanding/img/vector.png",
+    "notes/nebula-vsearch/understanding/img/processor.png",
+    "notes/nebula-vsearch/understanding/img/memory.png",
+    "notes/nebula-vsearch/understanding/img/LRU.png",
+    "notes/nebula-vsearch/understanding/img/folly_async.png",
+    "notes/nebula-vsearch/understanding/img/exec.png",
+    "notes/nebula-vsearch/understanding/img/doPut.png",
+    "notes/nebula-vsearch/understanding/img",
+    "notes/nebula-vsearch/understanding/8.a kv life.md",
+    "notes/nebula-vsearch/understanding/7.how to modify sql.md",
+    "notes/nebula-vsearch/understanding/6.raft-wal.md",
+    "notes/nebula-vsearch/understanding/5.nGQL life.md",
+    "notes/nebula-vsearch/understanding/4.folly future promise.md",
+    "notes/nebula-vsearch/understanding/3.concurrent lru cache.md",
+    "notes/nebula-vsearch/understanding/2.memory management.md",
+    "notes/nebula-vsearch/understanding/1.visitor pattern.md",
+    "notes/nebula-vsearch/implements/wal for vector type.md",
+    "notes/nebula-vsearch/implements/Match for vector property.md",
+    "notes/nebula-vsearch/implements/DDL for vector type.md",
+    "notes/nebula-vsearch/implements/DML for vector type.md",
+    "notes/nebula-vsearch/implements/Create Ann Index.md",
+    "notes/nebula-vsearch/上篇：初识 Nebula Graph —— 向量类型支持.md",
+    "notes/nebula-vsearch/understanding",
+    "notes/nebula-vsearch/summary.md",
+    "notes/nebula-vsearch/implements",
+    "notes/nebula-vsearch/img",
+    "notes/nebula-vsearch",
+    "notes/tiny-llm/RMSNorm & MLP.md",
+    "notes/tiny-llm/img",
+    "notes/tiny-llm/Group Query Attention.md",
+    "notes/tiny-llm/Batching Inference & KV Cache.md",
+    "notes/tiny-llm",
+    "projects/tiny-llm/Group Query Attention.md",
+    "projects/tiny-llm/RMSNorm & MLP.md",
+    "projects/tiny-llm/img",
+    "projects/tiny-llm/Batching Inference & KV Cache.md",
+    "projects/sglang/SGLang Schedular 技术变迁.md",
     "Untitled.canvas",
     "Untitled 1.canvas"
   ]

docs/notes/cuda/Roofline & Basic Analysis.md

@@ -0,0 +1,61 @@
+---
+title: Roofline & Basic Analysis
+date: 2025/10/20 23:15
+tags:
+  - LLMInference
+---
+
+# Roofline & Basic Analysis
+
+## AI
+
+## Roofline
+
+Roofline 模型（屋顶线模型）是一种用来**分析程序性能瓶颈**（计算受限还是带宽受限）的方法。  
+它把**计算性能**（FLOPs/s）和**访存性能**（Bytes/s）联系在一起，以可视化的方式展示性能上限。
+
+$$
+Achievable FLOPs=min(AI×Memory BW,Peak FLOPs)
+$$
+
+### 以 vector add 为例
+
+- 最简单的累加方式，每个 thread 负责一个线程的计算
+
+```cpp
+// FP32
+// ElementWise Add grid(N/256),
+// block(256) a: Nx1, b: Nx1, c: Nx1, c = elementwise_add(a, b)
+__global__ void vector_add_kernel(const float *a, const float *b, float *c,
+                                  int n) {
+  int idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (idx < n) {
+    c[idx] = a[idx] + b[idx];
+  }
+}
+```
+
+| 指标           | 数值             |
+| -------------- | ---------------- |
+| 每元素 FLOPs   | 1                |
+| 每元素 Bytes   | 12               |
+| AI             | 0.083 FLOPs/Byte |
+| Peak Bandwidth | 1008 GB/s        |
+| Peak Compute   | 82.6 TFLOPs/s    |
+
+```shell
+性能 (GFLOPs/s)
+↑
+|                   ────────────────  ← 82.6 TFLOPs/s (平顶线)
+|                  /
+|                 /
+|    o           /
+|   (VectorAdd) /
+|______________/__________________→ AI (FLOPs/Byte)
+                 0.083
+
+```
+
+- 左下角：访存主导（memory-bound）
+- 右上角：计算主导（compute-bound）
+- 中间交点：分界点（称为 **ridge point**）