[REVIEW] L1 distance support for iterative search CAGRA build

cpp/CMakeLists.txt

@@ -249,6 +249,18 @@ if(NOT BUILD_CPU_ONLY)
     src/neighbors/detail/cagra/compute_distance_standard_InnerProduct_uint8_uint32_dim128_t8.cu
     src/neighbors/detail/cagra/compute_distance_standard_InnerProduct_uint8_uint32_dim256_t16.cu
     src/neighbors/detail/cagra/compute_distance_standard_InnerProduct_uint8_uint32_dim512_t32.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim128_t8.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim256_t16.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim512_t32.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim128_t8.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim256_t16.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim512_t32.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim128_t8.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim256_t16.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim512_t32.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_uint8_uint32_dim128_t8.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_uint8_uint32_dim256_t16.cu
+    src/neighbors/detail/cagra/compute_distance_standard_L1_uint8_uint32_dim512_t32.cu
     src/neighbors/detail/cagra/compute_distance_standard_L2Expanded_float_uint32_dim128_t8.cu
     src/neighbors/detail/cagra/compute_distance_standard_L2Expanded_float_uint32_dim256_t16.cu
     src/neighbors/detail/cagra/compute_distance_standard_L2Expanded_float_uint32_dim512_t32.cu

cpp/src/neighbors/detail/cagra/compute_distance.cu

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -22,61 +22,73 @@ template struct instance_selector<
   standard_descriptor_spec<DistanceType::L2Expanded, 8, 128, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 8, 128, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 8, 128, float, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 8, 128, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 2, half, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 4, half, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 16, 256, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 16, 256, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 16, 256, float, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 16, 256, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 2, half, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 4, half, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 32, 512, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 32, 512, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 32, 512, float, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 32, 512, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 2, half, float, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 4, half, float, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 8, 128, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 8, 128, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 8, 128, half, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 8, 128, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 2, half, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 4, half, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 16, 256, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 16, 256, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 16, 256, half, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 16, 256, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 2, half, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 4, half, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 32, 512, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 32, 512, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 32, 512, half, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 32, 512, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 2, half, half, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 4, half, half, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 8, 128, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 8, 128, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 8, 128, int8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 8, 128, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 2, half, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 4, half, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 16, 256, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 16, 256, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 16, 256, int8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 16, 256, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 2, half, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 4, half, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 32, 512, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 32, 512, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 32, 512, int8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 32, 512, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 2, half, int8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 4, half, int8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 8, 128, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 8, 128, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 8, 128, uint8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 8, 128, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 2, half, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 8, 128, 8, 4, half, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 16, 256, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 16, 256, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 16, 256, uint8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 16, 256, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 2, half, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 16, 256, 8, 4, half, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::L2Expanded, 32, 512, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::InnerProduct, 32, 512, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::CosineExpanded, 32, 512, uint8_t, uint32_t, float>,
+  standard_descriptor_spec<DistanceType::L1, 32, 512, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 2, half, uint8_t, uint32_t, float>,
   vpq_descriptor_spec<DistanceType::L2Expanded, 32, 512, 8, 4, half, uint8_t, uint32_t, float>,
   standard_descriptor_spec<DistanceType::BitwiseHamming, 8, 128, uint8_t, uint32_t, float>,

cpp/src/neighbors/detail/cagra/compute_distance_00_generate.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 import datetime
 import os
@@ -19,14 +19,14 @@
  *
  */
 
-{{includes}}
+{includes}
 
-namespace cuvs::neighbors::cagra::detail {{{{
+namespace cuvs::neighbors::cagra::detail {{
 
 using namespace cuvs::distance;
-{{content}}
+{content}
 
-}}}}  // namespace cuvs::neighbors::cagra::detail
+}}  // namespace cuvs::neighbors::cagra::detail
 """
 
 mxdim_team = [(128, 8), (256, 16), (512, 32)]
@@ -65,7 +65,7 @@
 for type_path, (data_t, idx_t, distance_t) in search_types.items():
     for mxdim, team in mxdim_team:
         # CAGRA
-        for metric in ["L2Expanded", "InnerProduct", "CosineExpanded"]:
+        for metric in ["L2Expanded", "InnerProduct", "CosineExpanded", "L1"]:
             path = f"compute_distance_standard_{metric}_{type_path}_dim{mxdim}_t{team}.cu"
             includes = '#include "compute_distance_standard-impl.cuh"'
             params = f"{metric_prefix}{metric}, {team}, {mxdim}, {data_t}, {idx_t}, {distance_t}"

cpp/src/neighbors/detail/cagra/compute_distance_standard-impl.cuh

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
  * SPDX-License-Identifier: Apache-2.0
  */
 #pragma once
@@ -15,32 +15,37 @@
 namespace cuvs::neighbors::cagra::detail {
 namespace {
 template <typename DATA_T, typename DISTANCE_T, cuvs::distance::DistanceType Metric>
+  requires(Metric == cuvs::distance::DistanceType::L2Expanded)
 RAFT_DEVICE_INLINE_FUNCTION constexpr auto dist_op(DATA_T a, DATA_T b)
-  -> std::enable_if_t<Metric == cuvs::distance::DistanceType::L2Expanded, DISTANCE_T>
 {
   DISTANCE_T diff = a - b;
   return diff * diff;
 }
 
 template <typename DATA_T, typename DISTANCE_T, cuvs::distance::DistanceType Metric>
+  requires(Metric == cuvs::distance::DistanceType::InnerProduct ||
+           Metric == cuvs::distance::DistanceType::CosineExpanded)
 RAFT_DEVICE_INLINE_FUNCTION constexpr auto dist_op(DATA_T a, DATA_T b)
-  -> std::enable_if_t<Metric == cuvs::distance::DistanceType::InnerProduct ||
-                        Metric == cuvs::distance::DistanceType::CosineExpanded,
-                      DISTANCE_T>
 {
   return -static_cast<DISTANCE_T>(a) * static_cast<DISTANCE_T>(b);
 }
 
 template <typename DATA_T, typename DISTANCE_T, cuvs::distance::DistanceType Metric>
+  requires(Metric == cuvs::distance::DistanceType::BitwiseHamming && std::is_integral_v<DATA_T>)
 RAFT_DEVICE_INLINE_FUNCTION constexpr auto dist_op(DATA_T a, DATA_T b)
-  -> std::enable_if_t<Metric == cuvs::distance::DistanceType::BitwiseHamming &&
-                        std::is_integral_v<DATA_T>,
-                      DISTANCE_T>
 {
   // mask the result of xor for the integer promotion
   const auto v = (a ^ b) & 0xffu;
   return __popc(v);
 }
+
+template <typename DATA_T, typename DISTANCE_T, cuvs::distance::DistanceType Metric>
+  requires(Metric == cuvs::distance::DistanceType::L1)
+RAFT_DEVICE_INLINE_FUNCTION constexpr auto dist_op(DATA_T a, DATA_T b)
+{
+  DISTANCE_T diff = a - b;
+  return raft::abs(diff);
+}
 }  // namespace
 
 template <cuvs::distance::DistanceType Metric,

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim128_t8.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 8, 128, float, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim256_t16.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 16, 256, float, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_float_uint32_dim512_t32.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 32, 512, float, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim128_t8.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 8, 128, half, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim256_t16.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 16, 256, half, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_half_uint32_dim512_t32.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 32, 512, half, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim128_t8.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 8, 128, int8_t, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim256_t16.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 16, 256, int8_t, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail

cpp/src/neighbors/detail/cagra/compute_distance_standard_L1_int8_uint32_dim512_t32.cu

@@ -0,0 +1,22 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+/*
+ * NOTE: this file is generated by compute_distance_00_generate.py
+ *
+ * Make changes there and run in this directory:
+ *
+ * > python compute_distance_00_generate.py
+ *
+ */
+
+#include "compute_distance_standard-impl.cuh"
+
+namespace cuvs::neighbors::cagra::detail {
+
+using namespace cuvs::distance;
+template struct standard_descriptor_spec<DistanceType::L1, 32, 512, int8_t, uint32_t, float>;
+
+}  // namespace cuvs::neighbors::cagra::detail