docs: build curated resource hub

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: LessUp

docs/.vitepress/config.ts

@@ -36,7 +36,7 @@ export default withMermaid(defineConfig({
           { text: 'Home', link: '/en/', activeMatch: '^/en/$' },
           { text: 'Architecture', link: '/en/architecture/', activeMatch: '^/en/architecture' },
           { text: 'Methodology', link: '/en/methodology/', activeMatch: '^/en/(methodology|learning-path|kernel-.*|optimization-playbook)' },
-          { text: 'Resources', link: '/en/references', activeMatch: '^/en/(references|cuda-memory-cheatsheet|performance-casebook)' },
+          { text: 'Resources', link: '/en/resources/', activeMatch: '^/en/(resources|references|cuda-memory-cheatsheet|performance-casebook)' },
           { text: 'Validation', link: '/en/validation/', activeMatch: '^/en/(validation|benchmark-results)' },
           { text: 'Support', link: '/en/getting-started', activeMatch: '^/en/getting-started' },
         ],
@@ -68,9 +68,11 @@ export default withMermaid(defineConfig({
             {
               text: 'Resources',
               items: [
+                { text: 'Resources Hub', link: '/en/resources/' },
+                { text: 'Further Reading Routes', link: '/en/resources/further-reading' },
                 { text: 'CUDA Memory Cheat Sheet', link: '/en/cuda-memory-cheatsheet' },
                 { text: 'Performance Casebook', link: '/en/performance-casebook' },
-                { text: 'References', link: '/en/references' },
+                { text: 'Curated References', link: '/en/references' },
               ],
             },
             {
@@ -105,7 +107,7 @@ export default withMermaid(defineConfig({
           { text: '首页', link: '/zh/', activeMatch: '^/zh/$' },
           { text: '架构', link: '/zh/architecture/', activeMatch: '^/zh/architecture' },
           { text: '方法论', link: '/zh/methodology/', activeMatch: '^/zh/(methodology|learning-path|kernel-.*|optimization-playbook)' },
-          { text: '资源', link: '/zh/references', activeMatch: '^/zh/(references|cuda-memory-cheatsheet|performance-casebook)' },
+          { text: '资源', link: '/zh/resources/', activeMatch: '^/zh/(resources|references|cuda-memory-cheatsheet|performance-casebook)' },
           { text: '验证', link: '/zh/validation/', activeMatch: '^/zh/(validation|benchmark-results)' },
           { text: '支持', link: '/zh/getting-started', activeMatch: '^/zh/getting-started' },
         ],
@@ -137,9 +139,11 @@ export default withMermaid(defineConfig({
             {
               text: '资源',
               items: [
+                { text: '资源中心', link: '/zh/resources/' },
+                { text: '延伸阅读路线', link: '/zh/resources/further-reading' },
                 { text: 'CUDA 内存速查表', link: '/zh/cuda-memory-cheatsheet' },
                 { text: '性能案例库', link: '/zh/performance-casebook' },
-                { text: '参考文献', link: '/zh/references' },
+                { text: '参考资料清单', link: '/zh/references' },
               ],
             },
             {

docs/en/architecture/index.md

@@ -83,8 +83,9 @@ This is not just process documentation. It affects how the architecture is narra
 4. Read [Tensor Core Path](/en/architecture/tensor-core-path) before interpreting WMMA benchmark numbers.
 5. Use the existing kernel deep dives when you want implementation detail instead of system rationale.
 
-## Related references
+## Related resources
 
+- [Resources Hub](/en/resources/)
 - [Validation Overview](/en/validation/)
 - [Learning Path](/en/learning-path)
 - [Getting Started](/en/getting-started)

docs/en/architecture/memory-flow.md

@@ -129,3 +129,4 @@ Padding away bank conflicts and keeping fallback logic explicit are both example
 - [Double Buffer](/en/kernel-double-buffer)
 - [Tensor Core WMMA](/en/kernel-tensor-core)
 - [CUDA Memory Cheat Sheet](/en/cuda-memory-cheatsheet)
+- [Resources Hub](/en/resources/)

docs/en/cuda-memory-cheatsheet.md

@@ -4,15 +4,19 @@ title: CUDA Memory Cheat Sheet
 
 # CUDA Memory Cheat Sheet
 
-A compact memory reference for reading SGEMM kernels faster
+This quick sheet is now part of the [Resources Hub](/en/resources/). Use it when you need to re-orient yourself before reopening kernel code, profiler output, or WMMA constraints.
 
+## When this page is most useful
 
+- Right before reading [Memory Flow](/en/architecture/memory-flow) or [Tiled Kernel](/en/kernel-tiled) again.
+- When a benchmark changed and you need a fast checklist before blaming occupancy or Tensor Cores.
+- When you want to explain memory behavior in an interview without reopening the full CUDA manuals.
 
 ## Coalescing quick rules
 
 - Consecutive threads in a warp should touch consecutive addresses whenever possible.
 - Accessing `B[k * N + col]` with large `N` can become stride-heavy for neighboring threads.
-- Tiling is not just reuse; it is also a way to reshape access into coalesced loads.
+- Tiling is not only about reuse; it also reshapes access into more coalesced loads.
 
 ```mermaid
 flowchart LR
@@ -21,25 +25,34 @@ flowchart LR
     C --> D[Global memory writeback]
 ```
 
+## Shared-memory watchpoints
 
+| Question | Why it matters |
+|---|---|
+| Are threads writing a tile layout that later reads back contiguously? | Shared memory only helps when it fixes a global-memory access problem instead of creating a local one. |
+| Did padding or index remapping remove bank conflicts on the hot path? | Bank conflicts can erase the benefit of otherwise-good tiling choices. |
+| Did the extra shared-memory footprint change occupancy enough to matter? | Some tiling wins disappear if the launch geometry becomes too constrained. |
 
 ## Tensor Core memory notes
 
 | Topic | What to remember |
-|-------|------------------|
-| Alignment constraints | WMMA path expects dimensions aligned to 16 for efficient fragment handling |
-| Data conversion | End-to-end timing includes conversion and wrapper logic |
-| Safe behavior | Non-friendly shapes should fall back to FP32 path |
-| Reporting | Distinguish end-to-end and compute-only results |
-
-
+|---|---|
+| Alignment constraints | WMMA paths expect dimensions aligned to fragment-friendly sizes, typically 16, for efficient fragment handling. |
+| Data conversion | End-to-end timing includes conversion and wrapper logic, not just the fused matrix multiply. |
+| Safe behavior | Non-friendly shapes should fall back to the FP32 path instead of forcing a misleading WMMA result. |
+| Reporting | Distinguish end-to-end numbers from compute-only numbers before comparing implementations. |
 
 ## Fast checklist when reading a kernel
 
-1. Can I explain the global memory access order for one warp?
-2. Is shared memory layout conflict-aware?
+1. Can I explain the global-memory access order for one warp?
+2. Does shared-memory layout reduce conflicts rather than merely move data around?
 3. Are register accumulators bounded and intentional?
 4. Is Tensor Core fallback behavior explicit?
 5. Do benchmark labels match what the kernel path really measures?
 
----
+## Where to go next
+
+- [Resources Hub](/en/resources/)
+- [Further Reading Routes](/en/resources/further-reading)
+- [Curated References](/en/references)
+- [Diagnosis Loop](/en/methodology/diagnosis-loop)

docs/en/index.md

@@ -16,9 +16,9 @@ title: SGEMM Architecture Whitepaper
         the implementation ladder and the evidence model from the homepage alone.
       </p>
       <div class="home-action-row">
-        <a class="btn" href="/en/architecture/">Read the architecture map</a>
-        <a class="btn btn-outline" href="/en/methodology/">Follow the methodology</a>
-        <a class="btn btn-outline" href="/en/validation/">Check the validation boundary</a>
+        <a class="btn" href="./architecture/">Read the architecture map</a>
+        <a class="btn btn-outline" href="./methodology/">Follow the methodology</a>
+        <a class="btn btn-outline" href="./validation/">Check the validation boundary</a>
         <a class="btn btn-outline" href="https://github.com/LessUp/sgemm-optimization">GitHub</a>
       </div>
       <div class="home-kicker-row">
@@ -41,7 +41,7 @@ title: SGEMM Architecture Whitepaper
       <div class="signal-card">
         <div class="signal-title">Validation Boundary</div>
         <div class="signal-value">CI is not a GPU</div>
-        <div class="signal-note">Hosted automation proves build and structure health; local hardware proves runtime and performance.</div>
+        <div class="signal-note">Hosted automation proves docs and governance health; local hardware proves CUDA runtime and performance.</div>
       </div>
       <div class="signal-card">
         <div class="signal-title">Reader Model</div>
@@ -144,56 +144,56 @@ flowchart LR
     <h3>I need the big picture first</h3>
     <p>Start with the system view of the repository, then use validation context to interpret what the architecture is allowed to claim.</p>
     <div class="route-links">
-      <a href="/en/architecture/">Architecture</a>
-      <a href="/en/validation/">Validation overview</a>
+      <a href="./architecture/">Architecture</a>
+      <a href="./validation/">Validation overview</a>
     </div>
   </div>
   <div class="route-card">
     <h3>I want to learn the optimization ladder in order</h3>
     <p>Use the staged reading path when you want each performance concept to build on the previous kernel rather than jump straight to WMMA.</p>
     <div class="route-links">
-      <a href="/en/learning-path">Learning Path</a>
-      <a href="/en/architecture/">Architecture overview</a>
+      <a href="./learning-path">Learning Path</a>
+      <a href="./architecture/">Architecture overview</a>
     </div>
   </div>
   <div class="route-card">
     <h3>I want tuning heuristics and follow-up material</h3>
     <p>Use the methodology and resource surfaces together when you need actionable optimization guidance plus technical lineage.</p>
     <div class="route-links">
-      <a href="/en/methodology/diagnosis-loop">Diagnosis Loop</a>
-      <a href="/en/references">References</a>
+      <a href="./methodology/diagnosis-loop">Diagnosis Loop</a>
+      <a href="./resources/">Resources hub</a>
     </div>
   </div>
   <div class="route-card">
     <h3>I want to reproduce or audit the claims</h3>
     <p>Use the support and validation surfaces together to understand what to run locally, what CI already proves, and how results should be interpreted.</p>
     <div class="route-links">
-      <a href="/en/getting-started">Getting Started</a>
-      <a href="/en/validation/">Validation overview</a>
+      <a href="./getting-started">Getting Started</a>
+      <a href="./validation/">Validation overview</a>
     </div>
   </div>
 </div>
 
 ## Resource hub entry points
 
 <div class="knowledge-grid">
-  <a class="knowledge-card" href="/en/architecture/">
+  <a class="knowledge-card" href="./architecture/">
     <h3>Architecture</h3>
     <p>The structural map of the kernel ladder, interface boundaries, and the decisions that hold the implementation together.</p>
   </a>
-  <a class="knowledge-card" href="/en/methodology/">
+  <a class="knowledge-card" href="./methodology/">
     <h3>Methodology</h3>
     <p>The canonical optimization workflow: staged learning, benchmark discipline, and diagnosis logic tied back to the architecture.</p>
   </a>
-  <a class="knowledge-card" href="/en/references">
+  <a class="knowledge-card" href="./resources/">
     <h3>Resources</h3>
-    <p>The source trail behind the project: official docs, papers, and mature repositories mapped to concrete decisions.</p>
+    <p>A curated learning hub that explains which docs, papers, tools, and follow-up routes matter after each part of the whitepaper.</p>
   </a>
-  <a class="knowledge-card" href="/en/validation/">
+  <a class="knowledge-card" href="./validation/">
     <h3>Validation</h3>
     <p>The evidence surface for correctness budgets, benchmark scope, fallback behavior, and what the published numbers actually mean.</p>
   </a>
-  <a class="knowledge-card" href="/en/getting-started">
+  <a class="knowledge-card" href="./getting-started">
     <h3>Support</h3>
     <p>The practical entry point for cloning, building, testing, and running the project with the right local-versus-CI expectations.</p>
   </a>

docs/en/interview-playbook.md

@@ -89,4 +89,4 @@ Strong answer pattern:
 - [Kernel Ladder](/en/architecture/kernel-ladder)
 - [Validation Overview](/en/validation/)
 - [Learning Path](/en/learning-path)
-- [References](/en/references)
+- [Resources Hub](/en/resources/)

docs/en/project-highlights.md

@@ -28,4 +28,4 @@ This page is now a quick orientation surface. The full “why this design exists
 2. [Kernel Ladder](/en/architecture/kernel-ladder)
 3. [Validation Overview](/en/validation/)
 4. [Interview Playbook](/en/interview-playbook)
-5. [References](/en/references)
+5. [Resources Hub](/en/resources/)