The map
Every node in the nine months, grouped by section and ordered by week. All 124 are written; what changes is whether you have earned the right to open one.
00 Instrument 8 nodes
- Install the RISC-V toolchainW1 · written
- QEMU with RVV runningW1 · written
- Scalar matmul baselineW1 · written
- Timing harness with mediansW1 · written
- Elementwise and softmax baselinesW2 · written
- One-command benchmarkW2 · written
- Stock linalg.matmul end to endW2 · written
- Public repo and READMEW2 · written
01 First lowering 28 nodes
- MLIR dialects and operationsW3 · written
- Regions, blocks, and block argumentsW3 · written
- Traits and interfacesW3 · written
- RewritePattern and canonicalizationW3 · written
- Pass pipeline and analysesW3 · written
- Dialect conversionW3 · written
- linalg iteration spaceW4 · written
- linalg indexing mapsW4 · written
- Structured op contractW4 · written
- Tensors versus memrefsW4 · written
- Hand-lower matmul to scf.forW4 · written
- Tiling the iteration spaceW5 · written
- scf.forall tilingW5 · written
- Transform dialect tilingW5 · written
- One-shot bufferizationW5 · written
- Tile-size sweepW5 · written
- Vector dialectW6 · written
- linalg vectorizationW6 · written
- Fixed-width vector matmulW6 · written
- Vector to LLVM loweringW6 · written
- Scalable vector typesW7 · written
- Scalable vector loweringW7 · written
- RVV disassembly checkW7 · written
- vsetvli and LMUL semanticsW7 · written
- Rigorous benchmarkingW8 · written
- README speedup tableW8 · written
- Fixed versus scalable comparisonW8 · written
- Reproduce benchmark scriptW8 · written
02 Fast & real 32 nodes
- Register blockingW9 · written
- LMUL selectionW9 · written
- Vector register budgetW9 · written
- Microkernel qualityW9 · written
- Hand-written RVV intrinsics matmulW10 · written
- Compiler versus hand-tuned gapW10 · written
- Vector tail handlingW10 · written
- Codegen parameter sweepW10 · written
- Epilogue conceptW11 · written
- Bias fusionW11 · written
- Relu fusionW11 · written
- Fused matmul+bias+reluW11 · written
- Fusion legalityW12 · written
- Tile-level fusionW12 · written
- Fusion benchmarkW12 · written
- Layout propagationW12 · written
- Dynamic-shape IRW13 · written
- Dynamic tilingW13 · written
- Unranked handlingW13 · written
- Runtime-size validationW13 · written
- Dynamic vectorizationW14 · written
- Dynamic benchmarkW14 · written
- Shape specializationW14 · written
- Dynamic epilogue fusionW14 · written
- Roofline modelW15 · written
- Bottleneck evidenceW15 · written
- Next-lever analysisW15 · written
- Performance regression guardW15 · written
- Technical blog postW16 · written
- Reproducible artifactsW16 · written
- Remaining 2x storyW16 · written
- Publish checklistW16 · written
03 Extend 20 nodes
- Second-op decisionW17 · written
- Stable softmax numericsW17 · written
- Vector reduction loweringW18 · written
- Softmax loweringW18 · written
- Softmax benchmarkW18 · written
- Attention block decompositionW19 · written
- QK-transpose matmulW19 · written
- Attention softmax fusionW19 · written
- Probability-by-V matmulW20 · written
- Attention benchmarkW20 · written
- Lane B applications launchW20 · written
- Upstream target selectionW21 · written
- Upstream reproducerW21 · written
- Upstream patchW22 · written
- Upstream PR reviewW22 · written
- Upstream CI greenW23 · written
- Upstream regression testW23 · written
- Upstream merged or in reviewW24 · written
- Final writeupW25 · written
- Talk-ready slidesW26 · written
GPU 24 nodes
- GPU access and nvccW1 · written
- Vector-add kernelW1 · written
- CUDA execution modelW2 · written
- Coalesced global memoryW2 · written
- Shared-memory matmulW3 · written
- Register-blocked matmulW4 · written
- Closing the cuBLAS gapW4 · written
- Block reductionW5 · written
- Softmax kernelW5 · written
- Occupancy tuningW6 · written
- Nsight Compute analysisW6 · written
- Nsight Systems timelineW7 · written
- GPU rooflineW8 · written
- Triton programming modelW9 · written
- First Triton vector addW9 · written
- Triton matmulW10 · written
- Triton autotuningW11 · written
- Fused Triton softmaxW12 · written
- Online softmaxW13 · written
- FlashAttention-style kernelW13 · written
- Attention benchmarkW14 · written
- Triton codegen inspectionW14 · written
- Capstone kernelW17 · written
- GPU writeup and storyW20 · written
Theory 12 nodes
- Roofline paperW1 · written
- MLIR paperW2 · written
- linalg docsW2 · written
- Halide paperW4 · written
- TVM paperW6 · written
- Triton paperW12 · written
- FlashAttention paperW15 · written
- FlashAttention-2 paperW15 · written
- TorchInductor design docsW18 · written
- XLA and StableHLO overviewW18 · written
- Transformer primerW20 · written
- Vocabulary journalW20 · written