wehub-resource-sync
380 lines
14 KiB
C++
380 lines
14 KiB
C++
#include <helpers/ConstantTadHelper.h>
|
|
#include <helpers/Loops.h>
|
|
#include <helpers/OmpLaunchHelper.h>
|
|
#include <helpers/ShapeBuilders.h>
|
|
#include <loops/legacy_ops.h>
|
|
#include <loops/reduce_float.h>
|
|
#include <system/op_boilerplate.h>
|
|
#include <types/types.h>
|
|
#include <algorithm>
|
|
|
|
using namespace simdOps;
|
|
|
|
namespace functions {
|
|
namespace reduce {
|
|
|
|
// =============================================================================
|
|
// TYPE-SAFE UTILITIES FOR ALL NUMERIC TYPES INCLUDING FLOAT16
|
|
// =============================================================================
|
|
|
|
namespace SafeTypeUtils {
|
|
|
|
/**
|
|
* @brief Type-safe array initialization that works with float16 and all other types
|
|
*/
|
|
template<typename T>
|
|
SD_INLINE void initializeArray(T* array, size_t count) {
|
|
if constexpr (std::is_arithmetic_v<T>) {
|
|
// For arithmetic types including float16, use loop initialization
|
|
for (size_t i = 0; i < count; i++) {
|
|
array[i] = static_cast<T>(0);
|
|
}
|
|
} else {
|
|
// For non-arithmetic types, use default initialization
|
|
std::fill_n(array, count, T{});
|
|
}
|
|
}
|
|
|
|
/**
|
|
* @brief Safe type conversion for mixed-type operations
|
|
*/
|
|
template<typename From, typename To>
|
|
SD_INLINE constexpr To safeCast(const From& value) {
|
|
return static_cast<To>(value);
|
|
}
|
|
|
|
/**
|
|
* @brief Convert parameter arrays between types safely
|
|
*/
|
|
template<typename SourceType, typename TargetType>
|
|
SD_INLINE void convertParams(const SourceType* source, TargetType* target, size_t count = 8) {
|
|
if (source && target) {
|
|
for (size_t i = 0; i < count; i++) {
|
|
target[i] = safeCast<SourceType, TargetType>(source[i]);
|
|
}
|
|
}
|
|
}
|
|
|
|
/**
|
|
* @brief Determine appropriate parameter type for mixed operations
|
|
* For float16, use float for better compatibility; otherwise use Z type
|
|
*/
|
|
template<typename X, typename Z>
|
|
struct CompatibleParamType {
|
|
using type = typename std::conditional_t<
|
|
std::is_same_v<Z, float16> || std::is_same_v<Z, bfloat16>,
|
|
float, // Use float for half-precision types
|
|
Z // Use Z for all other types
|
|
>;
|
|
};
|
|
|
|
} // namespace SafeTypeUtils
|
|
|
|
// =============================================================================
|
|
// REDUCE FLOAT FUNCTION IMPLEMENTATION WITH FLOAT16 SUPPORT
|
|
// =============================================================================
|
|
|
|
template <typename X, typename Z>
|
|
template <typename OpType>
|
|
void SD_HOST ReduceFloatFunction<X, Z>::execScalar(const void *vx, const sd::LongType *xShapeInfo, void *vextraParams,
|
|
void *vz, const sd::LongType *zShapeInfo) {
|
|
auto x = reinterpret_cast<const X *>(vx);
|
|
auto z = reinterpret_cast<Z *>(vz);
|
|
|
|
// Convert to Z* for consistency with macro expectations
|
|
Z *extraParams = nullptr;
|
|
Z convertedParams[8];
|
|
|
|
if (vextraParams != nullptr) {
|
|
if constexpr (std::is_same_v<Z, X>) {
|
|
extraParams = reinterpret_cast<Z*>(vextraParams);
|
|
} else {
|
|
// Convert parameters to Z type
|
|
auto originalParams = reinterpret_cast<X*>(vextraParams);
|
|
for (int i = 0; i < 8; ++i) {
|
|
convertedParams[i] = static_cast<Z>(originalParams[i]);
|
|
}
|
|
extraParams = convertedParams;
|
|
}
|
|
}
|
|
|
|
const auto length = shape::length(xShapeInfo);
|
|
|
|
if (shape::isEmptyConst(xShapeInfo)) {
|
|
z[0] = static_cast<Z>(OpType::startingValue(x));
|
|
return;
|
|
}
|
|
|
|
if (sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) {
|
|
if (sd::ArrayOptions::arrayType(zShapeInfo) == sd::ArrayType::EMPTY) return;
|
|
const auto startingVal = static_cast<Z>(OpType::startingValue(x));
|
|
|
|
for (sd::LongType i = 0; i < length; i++) {
|
|
z[i] = startingVal;
|
|
}
|
|
return;
|
|
}
|
|
|
|
auto startingValue = static_cast<typename OpType::InterType>(OpType::startingValue(x));
|
|
int maxThreads = sd::math::sd_min<int>(64, sd::Environment::getInstance().maxThreads());
|
|
typename OpType::InterType intermediate[64];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (auto e = 0; e < maxThreads; e++) {
|
|
intermediate[e] = startingValue;
|
|
}
|
|
|
|
sd::LongType xRank = shape::rank(xShapeInfo);
|
|
sd::LongType* xShape = shape::shapeOf(xShapeInfo);
|
|
sd::LongType* xStride = shape::stride(xShapeInfo);
|
|
|
|
if(shape::isViewConst(xShapeInfo)) {
|
|
auto func = PRAGMA_THREADS_FOR {
|
|
for (auto i = start; i < stop; i++) {
|
|
sd::LongType coords[SD_MAX_RANK];
|
|
INDEX2COORDS(i, xRank, xShape, coords);
|
|
sd::LongType indexOffset;
|
|
COORDS2INDEX(xRank, xStride, coords, indexOffset);
|
|
|
|
auto opResult = OpType::op(x[indexOffset], extraParams);
|
|
intermediate[thread_id] = OpType::update(
|
|
intermediate[thread_id],
|
|
opResult,
|
|
extraParams
|
|
);
|
|
}
|
|
};
|
|
maxThreads = samediff::Threads::parallel_for(func, 0, length, 1, maxThreads);
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int e = 1; e < maxThreads; e++) {
|
|
intermediate[0] = OpType::merge(intermediate[0], intermediate[e], extraParams);
|
|
}
|
|
|
|
z[0] = OpType::postProcess(intermediate[0], length, extraParams);
|
|
} else {
|
|
auto func = PRAGMA_THREADS_FOR {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto opResult = OpType::op(x[i], extraParams);
|
|
intermediate[thread_id] = OpType::update(
|
|
intermediate[thread_id],
|
|
opResult,
|
|
extraParams
|
|
);
|
|
}
|
|
};
|
|
maxThreads = samediff::Threads::parallel_for(func, 0, length, 1, maxThreads);
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (int e = 1; e < maxThreads; e++) {
|
|
intermediate[0] = OpType::merge(intermediate[0], intermediate[e], extraParams);
|
|
}
|
|
|
|
z[0] = OpType::postProcess(intermediate[0], length, extraParams);
|
|
}
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
template <typename OpType>
|
|
Z SD_HOST ReduceFloatFunction<X, Z>::execScalar(const void *vx, const sd::LongType *xShapeInfo, void *vextraParams) {
|
|
auto x = reinterpret_cast<const X *>(vx);
|
|
|
|
// Convert to Z* for compatibility with OpType::op
|
|
Z *extraParams = nullptr;
|
|
Z convertedParams[8];
|
|
|
|
if (vextraParams != nullptr) {
|
|
if constexpr (std::is_same_v<Z, X>) {
|
|
extraParams = reinterpret_cast<Z*>(vextraParams);
|
|
} else {
|
|
// Convert the parameters to Z type
|
|
auto originalParams = reinterpret_cast<X*>(vextraParams);
|
|
for (int i = 0; i < 8; ++i) {
|
|
convertedParams[i] = static_cast<Z>(originalParams[i]);
|
|
}
|
|
extraParams = convertedParams;
|
|
}
|
|
}
|
|
|
|
const sd::LongType length = shape::length(xShapeInfo);
|
|
auto startingValue = static_cast<typename OpType::InterType>(OpType::startingValue(x));
|
|
|
|
sd::LongType xRank = shape::rank(xShapeInfo);
|
|
sd::LongType *xShape = shape::shapeOf(xShapeInfo);
|
|
sd::LongType *xStride = shape::stride(xShapeInfo);
|
|
|
|
for (sd::LongType i = 0; i < length; i++) {
|
|
sd::LongType coords[SD_MAX_RANK];
|
|
INDEX2COORDS(i, xRank, xShape, coords);
|
|
sd::LongType offset;
|
|
COORDS2INDEX(xRank, xStride, coords, offset);
|
|
|
|
auto opResult = OpType::op(x[offset], extraParams);
|
|
startingValue = OpType::update(startingValue, opResult, extraParams);
|
|
}
|
|
|
|
return OpType::postProcess(startingValue, length, extraParams);
|
|
}
|
|
template <typename X, typename Z>
|
|
template <typename OpType>
|
|
void SD_HOST ReduceFloatFunction<X, Z>::exec(sd::memory::Workspace *workspace, const void *vx,
|
|
const sd::LongType *xShapeInfo, void *vextraParams, void *vz,
|
|
const sd::LongType *zShapeInfo, const sd::LongType *dims) {
|
|
const X *x = reinterpret_cast<const X *>(vx);
|
|
Z *z = reinterpret_cast<Z *>(vz);
|
|
|
|
using CompatibleParamType = typename SafeTypeUtils::CompatibleParamType<X, Z>::type;
|
|
CompatibleParamType *compatibleExtraParams = nullptr;
|
|
CompatibleParamType convertedParams[8];
|
|
SafeTypeUtils::initializeArray(convertedParams, 8);
|
|
|
|
if (vextraParams != nullptr) {
|
|
if constexpr (std::is_same_v<X, CompatibleParamType>) {
|
|
compatibleExtraParams = reinterpret_cast<CompatibleParamType*>(vextraParams);
|
|
} else {
|
|
SafeTypeUtils::convertParams(reinterpret_cast<X*>(vextraParams), convertedParams, 8);
|
|
compatibleExtraParams = convertedParams;
|
|
}
|
|
}
|
|
|
|
const int xRank = shape::rank(xShapeInfo);
|
|
const int zRank = shape::rank(zShapeInfo);
|
|
|
|
if (sd::ArrayOptions::arrayType(xShapeInfo) == sd::ArrayType::EMPTY) {
|
|
const auto startingVal = std::is_same<OpType, simdOps::Mean<X, Z>>::value
|
|
? sd::DataTypeUtils::nanOrZero<Z>()
|
|
: SafeTypeUtils::safeCast<X, Z>(OpType::startingValue(x));
|
|
const auto zLen = shape::length(zShapeInfo);
|
|
if (z != nullptr)
|
|
for (sd::LongType i = 0; i < zLen; i++) z[i] = startingVal;
|
|
return;
|
|
}
|
|
|
|
if (shape::length(zShapeInfo) == 1) {
|
|
z[0] = execScalar<OpType>(x, xShapeInfo, compatibleExtraParams);
|
|
return;
|
|
}
|
|
|
|
if (OpType::requiresSpecialAccumulation) {
|
|
// FIXED: Handle execSpecial with flexible parameter types
|
|
// The enhanced macro provides template overloads that accept any arithmetic type
|
|
|
|
if constexpr (std::is_same_v<CompatibleParamType, sd::LongType>) {
|
|
// Direct call for sd::LongType parameters
|
|
OpType::execSpecial(x, xShapeInfo, compatibleExtraParams, z, zShapeInfo,
|
|
const_cast<sd::LongType *>(dims) + zRank, xRank - zRank,
|
|
nullptr, nullptr);
|
|
} else {
|
|
// Convert to sd::LongType for operations that specifically need it
|
|
sd::LongType longExtraParams[8];
|
|
SafeTypeUtils::initializeArray(longExtraParams, 8);
|
|
|
|
if (compatibleExtraParams != nullptr) {
|
|
SafeTypeUtils::convertParams(compatibleExtraParams, longExtraParams, 8);
|
|
}
|
|
|
|
// Use template overload that accepts sd::LongType*
|
|
OpType::execSpecial(x, xShapeInfo, longExtraParams, z, zShapeInfo,
|
|
const_cast<sd::LongType *>(dims) + zRank, xRank - zRank,
|
|
nullptr, nullptr);
|
|
}
|
|
return;
|
|
}
|
|
|
|
#ifdef SD_LOOPS_INLINED
|
|
sd::ReductionLoops<X, Z, CompatibleParamType>::template loopReduce<OpType>(workspace, x, xShapeInfo, z, zShapeInfo, dims, compatibleExtraParams);
|
|
#else
|
|
sd::ReductionFloatLoops<X, Z>::template innerloopReduce<OpType>(workspace, x, xShapeInfo, z, zShapeInfo, dims, compatibleExtraParams);
|
|
#endif
|
|
}
|
|
|
|
|
|
template <typename X, typename Z>
|
|
Z ReduceFloatFunction<X, Z>::execScalar(const int opNum, const void *x, const sd::LongType *xShapeInfo,
|
|
void *extraParams) {
|
|
RETURNING_DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams), REDUCE_FLOAT_OPS);
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
void ReduceFloatFunction<X, Z>::execScalar(const int opNum, const void *x, const sd::LongType *xShapeInfo,
|
|
void *extraParams, void *z, const sd::LongType *zShapeInfo) {
|
|
DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo), REDUCE_FLOAT_OPS);
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
template <typename OpType>
|
|
void SD_HOST ReduceFloatFunction<X, Z>::exec(const void *x, const sd::LongType *xShapeInfo, void *extraParams,
|
|
void *vresult, const sd::LongType *resultShapeInfo) {
|
|
auto z = reinterpret_cast<Z *>(vresult);
|
|
z[0] = execScalar<OpType>(x, xShapeInfo, extraParams);
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
template <typename OpType>
|
|
Z SD_HOST ReduceFloatFunction<X, Z>::execScalar(const void *vx, sd::LongType xEws, sd::LongType length,
|
|
void *vextraParams) {
|
|
auto x = reinterpret_cast<const X *>(vx);
|
|
|
|
using CompatibleParamType = typename SafeTypeUtils::CompatibleParamType<X, Z>::type;
|
|
CompatibleParamType *compatibleExtraParams = nullptr;
|
|
CompatibleParamType convertedParams[8];
|
|
SafeTypeUtils::initializeArray(convertedParams, 8);
|
|
|
|
if (vextraParams != nullptr) {
|
|
if constexpr (std::is_same_v<X, CompatibleParamType>) {
|
|
compatibleExtraParams = reinterpret_cast<CompatibleParamType*>(vextraParams);
|
|
} else {
|
|
SafeTypeUtils::convertParams(reinterpret_cast<X*>(vextraParams), convertedParams, 8);
|
|
compatibleExtraParams = convertedParams;
|
|
}
|
|
}
|
|
|
|
int maxThreads = sd::math::sd_min<int>(64, sd::Environment::getInstance().maxThreads());
|
|
using InterType = typename OpType::InterType;
|
|
InterType intermediate[64];
|
|
|
|
PRAGMA_OMP_SIMD
|
|
for (auto e = 0; e < maxThreads; e++) {
|
|
intermediate[e] = SafeTypeUtils::safeCast<X, InterType>(OpType::startingValue(x));
|
|
}
|
|
|
|
auto func = PRAGMA_THREADS_FOR {
|
|
if (xEws == 1) {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto opResult = OpType::op(x[i], compatibleExtraParams);
|
|
intermediate[thread_id] = OpType::update(
|
|
intermediate[thread_id],
|
|
SafeTypeUtils::safeCast<decltype(opResult), InterType>(opResult),
|
|
compatibleExtraParams
|
|
);
|
|
}
|
|
} else {
|
|
for (auto i = start; i < stop; i++) {
|
|
auto opResult = OpType::op(x[i * xEws], compatibleExtraParams);
|
|
intermediate[thread_id] = OpType::update(
|
|
intermediate[thread_id],
|
|
SafeTypeUtils::safeCast<decltype(opResult), InterType>(opResult),
|
|
compatibleExtraParams
|
|
);
|
|
}
|
|
}
|
|
};
|
|
|
|
maxThreads = samediff::Threads::parallel_for(func, 0, length, 1, maxThreads);
|
|
|
|
for (int e = 1; e < maxThreads; e++)
|
|
intermediate[0] = OpType::update(intermediate[0], intermediate[e], compatibleExtraParams);
|
|
|
|
return SafeTypeUtils::safeCast<InterType, Z>(OpType::postProcess(intermediate[0], length, compatibleExtraParams));
|
|
}
|
|
|
|
template <typename X, typename Z>
|
|
void ReduceFloatFunction<X, Z>::exec(int opNum, sd::memory::Workspace *workspace, const void *vx,
|
|
const sd::LongType *xShapeInfo, void *vextraParams, void *vz,
|
|
const sd::LongType *zShapeInfo, const sd::LongType *dims) {
|
|
DISPATCH_BY_OPNUM_TT(exec, PARAMS(workspace, vx, xShapeInfo, vextraParams, vz, zShapeInfo, dims), REDUCE_FLOAT_OPS);
|
|
}
|
|
|
|
|
|
} // namespace reduce
|
|
} // namespace functions
|