(shapes, output_shape, inputs, ctx: EinsumContext)
| 274 | |
| 275 | |
| 276 | def einsum_impl(shapes, output_shape, inputs, ctx: EinsumContext): |
| 277 | assert len(shapes) == len(inputs), "input size mismatch" |
| 278 | dim2firstop = dict() |
| 279 | dim2lastop = dict() |
| 280 | dims = set() |
| 281 | for i, shape in enumerate(shapes): |
| 282 | for dim in shape: |
| 283 | if dim not in output_shape: |
| 284 | dim2lastop[dim] = i |
| 285 | if dim not in dim2firstop: |
| 286 | dim2firstop[dim] = i |
| 287 | dims.add(dim) |
| 288 | result = EinsumOperand(shapes[0], inputs[0], ctx).dedup_dim() |
| 289 | for i in range(1, len(inputs)): |
| 290 | rhs = EinsumOperand(shapes[i], inputs[i], ctx).dedup_dim() |
| 291 | lshape = result.shape |
| 292 | batch_dims: Tuple[EinsumDimension, ...] = () |
| 293 | sum_dims: Tuple[EinsumDimension, ...] = () |
| 294 | left_dims: Tuple[EinsumDimension, ...] = () |
| 295 | right_dims: Tuple[EinsumDimension, ...] = () |
| 296 | reduce_dims: Tuple[EinsumDimension, ...] = () |
| 297 | lshape = result.shape |
| 298 | rshape = rhs.shape |
| 299 | for dim in lshape: |
| 300 | lastop = dim2lastop.get(dim, -1) |
| 301 | if lastop == i - 1: |
| 302 | reduce_dims = reduce_dims + (dim,) |
| 303 | elif lastop == i: |
| 304 | sum_dims = sum_dims + (dim,) |
| 305 | else: |
| 306 | if dim in rshape: |
| 307 | batch_dims = batch_dims + (dim,) |
| 308 | else: |
| 309 | left_dims = left_dims + (dim,) |
| 310 | for dim in rshape: |
| 311 | if dim not in lshape: |
| 312 | lastop = dim2lastop.get(dim, -1) |
| 313 | if lastop == i: |
| 314 | reduce_dims = reduce_dims + (dim,) |
| 315 | else: |
| 316 | right_dims = right_dims + (dim,) |
| 317 | result = result.reduce( |
| 318 | tuple(filter(lambda x: x not in reduce_dims, result.shape)) |
| 319 | ) |
| 320 | rhs = rhs.reduce(tuple(filter(lambda x: x not in reduce_dims, rhs.shape))) |
| 321 | result = einsum_matmul(result, rhs, batch_dims, sum_dims, left_dims, right_dims) |
| 322 | result = result.reduce( |
| 323 | tuple(filter(lambda x: x in output_shape, result.shape)) |
| 324 | ).transpose(output_shape) |
| 325 | return result._tracer |
| 326 | |
| 327 | |
| 328 | def diag_plane_interpret(inp: Tensor, inp_ndim: int, axes: List[int]) -> Tensor: |
no test coverage detected