diff options
author | James Ward <james.ward@arm.com> | 2023-01-23 17:13:37 +0000 |
---|---|---|
committer | James Ward <james.ward@arm.com> | 2023-01-31 11:45:21 +0000 |
commit | 736fd1a7e4083153ccc4cf360b44dd07b6788494 (patch) | |
tree | 42f34388dddb504650be0e17dbda8c9073223313 /reference_model/src/ops | |
parent | 2138a19ae830ec7d9ce5b15f15cbd7a22864bb8f (diff) | |
download | reference_model-736fd1a7e4083153ccc4cf360b44dd07b6788494.tar.gz |
Create MI tests for Type Conversion: CAST
* Add exclusion regex's to conformance generation
Signed-off-by: James Ward <james.ward@arm.com>
Change-Id: I15bef7451efd5662065060242d35bd7fa3381487
Diffstat (limited to 'reference_model/src/ops')
-rw-r--r-- | reference_model/src/ops/op_factory.cc | 7 | ||||
-rw-r--r-- | reference_model/src/ops/type_conversion.cc | 79 | ||||
-rw-r--r-- | reference_model/src/ops/type_conversion.h | 104 |
3 files changed, 184 insertions, 6 deletions
diff --git a/reference_model/src/ops/op_factory.cc b/reference_model/src/ops/op_factory.cc index 0d56161..76cf666 100644 --- a/reference_model/src/ops/op_factory.cc +++ b/reference_model/src/ops/op_factory.cc @@ -469,26 +469,33 @@ GraphNode* OpFactory::newOp(SubgraphTraverser* sgt, DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP16); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, BF16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BOOL); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT8); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP16); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BF16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BOOL); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT8); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP16); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BF16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT8); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT32); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, FP32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT8); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT32); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, FP32); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT8); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT16); DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT32); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, FP16); + DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, BF16); break; case Op_RESCALE: DEF_FACTORY_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT8); diff --git a/reference_model/src/ops/type_conversion.cc b/reference_model/src/ops/type_conversion.cc index f266675..dbedbad 100644 --- a/reference_model/src/ops/type_conversion.cc +++ b/reference_model/src/ops/type_conversion.cc @@ -15,6 +15,7 @@ #include "type_conversion.h" #include "quant_util.h" +#include "arith_util.h" #include "template_types.h" #include <cmath> #include "half.hpp" @@ -307,30 +308,88 @@ CastHelper<DType_BOOL, OutDtype>::CastHelper() template <DType InDtype> CastHelper<InDtype, DType_FP16>::CastHelper() { + // Integer data converted to fp16 (stored as fp32) fcn = [](InEigenType in) -> float { - half_float::half out = half_float::half_cast<half_float::half, InEigenType>(in); // Cast to half_float - return half_float::half_cast<float, half_float::half>(out); // Cast to float (underlying FP16 EigenType) + half_float::half h = half_float::half(in); + float out = half_float::half_cast<float, half_float::half>(h); + return out; + }; +} + +CastHelper<DType_FP32, DType_FP16>::CastHelper() +{ + // fp32 data converted to fp16 (stored as fp32) + fcn = [](float in) -> float { + float out = fpTrunc<DType_FP16>(in); // truncate required for conversion from higher precision + return out; + }; +} + +template <DType InDtype> +CastHelper<InDtype, DType_BF16>::CastHelper() +{ + // Integer data converted to bf16 (stored as fp32) + fcn = [](InEigenType in) -> float { + float out = (float)in; // default cast to float is round_to_nearest_float() + return out; + }; +} + +CastHelper<DType_FP32, DType_BF16>::CastHelper() +{ + // fp32 data converted to bf16 (stored as fp32) + fcn = [](float in) -> float { + return fpTrunc<DType_BF16>(in); // truncate required for conversions from higher precision }; } template <DType OutDtype> CastHelper<DType_FP16, OutDtype>::CastHelper() { - // Assuming InEigenType = float. + // fp16 data (stored as fp32) converted to integer fcn = [](float in) -> OutEigenType { - // Perform initial rounding in half-precision then cast back to float - half_float::half h = half_float::half_cast<half_float::half, float>(in); + // Cast from float representation back to half_float before rounding + half_float::half h = half_float::half(in); h = std::round(h); - OutEigenType out = half_float::half_cast<float, half_float::half>(h); + OutEigenType out = half_float::half_cast<OutEigenType, half_float::half>(h); out = std::max<OutEigenType>(out, OutMin); out = std::min<OutEigenType>(out, OutMax); return out; }; } +CastHelper<DType_FP16, DType_FP32>::CastHelper() +{ + // No-op since fp16 values treated internally as their fp32 representation + fcn = [](float in) -> OutEigenType { + return in; + }; +} + +template <DType OutDtype> +CastHelper<DType_BF16, OutDtype>::CastHelper() +{ + // bf16 data (stored as fp32) converted to integer + fcn = [](float in) -> OutEigenType { + OutEigenType out = std::round(in); + out = std::max<OutEigenType>(out, OutMin); + out = std::min<OutEigenType>(out, OutMax); + return out; + }; +} + +CastHelper<DType_BF16, DType_FP32>::CastHelper() +{ + // No-op since bf16 values treated as truncated fp32 internally + fcn = [](InEigenType in) -> OutEigenType { + return in; + }; +} + template <DType InDtype> CastHelper<InDtype, DType_FP32>::CastHelper() { + // Integer data converted to fp32 fcn = [](InEigenType in) -> float { float out = (OutEigenType)in; // default cast to float is round_to_nearest_float() return out; @@ -340,6 +399,7 @@ CastHelper<InDtype, DType_FP32>::CastHelper() template <DType OutDtype> CastHelper<DType_FP32, OutDtype>::CastHelper() { + // fp32 data converted to integer fcn = [](float in) -> OutEigenType { OutEigenType out = std::round(in); out = std::max<OutEigenType>(out, OutMin); @@ -356,26 +416,33 @@ DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, BOOL); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, INT32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP16); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, BF16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT8, FP32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BOOL); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, INT32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP16); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, BF16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT16, FP32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BOOL); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, INT16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP16); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, BF16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, INT32, FP32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, INT32); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP16, FP32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, INT32); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, BF16, FP32); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, INT32); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, FP16); +DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpCast, FP32, BF16); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT8); DEF_INSTANTIATE_RANK0_6_ONE_RANK_TWO_TYPE(OpRescale, INT8, INT16); diff --git a/reference_model/src/ops/type_conversion.h b/reference_model/src/ops/type_conversion.h index b0de30c..e2fc6e2 100644 --- a/reference_model/src/ops/type_conversion.h +++ b/reference_model/src/ops/type_conversion.h @@ -136,6 +136,76 @@ private: FcnType fcn; }; +template <> +class CastHelper<DType_FP32, DType_FP16> +{ +public: + using InEigenType = typename GetEigenType<DType_FP32>::type; + using OutEigenType = typename GetEigenType<DType_FP16>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + +template <DType InDtype> +class CastHelper<InDtype, DType_BF16> +{ +public: + using InEigenType = typename GetEigenType<InDtype>::type; + using OutEigenType = typename GetEigenType<DType_BF16>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + +template <DType OutDtype> +class CastHelper<DType_BF16, OutDtype> +{ +public: + using InEigenType = typename GetEigenType<DType_BF16>::type; + using OutEigenType = typename GetEigenType<OutDtype>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + static constexpr int32_t OutMin = GetQMin<OutDtype>::value; + static constexpr int32_t OutMax = GetQMax<OutDtype>::value; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + +template <> +class CastHelper<DType_FP32, DType_BF16> +{ +public: + using InEigenType = typename GetEigenType<DType_FP32>::type; + using OutEigenType = typename GetEigenType<DType_BF16>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + template <DType InDtype> class CastHelper<InDtype, DType_FP32> { @@ -153,6 +223,40 @@ private: FcnType fcn; }; +template <> +class CastHelper<DType_FP16, DType_FP32> +{ +public: + using InEigenType = typename GetEigenType<DType_FP16>::type; + using OutEigenType = typename GetEigenType<DType_FP32>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + +template <> +class CastHelper<DType_BF16, DType_FP32> +{ +public: + using InEigenType = typename GetEigenType<DType_BF16>::type; + using OutEigenType = typename GetEigenType<DType_FP32>::type; + using FcnType = std::function<OutEigenType(InEigenType)>; + CastHelper(); + const FcnType& get_fcn() const + { + return fcn; + } + +private: + FcnType fcn; +}; + template <DType OutDtype> class CastHelper<DType_FP32, OutDtype> { |