Skip to content

[ESIMD] Allow full autodeduction for scatter USM APIs accepting simd_view #13941

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 29, 2024
Merged

[ESIMD] Allow full autodeduction for scatter USM APIs accepting simd_view #13941

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
229 changes: 221 additions & 8 deletions sycl/include/sycl/ext/intel/esimd/memory.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -727,10 +727,46 @@ scatter(T *p, simd<OffsetT, N / VS> byte_offsets, simd<T, N> vals,
}
}

// template <typename T, int N, int VS = 1, typename OffsetT,
// typename PropertyListT = empty_properties_t>
// void scatter(T *p, simd<OffsetT, N / VS> byte_offsets, simd<T, N> vals,
// PropertyListT props = {}); // (usm-sc-2)
/// template <int VS = 1, typename OffsetT, typename ValuesSimdViewT, typename
/// T, int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
/// typename PropertyListT = empty_properties_t>
/// void scatter(T *p, simd<OffsetT, N / VS> byte_offsets, ValuesSimdViewT vals,
/// simd_mask<N / VS> mask, PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector. Access to
/// any element's memory location can be disabled via the input mask.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param mask The access mask.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS = 1, typename OffsetT, typename ValuesSimdViewT, typename T,
int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<ValuesSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, simd<OffsetT, N / VS> byte_offsets, ValuesSimdViewT vals,
simd_mask<N / VS> mask, PropertyListT props = {}) {
scatter<T, N, VS>(p, byte_offsets, vals.read(), mask, props);
}

/// template <typename T, int N, int VS = 1, typename OffsetT,
/// typename PropertyListT = empty_properties_t>
/// void scatter(T *p, simd<OffsetT, N / VS> byte_offsets, simd<T, N> vals,
/// PropertyListT props = {}); // (usm-sc-2)
///
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
Expand Down Expand Up @@ -759,10 +795,80 @@ scatter(T *p, simd<OffsetT, N / VS> byte_offsets, simd<T, N> vals,
scatter<T, N, VS>(p, byte_offsets, vals, Mask, props);
}

// template <typename T, int N, int VS = 1, typename OffsetSimdViewT,
// typename PropertyListT = empty_properties_t>
// void scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
// simd_mask<N / VS> mask, PropertyListT props = {}); // (usm-sc-3)
/// template <int VS = 1, typename OffsetSimdViewT, typename ValuesSimdViewT,
/// typename T, int N = ValuesSimdViewT::getSizeX() *
/// ValuesSimdViewT::getSizeY(), typename PropertyListT = empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, ValuesSimdViewT vals,
/// simd_mask<N / VS> mask, PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param mask The access mask.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS = 1, typename OffsetSimdViewT, typename ValuesSimdViewT, typename T,
int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<ValuesSimdViewT> &&
detail::is_simd_view_type_v<OffsetSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, OffsetSimdViewT byte_offsets, ValuesSimdViewT vals,
simd_mask<N / VS> mask, PropertyListT props = {}) {
scatter<T, N, VS>(p, byte_offsets.read(), vals.read(), mask, props);
}

/// template <int VS = 1, typename OffsetT, typename ValuesSimdViewT, typename
/// T, int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
/// typename PropertyListT = empty_properties_t>
/// void scatter(T *p, simd<OffsetT, N / VS> byte_offsets, ValuesSimdViewT vals,
/// PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS = 1, typename OffsetT, typename ValuesSimdViewT, typename T,
int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<ValuesSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, simd<OffsetT, N / VS> byte_offsets, ValuesSimdViewT vals,
PropertyListT props = {}) {
scatter<T, N, VS>(p, byte_offsets, vals.read(), props);
}

/// template <typename T, int N, int VS = 1, typename OffsetSimdViewT,
/// typename PropertyListT = empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
/// simd_mask<N / VS> mask, PropertyListT props = {}); // (usm-sc-3)
///
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
Expand Down Expand Up @@ -794,6 +900,75 @@ scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
scatter<T, N, VS>(p, byte_offsets.read(), vals, mask, props);
}

/// template <int VS, typename OffsetSimdViewT, typename T, int N, typename
/// PropertyListT = empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, simd<T,N> vals,
/// simd_mask<N / VS> mask, PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector. Access to
/// any element's memory location can be disabled via the input mask.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes
/// represented as a 'simd_view' object.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param mask The access mask.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS, typename OffsetSimdViewT, typename T, int N,
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<OffsetSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
simd_mask<N / VS> mask, PropertyListT props = {}) {
scatter<T, N, VS>(p, byte_offsets.read(), vals, mask, props);
}

/// template <int VS, typename OffsetSimdViewT, typename T, int N, typename
/// PropertyListT = empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, simd<T,N> vals,
/// PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector. Access to
/// any element's memory location can be disabled via the input mask.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes
/// represented as a 'simd_view' object.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS, typename OffsetSimdViewT, typename T, int N,
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<OffsetSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
PropertyListT props = {}) {
scatter<T, N, VS>(p, byte_offsets.read(), vals, props);
}

/// template <typename T, int N, int VS = 1, typename OffsetSimdViewT,
/// typename PropertyListT = empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
Expand Down Expand Up @@ -828,6 +1003,44 @@ scatter(T *p, OffsetSimdViewT byte_offsets, simd<T, N> vals,
scatter<T, N, VS>(p, byte_offsets.read(), vals, Mask, props);
}

/// template <int VS = 1, typename OffsetSimdViewT, typename
/// ValuesSimdViewT, typename T, int N = ValuesSimdViewT::getSizeX() *
/// ValuesSimdViewT::getSizeY(), typename PropertyListT =
/// empty_properties_t>
/// void scatter(T *p, OffsetSimdViewT byte_offsets, ValuesSimdViewT vals,
/// PropertyListT props = {});
///
/// Variation of the API that allows to use \c simd_view without specifying \c T
/// and \c N template parameters.
/// Writes ("scatters") elements of the input vector to different memory
/// locations. Each memory location is base address plus an offset - a
/// value of the corresponding element in the input offset vector.
/// @tparam VS Vector size. It can also be read as the number of writes per each
/// address. The parameter 'N' must be divisible by 'VS'. (VS > 1) is supported
/// only on DG2 and PVC and only for 4- and 8-byte element vectors.
/// @param p The base address.
/// @param byte_offsets the vector of 32-bit or 64-bit offsets in bytes
/// represented as a 'simd_view' object.
/// For each i, ((byte*)p + byte_offsets[i]) must be element size aligned.
/// If the alignment property is not passed, then it is assumed that each
/// accessed address is aligned by element-size.
/// @param vals The vector to scatter.
/// @param props The optional compile-time properties. Only 'alignment'
/// and cache hint properties are used.
template <
int VS = 1, typename OffsetSimdViewT, typename ValuesSimdViewT, typename T,
int N = ValuesSimdViewT::getSizeX() * ValuesSimdViewT::getSizeY(),
typename PropertyListT = ext::oneapi::experimental::empty_properties_t>
__ESIMD_API std::enable_if_t<
detail::is_simd_view_type_v<OffsetSimdViewT> &&
detail::is_simd_view_type_v<ValuesSimdViewT> &&
ext::oneapi::experimental::is_property_list_v<PropertyListT>>
scatter(T *p, OffsetSimdViewT byte_offsets, ValuesSimdViewT vals,
PropertyListT props = {}) {
simd_mask<N / VS> Mask = 1;
scatter<T, N, VS>(p, byte_offsets.read(), vals.read(), Mask, props);
}

/// A variation of \c scatter API with \c offsets represented as scalar.
///
/// @tparam Tx Element type, must be of size 4 or less.
Expand Down
37 changes: 35 additions & 2 deletions sycl/test/esimd/memory_properties_scatter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,7 +96,7 @@ test_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,

scatter(ptrf, ioffset_n32, usm, props_align4);

// CHECK-COUNT-14: call void @llvm.genx.lsc.store.stateless.v32i1.v32i64.v32i32(<32 x i1> {{[^)]+}}, i8 4, i8 1, i8 1, i16 1, i32 0, i8 3, i8 1, i8 1, i8 0, <32 x i64> {{[^)]+}}, <32 x i32> {{[^)]+}}, i32 0)
// CHECK-COUNT-22: call void @llvm.genx.lsc.store.stateless.v32i1.v32i64.v32i32(<32 x i1> {{[^)]+}}, i8 4, i8 1, i8 1, i16 1, i32 0, i8 3, i8 1, i8 1, i8 0, <32 x i64> {{[^)]+}}, <32 x i32> {{[^)]+}}, i32 0)
scatter(ptrf, ioffset_n32, usm, mask_n32, props_cache_load);
scatter(ptrf, ioffset_n32, usm, props_cache_load);

Expand All @@ -110,6 +110,12 @@ test_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,
props_cache_load);
scatter<float, 32>(ptrf, ioffset_n32_view, usm_view, props_cache_load);

scatter(ptrf, ioffset_n32, usm_view, mask_n32, props_cache_load);
scatter(ptrf, ioffset_n32, usm_view, props_cache_load);

scatter(ptrf, ioffset_n32_view, usm_view, mask_n32, props_cache_load);
scatter(ptrf, ioffset_n32_view, usm_view, props_cache_load);

scatter(ptrf, ioffset_n32_view.select<32, 1>(), usm, mask_n32,
props_cache_load);
scatter(ptrf, ioffset_n32_view.select<32, 1>(), usm, props_cache_load);
Expand All @@ -123,9 +129,17 @@ test_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,
usm_view.select<32, 1>(), mask_n32, props_cache_load);
scatter<float, 32>(ptrf, ioffset_n32_view.select<32, 1>(),
usm_view.select<32, 1>(), props_cache_load);
scatter(ptrf, ioffset_n32, usm_view.select<32, 1>(), mask_n32,
props_cache_load);
scatter(ptrf, ioffset_n32, usm_view.select<32, 1>(), props_cache_load);

scatter(ptrf, ioffset_n32_view.select<32, 1>(), usm_view.select<32, 1>(),
mask_n32, props_cache_load);
scatter(ptrf, ioffset_n32_view.select<32, 1>(), usm_view.select<32, 1>(),
props_cache_load);

// VS > 1
// CHECK-COUNT-14: call void @llvm.genx.lsc.store.stateless.v16i1.v16i64.v32i32(<16 x i1> {{[^)]+}}, i8 4, i8 1, i8 1, i16 1, i32 0, i8 3, i8 2, i8 1, i8 0, <16 x i64> {{[^)]+}}, <32 x i32> {{[^)]+}}, i32 0)
// CHECK-COUNT-24: call void @llvm.genx.lsc.store.stateless.v16i1.v16i64.v32i32(<16 x i1> {{[^)]+}}, i8 4, i8 1, i8 1, i16 1, i32 0, i8 3, i8 2, i8 1, i8 0, <16 x i64> {{[^)]+}}, <32 x i32> {{[^)]+}}, i32 0)
scatter<float, 32, 2>(ptrf, ioffset_n16, usm, mask_n16, props_cache_load);

scatter<float, 32, 2>(ptrf, ioffset_n16, usm, props_cache_load);
Expand All @@ -147,6 +161,16 @@ test_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,
scatter<float, 32, 2>(ptrf, ioffset_n16_view.select<16, 1>(), usm,
props_cache_load);

scatter<2>(ptrf, ioffset_n16, usm_view, mask_n16, props_cache_load);
scatter<2>(ptrf, ioffset_n16, usm_view, props_cache_load);

scatter<2>(ptrf, ioffset_n16_view, usm_view, mask_n16, props_cache_load);
scatter<2>(ptrf, ioffset_n16_view, usm_view, props_cache_load);

scatter<2>(ptrf, ioffset_n16_view.select<16, 1>(), usm, mask_n16,
props_cache_load);
scatter<2>(ptrf, ioffset_n16_view.select<16, 1>(), usm, props_cache_load);

scatter<float, 32, 2>(ptrf, ioffset_n16, usm_view.select<32, 1>(), mask_n16,
props_cache_load);
scatter<float, 32, 2>(ptrf, ioffset_n16, usm_view.select<32, 1>(),
Expand All @@ -157,6 +181,15 @@ test_scatter(AccType &acc, LocalAccType &local_acc, float *ptrf,
scatter<float, 32, 2>(ptrf, ioffset_n16_view.select<16, 1>(),
usm_view.select<32, 1>(), props_cache_load);

scatter<2>(ptrf, ioffset_n16, usm_view.select<32, 1>(), mask_n16,
props_cache_load);
scatter<2>(ptrf, ioffset_n16, usm_view.select<32, 1>(), props_cache_load);

scatter<2>(ptrf, ioffset_n16_view.select<16, 1>(), usm_view.select<32, 1>(),
mask_n16, props_cache_load);
scatter<2>(ptrf, ioffset_n16_view.select<16, 1>(), usm_view.select<32, 1>(),
props_cache_load);

// CHECK-COUNT-14: call void @llvm.genx.lsc.store.stateless.v16i1.v16i64.v32i32(<16 x i1> {{[^)]+}}, i8 4, i8 0, i8 0, i16 1, i32 0, i8 3, i8 2, i8 1, i8 0, <16 x i64> {{[^)]+}}, <32 x i32> {{[^)]+}}, i32 0)
scatter<float, 32, 2>(ptrf, ioffset_n16, usm, mask_n16);

Expand Down