| 382 | template <Path CompiledPaths, typename LhsScalar, typename RhsScalar, |
| 383 | typename DstScalar, typename Spec> |
| 384 | void DispatchMul(const Matrix<LhsScalar>& lhs, const Matrix<RhsScalar>& rhs, |
| 385 | const Spec& spec, Context* context, Matrix<DstScalar>* dst) { |
| 386 | static_assert(CompiledPaths != Path::kNone, "Must compile at least one Path"); |
| 387 | static_assert((CompiledPaths & ~kAllPaths) == Path::kNone, |
| 388 | "CompiledPaths must be a subset of ruy::kAllPaths"); |
| 389 | |
| 390 | gemmlowp::ScopedProfilingLabel label("Mul"); |
| 391 | |
| 392 | EnforceLayoutSupport<Spec>(lhs.layout, rhs.layout, dst->layout); |
| 393 | EnforceZeroPointSupport<Spec>(lhs.zero_point, rhs.zero_point, |
| 394 | dst->zero_point); |
| 395 | EnforceDstSpecSupport<Spec>(spec, dst->zero_point); |
| 396 | |
| 397 | // This should be a constant, for a given machine and CompiledPaths. |
| 398 | // There is a back door to override it for testing, but in production it will |
| 399 | // always be the "best" Path. I.e. the one with the newest SIMD instructions |
| 400 | // available on the present machine, and avoiding Path::kReference unless |
| 401 | // no other path is compiled. |
| 402 | // |
| 403 | // Unfortunately, it is not a *static* constant, since it depends on runtime |
| 404 | // detection of the available SIMD instructions. |
| 405 | Path the_path = context->GetPathToTake<CompiledPaths>(); |
| 406 | |
| 407 | // Production code should probably never execute Path::kReference. |
| 408 | // Path::kReference implements a Mul, not a TrMul like the rest of Ruy, so if |
| 409 | // that's what we need to do, then get it out of the way before going down the |
| 410 | // TrMul path. |
| 411 | if (the_path == Path::kReference) { |
| 412 | constexpr bool ReferenceMulIsEnabled = |
| 413 | (CompiledPaths & Path::kReference) != Path::kNone; |
| 414 | CompileTimeEnabledReferenceMul<ReferenceMulIsEnabled>::Run(lhs, rhs, spec, |
| 415 | dst); |
| 416 | return; |
| 417 | } |
| 418 | |
| 419 | // As described in the comment at the top of this file, Ruy internally |
| 420 | // converts Mul into TrMul. We handle that here. |
| 421 | // |
| 422 | // This is Ruy's main code path. |
| 423 | constexpr Path TrMulCompiledPaths = CompiledPaths & ~Path::kReference; |
| 424 | Matrix<LhsScalar> transposed_lhs(lhs); |
| 425 | Transpose(&transposed_lhs); |
| 426 | TrMulParams params; |
| 427 | CreateTrMulParams<TrMulCompiledPaths>(transposed_lhs, rhs, spec, context, dst, |
| 428 | the_path, ¶ms); |
| 429 | TrMul(¶ms, context); |
| 430 | } |
| 431 | |
| 432 | } // namespace ruy |
| 433 | |