cuda-cccl 0.3.3__cp313-cp313-win_amd64.whl

cuda/cccl/headers/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh

@@ -0,0 +1,1010 @@
+/******************************************************************************
+ * Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in the
+ *       documentation and/or other materials provided with the distribution.
+ *     * Neither the name of the NVIDIA CORPORATION nor the
+ *       names of its contributors may be used to endorse or promote products
+ *       derived from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY
+ * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ ******************************************************************************/
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/agent/agent_scan_by_key.cuh>
+#include <cub/agent/single_pass_scan_operators.cuh>
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_scan.cuh>
+#include <cub/block/block_store.cuh>
+#include <cub/thread/thread_operators.cuh>
+#include <cub/util_device.cuh>
+#include <cub/util_math.cuh>
+#include <cub/util_type.cuh>
+
+CUB_NAMESPACE_BEGIN
+
+namespace detail::scan_by_key
+{
+enum class primitive_accum
+{
+  no,
+  yes
+};
+enum class primitive_op
+{
+  no,
+  yes
+};
+enum class val_size
+{
+  _1,
+  _2,
+  _4,
+  _8,
+  _16,
+  unknown
+};
+enum class key_size
+{
+  _1,
+  _2,
+  _4,
+  _8,
+  _16,
+  unknown
+};
+
+template <class AccumT>
+constexpr primitive_accum is_primitive_accum()
+{
+  return is_primitive<AccumT>::value ? primitive_accum::yes : primitive_accum::no;
+}
+
+template <class ScanOpT>
+constexpr primitive_op is_primitive_op()
+{
+  return basic_binary_op_t<ScanOpT>::value ? primitive_op::yes : primitive_op::no;
+}
+
+template <class ValueT>
+constexpr val_size classify_val_size()
+{
+  return sizeof(ValueT) == 1 ? val_size::_1
+       : sizeof(ValueT) == 2 ? val_size::_2
+       : sizeof(ValueT) == 4 ? val_size::_4
+       : sizeof(ValueT) == 8 ? val_size::_8
+       : sizeof(ValueT) == 16
+         ? val_size::_16
+         : val_size::unknown;
+}
+
+template <class KeyT>
+constexpr key_size classify_key_size()
+{
+  return sizeof(KeyT) == 1 ? key_size::_1
+       : sizeof(KeyT) == 2 ? key_size::_2
+       : sizeof(KeyT) == 4 ? key_size::_4
+       : sizeof(KeyT) == 8 ? key_size::_8
+       : sizeof(KeyT) == 16
+         ? key_size::_16
+         : key_size::unknown;
+}
+
+template <class KeyT,
+          class ValueT,
+          primitive_op PrimitiveOp,
+          key_size KeySize                     = classify_key_size<KeyT>(),
+          val_size ValueSize                   = classify_val_size<ValueT>(),
+          primitive_accum PrimitiveAccumulator = is_primitive_accum<ValueT>()>
+struct sm80_tuning;
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = no_delay_constructor_t<795>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 288;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<825>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<640>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<124, 1040>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 19;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1095>;
+};
+
+template <class KeyT>
+struct sm80_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+    : sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 8;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = no_delay_constructor_t<1070>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 320;
+  static constexpr int items                           = 14;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<625>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1055>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 160;
+  static constexpr int items                           = 17;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<160, 695>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 160;
+  static constexpr int items                           = 14;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1105>;
+};
+
+template <class KeyT>
+struct sm80_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+    : sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = no_delay_constructor_t<1130>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1130>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1140>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 9;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<888, 635>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 17;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1100>;
+};
+
+template <class KeyT>
+struct sm80_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+    : sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 11;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1120>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1115>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 13;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<24, 1060>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1160>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 320;
+  static constexpr int items                           = 8;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<220>;
+};
+
+template <class KeyT>
+struct sm80_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+    : sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 7;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<144, 1120>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 7;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<364, 780>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 7;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1170>;
+};
+
+template <class KeyT, class ValueT>
+struct sm80_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1030>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1160>;
+};
+
+template <class KeyT>
+struct sm80_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+    : sm80_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT,
+          class ValueT,
+          primitive_op PrimitiveOp,
+          key_size KeySize                     = classify_key_size<KeyT>(),
+          val_size ValueSize                   = classify_val_size<ValueT>(),
+          primitive_accum PrimitiveAccumulator = is_primitive_accum<ValueT>()>
+struct sm90_tuning;
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = no_delay_constructor_t<650>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 16;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<124, 995>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<488, 545>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<488, 1070>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 23;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<936, 1105>;
+};
+
+template <class KeyT>
+struct sm90_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+    : sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_1, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = fixed_delay_constructor_t<136, 785>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 20;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<445>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 22;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<312, 865>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<352, 1170>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 23;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<504, 1190>;
+};
+
+template <class KeyT>
+struct sm90_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+    : sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_2, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = no_delay_constructor_t<850>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 14;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<128, 965>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 288;
+  static constexpr int items                           = 14;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<700, 1005>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 14;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<556, 1195>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 23;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<512, 1030>;
+};
+
+template <class KeyT>
+struct sm90_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+    : sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_4, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 12;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_DIRECT;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_DIRECT;
+  using delay_constructor                              = fixed_delay_constructor_t<504, 1010>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<420, 970>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<500, 1125>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 11;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<600, 930>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 15;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<364, 1085>;
+};
+
+template <class KeyT>
+struct sm90_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+    : sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_8, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_1, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 7;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<500, 975>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_2, primitive_accum::yes>
+{
+  static constexpr int threads                         = 224;
+  static constexpr int items                           = 10;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<164, 1075>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_4, primitive_accum::yes>
+{
+  static constexpr int threads                         = 256;
+  static constexpr int items                           = 9;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<268, 1120>;
+};
+
+template <class KeyT, class ValueT>
+struct sm90_tuning<KeyT, ValueT, primitive_op::yes, key_size::_16, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int threads                         = 192;
+  static constexpr int items                           = 9;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<320, 1200>;
+};
+
+#if _CCCL_HAS_INT128()
+template <class KeyT>
+struct sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+{
+  static constexpr int threads                         = 128;
+  static constexpr int items                           = 23;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<364, 1050>;
+};
+
+template <class KeyT>
+struct sm90_tuning<KeyT, __uint128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+    : sm90_tuning<KeyT, __int128_t, primitive_op::yes, key_size::_16, val_size::_16, primitive_accum::no>
+{};
+#endif
+
+template <class KeyT,
+          class ValueT,
+          primitive_op PrimitiveOp,
+          key_size KeySize                     = classify_key_size<KeyT>(),
+          val_size ValueSize                   = classify_val_size<ValueT>(),
+          primitive_accum PrimitiveAccumulator = is_primitive_accum<ValueT>()>
+struct sm100_tuning;
+
+// key_size = 8 bits
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_1, primitive_accum::yes>
+{
+  // ipt_13.tpb_288.ns_420.dcid_0.l2w_745.trp_1.ld_0 1.030222  0.998162  1.027506  1.068348
+  static constexpr int items                           = 13;
+  static constexpr int threads                         = 288;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<745>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_2, primitive_accum::yes>
+{
+  // ipt_13.tpb_288.ns_388.dcid_1.l2w_570.trp_1.ld_0 1.228612   1.0  1.216841  1.416167
+  static constexpr int items                           = 13;
+  static constexpr int threads                         = 288;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<388, 570>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_4, primitive_accum::yes>
+{
+  // ipt_19.tpb_224.ns_1028.dcid_5.l2w_910.trp_1.ld_1 1.163440   1.0  1.146400  1.260684
+  static constexpr int items                           = 19;
+  static constexpr int threads                         = 224;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_jitter_window_constructor_t<1028, 910>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_1, val_size::_8, primitive_accum::yes>
+{
+  // ipt_18.tpb_192.ns_432.dcid_1.l2w_1035.trp_1.ld_1 1.177638  0.985417  1.157164  1.296477
+  static constexpr int items                           = 18;
+  static constexpr int threads                         = 192;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<432, 1035>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+// key_size = 16 bits
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_1, primitive_accum::yes>
+{
+  // ipt_12.tpb_384.ns_1900.dcid_0.l2w_840.trp_1.ld_0 1.010828  0.985782  1.007993  1.048859
+  static constexpr int items                           = 12;
+  static constexpr int threads                         = 384;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1900>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_2, primitive_accum::yes>
+{
+  // ipt_14.tpb_160.ns_1736.dcid_7.l2w_170.trp_1.ld_0 1.095207  1.065061  1.100302  1.142857
+  static constexpr int items                           = 14;
+  static constexpr int threads                         = 160;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_constructor_t<1736, 170>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_4, primitive_accum::yes>
+{
+  // ipt_14.tpb_160.ns_336.dcid_1.l2w_805.trp_1.ld_0 1.119313  1.095238  1.122013  1.148681
+  static constexpr int items                           = 14;
+  static constexpr int threads                         = 160;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = fixed_delay_constructor_t<336, 805>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_2, val_size::_8, primitive_accum::yes>
+{
+  static constexpr int items                           = 13;
+  static constexpr int threads                         = 224;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backoff_constructor_t<348, 735>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+// key_size = 32 bits
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_1, primitive_accum::yes>
+{
+  // todo(gonidlelis): Significant regression. Search more workloads.
+  // ipt_20.tpb_224.ns_1436.dcid_7.l2w_155.trp_1.ld_1 1.135878  0.866667  1.106600  1.339708
+  static constexpr int items                           = 20;
+  static constexpr int threads                         = 224;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_constructor_t<1436, 155>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_2, primitive_accum::yes>
+{
+  // ipt_13.tpb_288.ns_620.dcid_7.l2w_925.trp_1.ld_2 1.050929  1.000000  1.047178  1.115809
+  static constexpr int items                           = 13;
+  static constexpr int threads                         = 288;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_constructor_t<620, 925>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_4, primitive_accum::yes>
+{
+  // ipt_20.tpb_224.ns_1856.dcid_5.l2w_280.trp_1.ld_1 1.247248  1.000000  1.220196  1.446328
+  static constexpr int items                           = 20;
+  static constexpr int threads                         = 224;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_jitter_window_constructor_t<1856, 280>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_4, val_size::_8, primitive_accum::yes>
+{
+  // ipt_14.tpb_224.ns_464.dcid_2.l2w_680.trp_1.ld_1 1.070831  1.002088  1.064736  1.105437
+  static constexpr int items                           = 14;
+  static constexpr int threads                         = 224;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backoff_constructor_t<464, 860>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+// key_size = 64 bits
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_1, primitive_accum::yes>
+{
+  // ipt_12.tpb_160.ns_532.dcid_0.l2w_850.trp_1.ld_0 1.041966  1.000000  1.037010  1.078399
+  static constexpr int items                           = 12;
+  static constexpr int threads                         = 160;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<532>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_2, primitive_accum::yes>
+{
+  // todo(gonidlelis): Significant regression. Search more workloads.
+  // ipt_15.tpb_288.ns_988.dcid_7.l2w_335.trp_1.ld_0 1.064413  0.866667  1.045946  1.116803
+  static constexpr int items                           = 15;
+  static constexpr int threads                         = 288;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_constructor_t<988, 335>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_4, primitive_accum::yes>
+{
+  // ipt_22.tpb_160.ns_1032.dcid_5.l2w_505.trp_1.ld_2 1.184805  1.000000  1.164843  1.338536
+  static constexpr int items                           = 22;
+  static constexpr int threads                         = 160;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = exponential_backon_jitter_window_constructor_t<1032, 505>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_CA;
+};
+
+template <class KeyT, class ValueT>
+struct sm100_tuning<KeyT, ValueT, primitive_op::yes, key_size::_8, val_size::_8, primitive_accum::yes>
+{
+  // ipt_23.tpb_256.ns_1232.dcid_0.l2w_810.trp_1.ld_0 1.067631  1.000000  1.059607  1.135646
+  static constexpr int items                           = 23;
+  static constexpr int threads                         = 256;
+  static constexpr BlockLoadAlgorithm load_algorithm   = BLOCK_LOAD_WARP_TRANSPOSE;
+  static constexpr BlockStoreAlgorithm store_algorithm = BLOCK_STORE_WARP_TRANSPOSE;
+  using delay_constructor                              = no_delay_constructor_t<1232>;
+  static constexpr CacheLoadModifier load_modifier     = LOAD_DEFAULT;
+};
+
+template <typename KeysInputIteratorT, typename AccumT, typename ValueT, typename ScanOpT>
+struct policy_hub
+{
+  using key_t                               = it_value_t<KeysInputIteratorT>;
+  static constexpr int max_input_bytes      = static_cast<int>((::cuda::std::max) (sizeof(key_t), sizeof(AccumT)));
+  static constexpr int combined_input_bytes = static_cast<int>(sizeof(key_t) + sizeof(AccumT));
+
+  struct Policy500 : ChainedPolicy<500, Policy500, Policy500>
+  {
+    static constexpr int nominal_4b_items_per_thread = 6;
+    static constexpr int items_per_thread =
+      max_input_bytes <= 8 ? 6 : Nominal4BItemsToItemsCombined(nominal_4b_items_per_thread, combined_input_bytes);
+
+    using ScanByKeyPolicyT =
+      AgentScanByKeyPolicy<128,
+                           items_per_thread,
+                           BLOCK_LOAD_WARP_TRANSPOSE,
+                           LOAD_CA,
+                           BLOCK_SCAN_WARP_SCANS,
+                           BLOCK_STORE_WARP_TRANSPOSE,
+                           default_reduce_by_key_delay_constructor_t<AccumT, int>>;
+  };
+
+  template <CacheLoadModifier LoadModifier, typename DelayConstructurValueT>
+  struct DefaultPolicy
+  {
+    static constexpr int nominal_4b_items_per_thread = 9;
+    static constexpr int items_per_thread =
+      max_input_bytes <= 8 ? 9 : Nominal4BItemsToItemsCombined(nominal_4b_items_per_thread, combined_input_bytes);
+
+    using ScanByKeyPolicyT =
+      AgentScanByKeyPolicy<256,
+                           items_per_thread,
+                           BLOCK_LOAD_WARP_TRANSPOSE,
+                           LoadModifier,
+                           BLOCK_SCAN_WARP_SCANS,
+                           BLOCK_STORE_WARP_TRANSPOSE,
+                           default_reduce_by_key_delay_constructor_t<DelayConstructurValueT, int>>;
+  };
+
+  struct Policy520
+      : DefaultPolicy<LOAD_CA, AccumT>
+      , ChainedPolicy<520, Policy520, Policy500>
+  {};
+
+  // Use values from tuning if a specialization exists, otherwise pick the default
+  template <typename Tuning>
+  static auto select_agent_policy(int)
+    -> AgentScanByKeyPolicy<Tuning::threads,
+                            Tuning::items,
+                            Tuning::load_algorithm,
+                            LOAD_DEFAULT,
+                            BLOCK_SCAN_WARP_SCANS,
+                            Tuning::store_algorithm,
+                            typename Tuning::delay_constructor>;
+
+  template <typename Tuning>
+  // FIXME(bgruber): should we rather use `AccumT` instead of `ValueT` like the other default policies?
+  static auto select_agent_policy(long) -> typename DefaultPolicy<LOAD_DEFAULT, ValueT>::ScanByKeyPolicyT;
+
+  struct Policy800 : ChainedPolicy<800, Policy800, Policy520>
+  {
+    using ScanByKeyPolicyT = decltype(select_agent_policy<sm80_tuning<key_t, ValueT, is_primitive_op<ScanOpT>()>>(0));
+  };
+
+  struct Policy860
+      : DefaultPolicy<LOAD_CA, AccumT>
+      , ChainedPolicy<860, Policy860, Policy800>
+  {};
+
+  struct Policy900 : ChainedPolicy<900, Policy900, Policy860>
+  {
+    using ScanByKeyPolicyT = decltype(select_agent_policy<sm90_tuning<key_t, ValueT, is_primitive_op<ScanOpT>()>>(0));
+  };
+
+  struct Policy1000 : ChainedPolicy<1000, Policy1000, Policy900>
+  {
+    // Use values from tuning if a specialization exists, otherwise pick Policy900
+    template <typename Tuning>
+    static auto select_agent_policy100(int)
+      -> AgentScanByKeyPolicy<Tuning::threads,
+                              Tuning::items,
+                              Tuning::load_algorithm,
+                              Tuning::load_modifier,
+                              BLOCK_SCAN_WARP_SCANS,
+                              Tuning::store_algorithm,
+                              typename Tuning::delay_constructor>;
+
+    template <typename Tuning>
+    // FIXME(bgruber): should we rather use `AccumT` instead of `ValueT` like the other default policies?
+    static auto select_agent_policy100(long) -> typename Policy900::ScanByKeyPolicyT;
+
+    using ScanByKeyPolicyT =
+      decltype(select_agent_policy100<sm100_tuning<key_t, ValueT, is_primitive_op<ScanOpT>()>>(0));
+  };
+
+  using MaxPolicy = Policy1000;
+};
+} // namespace detail::scan_by_key
+
+CUB_NAMESPACE_END