From b3e9465c30c92977f76ad956334f7d1e7b9352f2 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 12:34:34 -0500
Subject: [PATCH 1/6] Implementations of reductions for contigous case must
take offsets into account
---
.../libtensor/include/kernels/reductions.hpp | 22 +++++++++++--------
1 file changed, 13 insertions(+), 9 deletions(-)
diff --git a/dpctl/tensor/libtensor/include/kernels/reductions.hpp b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
index b9e2918c8c..d5fddce6ed 100644
--- a/dpctl/tensor/libtensor/include/kernels/reductions.hpp
+++ b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
@@ -1339,7 +1339,7 @@ sycl::event reduction_over_group_temps_strided_impl(
static_cast<py::ssize_t>(remaining_reduction_nelems)};
ResIndexerT res_iter_indexer{iter_nd, iter_res_offset,
/* shape */ iter_shape_and_strides,
- /*s trides */ iter_shape_and_strides +
+ /* strides */ iter_shape_and_strides +
2 * iter_nd};
InputOutputIterIndexerT in_out_iter_indexer{inp_indexer,
@@ -1424,8 +1424,9 @@ sycl::event reduction_axis1_over_group_temps_contig_impl(
py::ssize_t reduction_arg_offset,
const std::vector<sycl::event> &depends)
{
- const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp);
- resTy *res_tp = reinterpret_cast<resTy *>(res_cp);
+ const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp) +
+ iter_arg_offset + reduction_arg_offset;
+ resTy *res_tp = reinterpret_cast<resTy *>(res_cp) + iter_res_offset;
constexpr resTy identity_val = su_ns::Identity<ReductionOpT, resTy>::value;
@@ -1767,8 +1768,9 @@ sycl::event reduction_axis0_over_group_temps_contig_impl(
py::ssize_t reduction_arg_offset,
const std::vector<sycl::event> &depends)
{
- const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp);
- resTy *res_tp = reinterpret_cast<resTy *>(res_cp);
+ const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp) +
+ iter_arg_offset + reduction_arg_offset;
+ resTy *res_tp = reinterpret_cast<resTy *>(res_cp) + iter_res_offset;
constexpr resTy identity_val = su_ns::Identity<ReductionOpT, resTy>::value;
@@ -4258,8 +4260,9 @@ sycl::event search_axis1_over_group_temps_contig_impl(
py::ssize_t reduction_arg_offset,
const std::vector<sycl::event> &depends)
{
- const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp);
- resTy *res_tp = reinterpret_cast<resTy *>(res_cp);
+ const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp) +
+ iter_arg_offset + reduction_arg_offset;
+ resTy *res_tp = reinterpret_cast<resTy *>(res_cp) + iter_res_offset;
constexpr argTy identity_val = su_ns::Identity<ReductionOpT, argTy>::value;
constexpr resTy idx_identity_val = su_ns::Identity<IndexOpT, resTy>::value;
@@ -4635,8 +4638,9 @@ sycl::event search_axis0_over_group_temps_contig_impl(
py::ssize_t reduction_arg_offset,
const std::vector<sycl::event> &depends)
{
- const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp);
- resTy *res_tp = reinterpret_cast<resTy *>(res_cp);
+ const argTy *arg_tp = reinterpret_cast<const argTy *>(arg_cp) +
+ iter_arg_offset + reduction_arg_offset;
+ resTy *res_tp = reinterpret_cast<resTy *>(res_cp) + iter_res_offset;
constexpr argTy identity_val = su_ns::Identity<ReductionOpT, argTy>::value;
constexpr resTy idx_identity_val = su_ns::Identity<IndexOpT, resTy>::value;
From c63c5451ec92620956d106df8f68ec5d8bc58680 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 12:35:11 -0500
Subject: [PATCH 2/6] Expand test to cover non-contig. input that can be
simplified into one
---
dpctl/tests/test_tensor_sum.py | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/dpctl/tests/test_tensor_sum.py b/dpctl/tests/test_tensor_sum.py
index a4e202f073..fbfd9547e1 100644
--- a/dpctl/tests/test_tensor_sum.py
+++ b/dpctl/tests/test_tensor_sum.py
@@ -43,6 +43,7 @@ def test_sum_arg_dtype_default_output_dtype_matrix(arg_dtype):
q = get_queue_or_skip()
skip_if_dtype_not_supported(arg_dtype, q)
+ # test reduction for C-contiguous input
m = dpt.ones(100, dtype=arg_dtype)
r = dpt.sum(m)
@@ -55,12 +56,20 @@ def test_sum_arg_dtype_default_output_dtype_matrix(arg_dtype):
assert r.dtype.kind == "f"
elif m.dtype.kind == "c":
assert r.dtype.kind == "c"
+
assert dpt.all(r == 100)
+ # test reduction for strided input
m = dpt.ones(200, dtype=arg_dtype)[:1:-2]
r = dpt.sum(m)
assert dpt.all(r == 99)
+ # test reduction for strided input which can be simplified
+ # to contiguous computation
+ m = dpt.ones(100, dtype=arg_dtype)
+ r = dpt.sum(dpt.flip(m))
+ assert dpt.all(r == 100)
+
@pytest.mark.parametrize("arg_dtype", _all_dtypes)
@pytest.mark.parametrize("out_dtype", _all_dtypes[1:])
From e92d1f9ad3a138c9c85e181c0e31e7293a4b8eb2 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 12:43:05 -0500
Subject: [PATCH 3/6] Add tests for strided input where contig implementation
is applicable
---
dpctl/tests/test_usm_ndarray_reductions.py | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/dpctl/tests/test_usm_ndarray_reductions.py b/dpctl/tests/test_usm_ndarray_reductions.py
index 73cf9459a7..d500ce26b6 100644
--- a/dpctl/tests/test_usm_ndarray_reductions.py
+++ b/dpctl/tests/test_usm_ndarray_reductions.py
@@ -169,6 +169,14 @@ def test_search_reduction_kernels(arg_dtype):
m = dpt.argmax(x)
assert m == idx
+ m = dpt.argmax(dpt.flip(x))
+ assert m == x.size - 1 - idx
+
+ y = dpt.ones(2 * x.size, dtype=arg_dtype, sycl_queue=q)
+ y[::2] = x
+ m = dpt.argmax(y)
+ assert m == 2 * idx
+
x = dpt.reshape(x, (24, 1025))
x[idx_tup[0], :] = 3
From 702b707250d9ab226704d6ce3fee7e2307d0fdbb Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 10:59:54 -0700
Subject: [PATCH 4/6] Added comments to the test file
---
dpctl/tests/test_usm_ndarray_reductions.py | 3 +++
1 file changed, 3 insertions(+)
diff --git a/dpctl/tests/test_usm_ndarray_reductions.py b/dpctl/tests/test_usm_ndarray_reductions.py
index d500ce26b6..56059e54b8 100644
--- a/dpctl/tests/test_usm_ndarray_reductions.py
+++ b/dpctl/tests/test_usm_ndarray_reductions.py
@@ -169,9 +169,12 @@ def test_search_reduction_kernels(arg_dtype):
m = dpt.argmax(x)
assert m == idx
+ # test case of strided input mapping to contig
+ # implementation
m = dpt.argmax(dpt.flip(x))
assert m == x.size - 1 - idx
+ # test case of strided implementation
y = dpt.ones(2 * x.size, dtype=arg_dtype, sycl_queue=q)
y[::2] = x
m = dpt.argmax(y)
From dcb566a424a902af8ccb7e96021a3899e98c925b Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 15:24:17 -0500
Subject: [PATCH 5/6] Corrected logical error in can_use_reduce_over_group
trait implementation
---
dpctl/tensor/libtensor/include/kernels/reductions.hpp | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)
diff --git a/dpctl/tensor/libtensor/include/kernels/reductions.hpp b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
index d5fddce6ed..884a7c5461 100644
--- a/dpctl/tensor/libtensor/include/kernels/reductions.hpp
+++ b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
@@ -50,12 +50,18 @@ namespace tensor
namespace kernels
{
+template <typename ReductionOpT, typename T> struct needs_workaround
+{
+ static constexpr bool value =
+ std::is_same_v<ReductionOpT, sycl::multiplies<T>> &&
+ (std::is_same_v<T, std::int64_t> || std::is_same_v<T, std::uint64_t>);
+};
+
template <typename ReductionOpT, typename T> struct can_use_reduce_over_group
{
static constexpr bool value =
sycl::has_known_identity<ReductionOpT, T>::value &&
- !std::is_same_v<T, std::int64_t> && !std::is_same_v<T, std::uint64_t> &&
- !std::is_same_v<ReductionOpT, sycl::multiplies<T>>;
+ !needs_workaround<ReductionOpT, T>::value;
};
template <typename argT,
From bfba152db0ee15e2caaeac9c2dda6c74da628645 Mon Sep 17 00:00:00 2001
From: Oleksandr Pavlyk <oleksandr.pavlyk@intel.com>
Date: Thu, 26 Oct 2023 17:42:26 -0500
Subject: [PATCH 6/6] The taper optimization in tree-reduction which causes
problem with CUDA
The optimization should not use max-work-group-size, to allow RT some
of the SLM memory.
---
.../libtensor/include/kernels/reductions.hpp | 21 +++++++++++--------
1 file changed, 12 insertions(+), 9 deletions(-)
diff --git a/dpctl/tensor/libtensor/include/kernels/reductions.hpp b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
index 884a7c5461..40e5bd282d 100644
--- a/dpctl/tensor/libtensor/include/kernels/reductions.hpp
+++ b/dpctl/tensor/libtensor/include/kernels/reductions.hpp
@@ -1094,7 +1094,7 @@ sycl::event reduction_over_group_temps_strided_impl(
// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
- max_max_wg, d.get_info<sycl::info::device::max_work_group_size>());
+ max_max_wg, d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {
@@ -1444,7 +1444,7 @@ sycl::event reduction_axis1_over_group_temps_contig_impl(
// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
- max_max_wg, d.get_info<sycl::info::device::max_work_group_size>());
+ max_max_wg, d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {
@@ -1788,7 +1788,7 @@ sycl::event reduction_axis0_over_group_temps_contig_impl(
// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
- max_max_wg, d.get_info<sycl::info::device::max_work_group_size>());
+ max_max_wg, d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {
@@ -3883,8 +3883,9 @@ sycl::event search_over_group_temps_strided_impl(
constexpr size_t preferrered_reductions_per_wi = 4;
// max_max_wg prevents running out of resources on CPU
- size_t max_wg = std::min(
- size_t(2048), d.get_info<sycl::info::device::max_work_group_size>());
+ size_t max_wg =
+ std::min(size_t(2048),
+ d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {
@@ -4279,8 +4280,9 @@ sycl::event search_axis1_over_group_temps_contig_impl(
constexpr size_t preferrered_reductions_per_wi = 8;
// max_max_wg prevents running out of resources on CPU
- size_t max_wg = std::min(
- size_t(2048), d.get_info<sycl::info::device::max_work_group_size>());
+ size_t max_wg =
+ std::min(size_t(2048),
+ d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {
@@ -4657,8 +4659,9 @@ sycl::event search_axis0_over_group_temps_contig_impl(
constexpr size_t preferrered_reductions_per_wi = 8;
// max_max_wg prevents running out of resources on CPU
- size_t max_wg = std::min(
- size_t(2048), d.get_info<sycl::info::device::max_work_group_size>());
+ size_t max_wg =
+ std::min(size_t(2048),
+ d.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t reductions_per_wi(preferrered_reductions_per_wi);
if (reduction_nelems <= preferrered_reductions_per_wi * max_wg) {