Skip to content

Commit d4650f9

Browse files
lukstafilukstafi
committed
Tiny tweak of CLAUDE.md
1 parent b6e2154 commit d4650f9

File tree

1 file changed

+6
-6
lines changed

1 file changed

+6
-6
lines changed

‎CLAUDE.md‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -141,10 +141,11 @@ opam install cudajit # for CUDA backend
141141
- `c_syntax.ml` provides a functor with default C code generation patterns
142142
- `cc_backend.ml` uses defaults from `c_syntax.ml` with minimal overrides
143143
- `cuda_backend.ml` overrides more functions for CUDA-specific syntax (e.g., `__float2half`)
144-
- Both backends must provide `convert_precision` for type conversions
144+
- `metal_backend.ml` overrides using MSL-specific syntax
145+
- Backends must provide `convert_precision` for type conversions
145146
- Builtin functions (e.g., type conversions) must be implemented in:
146147
- `builtins.c` for C backends
147-
- `builtins_cuda_small.ml` for CUDA backend
148+
- `builtins_cuda.ml` for CUDA backend, `builtins_metal.ml` form Metal backend
148149
- When adding new precision types, ensure conversion functions exist in all backend builtins
149150

150151
### Syntax Extensions
@@ -166,7 +167,7 @@ opam install cudajit # for CUDA backend
166167

167168
## Common Development Tasks
168169

169-
### Adding New Operations
170+
### Adding New Primitive Operations
170171

171172
1. Add primitive operation to `arrayjit/lib/ops.ml`
172173
2. Implement interpretation in the same file
@@ -176,7 +177,7 @@ opam install cudajit # for CUDA backend
176177
### Debugging Backend Discrepancies
177178

178179
When outputs differ between backends:
179-
1. Compare runtime logs in `<backend>-<stream>-<stream>.log` files (might require minimizing test tensors)
180+
1. Compare runtime logs in `<backend>-<device>-<stream>.log` files (might require minimizing test tensors)
180181
2. Check generated code in `build_files/*.c` vs `*.cu` / `*.metal` for differences
181182
3. Common issues:
182183
- Incorrect type conversion in `convert_precision` overrides
@@ -208,5 +209,4 @@ When outputs differ between backends:
208209

209210
- Virtual nodes are inlined automatically (controlled by `virtualize_max_visits`)
210211
- Scalar constants can be inlined via `inline_scalar_constexprs=true`
211-
- Memory sharing optimizations through cross-stream tensor nodes
212-
- Backend-specific optimization levels configurable per backend
212+
- Memory sharing optimizations through cross-stream tensor nodes

0 commit comments

Comments
 (0)