Repository navigation
[v0.2.5+] Support Operator #59
Description
PyGPUkit – Core Operator Coverage
1. Elementwise Operations
| Operation | FP32 | TF32 | FP16 | BF16 | Notes |
|---|---|---|---|---|---|
| add | ✅ | — | 🔜 | 🔜 | Implemented |
| sub | 🔜 | — | 🔜 | 🔜 | Planned |
| mul | ✅ | — | 🔜 | 🔜 | Implemented |
| div | 🔜 | — | 🔜 | 🔜 | Planned |
| exp | 🔜 | — | 🔜 | 🔜 | SFU-bound |
| log | 🔜 | — | 🔜 | 🔜 | SFU-bound |
| relu | 🔜 | — | 🔜 | 🔜 | Can be fused later |
2. GEMM Operations
| Operation | FP32 | TF32 | FP16 | BF16 | Notes |
|---|---|---|---|---|---|
| matmul | ✅ | ✅ | 🔜 | 🔜 | TensorCore on Ampere+ |
Current Performance (v0.2.3):
- FP32: 18 TFLOPS (RTX 3090 Ti)
- TF32: 27.38 TFLOPS (RTX 3090 Ti)
3. Reduction Operations
| Operation | FP32 | TF32 | FP16 | BF16 | Notes |
|---|---|---|---|---|---|
| sum | 🔜 | — | 🔜 | 🔜 | Tree-based reduction |
| mean | 🔜 | — | 🔜 | 🔜 | sum + scale |
| max | 🔜 | — | 🔜 | 🔜 | Warp + block reduction |
4. Memory Operations
| Operation | Status | Notes |
|---|---|---|
| copy | ✅ | Device↔Device, Host↔Device |
| reshape / view | ✅ | Zero-copy metadata only |
| contiguous | 🔜 | Layout-aware kernel required |
Recommended Milestones
| Version | Focus |
|---|---|
| v0.2.4 | Driver-only runtime (no CUDA Toolkit) ✅ |
| v0.2.5 | JIT stabilization, cache persistence |
| v0.2.6 | Elementwise ops (sub, div, exp, log, relu) |
| v0.2.7 | Reductions + contiguous |
| v0.3.0 | FP16 / BF16 mixed precision |
Design Principle
GEMM proves performance.
Elementwise + memory prove usability.
Scheduler proves value.
Activity
feat(ops): add sub, div, exp, log, relu elementwise operations (#59)
v0.2.5: FP16/BF16 support, TF32 optimization, reduction ops (#61)
m96-chan commentedon Dec 15, 2025
Completed in v0.2.5
Operator Overloads Added
Additional Elementwise Ops
gpk.sub()- subtractiongpk.div()- divisiongpk.exp()- exponentialgpk.log()- natural logarithmgpk.relu()- ReLU activationReduction Operations
gpk.sum(a)- sum of all elementsgpk.mean(a)- mean of all elementsgpk.max(a)- maximum elementPR: #61