[FEA] Codepacking for IVF-flat

cpp/include/raft/neighbors/detail/ivf_flat_build.cuh

@@ -31,6 +31,7 @@
 #include <raft/neighbors/ivf_list_types.hpp>
 #include <raft/spatial/knn/detail/ann_utils.cuh>
 #include <raft/stats/histogram.cuh>
+#include <raft/util/fast_int_div.cuh>
 #include <raft/util/pow2_utils.cuh>
 
 #include <rmm/cuda_stream_view.hpp>
@@ -416,4 +417,133 @@ inline void fill_refinement_index(raft::resources const& handle,
                                          refinement_index->veclen());
   RAFT_CUDA_TRY(cudaPeekAtLastError());
 }
+
+
+
+// template <typename T, typename IdxT, uint32_t BlockSize>
+// __launch_bounds__(BlockSize) __global__ void unpack_list_data_kernel_float32(
+//   T* out_codes,
+//   T* in_list_data,
+//   uint32_t n_rows,
+//   uint32_t dim,
+//   uint32_t veclen)
+// {
+//   const IdxT i = IdxT(blockDim.x) * IdxT(blockIdx.x) + threadIdx.x;
+//   if (i >= n_rows * dim) { return; }
+
+//   auto col = i % kIndexGroupSize * veclen;
+//   auto row = i / (kIndexGroupSize * veclen);
+
+//   auto vec = 
+
+//   // The data is written in interleaved groups of `index::kGroupSize` vectors
+//   using interleaved_group = Pow2<kIndexGroupSize>;
+//   auto group_offset       = interleaved_group::roundDown(row);
+//   auto ingroup_id         = interleaved_group::mod(row) * veclen;
+
+//   // The value of 4 was chosen because for float_32 dtype, calculate_veclen returns 4
+//   auto within_group_offset = Pow2<4>::quot(col);
+
+//   // Interleave dimensions of the source vector while recording it.
+//   // NB: such `veclen` is selected, that `dim % veclen == 0`
+//   out_codes[] = in_list_data[i];
+// }
+
+// /**
+//  * Pack interleaved flat codes from an existing packed non-interleaved list by the given row offset.
+//  *
+//  * @param[out] codes flat codes, [n_rows, dim]
+//  * @param[in] list_data the packed ivf::list data.
+//  * @param[in] row_offset how many rows in the list to skip.
+//  * @param[in] stream
+//  */
+// template<typename T, typename IdxT>
+// inline void unpack_list_data_float32(
+//   raft::resources const& handle,
+//   device_matrix_view<T, uint32_t, row_major> codes,
+//   device_mdspan<T, typename list_spec<uint32_t, T, IdxT>::list_extents, row_major> list_data,
+//   uint32_t row_offset)
+// {
+//   auto stream = raft::resource::get_cuda_stream(handle);
+//   auto n_rows = codes.extent(0);
+//   if (n_rows == 0) { return; }
+
+//   auto dim = codes.extent(1);
+
+//   n_rows -= row_offset;
+//   constexpr uint32_t kBlockSize = 256;
+//   dim3 blocks(div_rounding_up_safe<uint32_t>(n_rows, kBlockSize), 1, 1);
+//   dim3 threads(kBlockSize, 1, 1);
+//   auto kernel = pack_list_data_kernel_float32<T, IdxT, kBlockSize>;
+//   kernel<<<blocks, threads, 0, stream>>>(codes.data_handle(),
+//   list_data.data_handle(),
+//   n_rows,
+//   dim,
+//   4);
+//   RAFT_CUDA_TRY(cudaPeekAtLastError());
+// }
+
+
+
+// template <typename T, typename IdxT, uint32_t BlockSize>
+// __launch_bounds__(BlockSize) __global__ void pack_list_data_kernel_float32(
+//   T* list_data,
+//   T* codes,
+//   uint32_t n_rows,
+//   uint32_t dim,
+//   uint32_t veclen)
+// {
+//   const IdxT i = IdxT(blockDim.x) * IdxT(blockIdx.x) + threadIdx.x;
+//   if (i >= n_rows * dim) { return; }
+
+//   auto col = i % dim;
+//   auto row = i / n_rows;
+
+//   // The data is written in interleaved groups of `index::kGroupSize` vectors
+//   using interleaved_group = Pow2<kIndexGroupSize>;
+//   auto group_offset       = interleaved_group::roundDown(row);
+//   auto ingroup_id         = interleaved_group::mod(row) * veclen;
+
+//   // The value of 4 was chosen because for float_32 dtype, calculate_veclen returns 4
+//   auto within_group_offset = Pow2<4>::quot(col);
+
+//   // Interleave dimensions of the source vector while recording it.
+//   // NB: such `veclen` is selected, that `dim % veclen == 0`
+//   list_data[group_offset * dim + within_group_offset * kIndexGroupSize * veclen + ingroup_id + col % veclen] = codes[i];
+// }
+
+// /**
+//  * Pack interleaved flat codes from an existing packed non-interleaved list by the given row offset.
+//  *
+//  * @param[out] codes flat codes, [n_rows, dim]
+//  * @param[in] list_data the packed ivf::list data.
+//  * @param[in] row_offset how many rows in the list to skip.
+//  * @param[in] stream
+//  */
+// template<typename T, typename IdxT>
+// inline void pack_list_data_float32(
+//   raft::resources const& handle,
+//   device_mdspan<T, typename list_spec<uint32_t, T, IdxT>::list_extents, row_major> list_data,
+//   device_matrix_view<const T, uint32_t, row_major> codes,
+//   uint32_t row_offset)
+// {
+//   auto stream = raft::resource::get_cuda_stream(handle);
+//   auto n_rows = codes.extent(0);
+//   if (n_rows == 0) { return; }
+
+//   auto dim = codes.extent(1);
+
+//   n_rows -= row_offset;
+//   constexpr uint32_t kBlockSize = 256;
+//   dim3 blocks(div_rounding_up_safe<uint32_t>(n_rows, kBlockSize), 1, 1);
+//   dim3 threads(kBlockSize, 1, 1);
+//   auto kernel = pack_list_data_kernel_float32<T, IdxT, kBlockSize>;
+//   kernel<<<blocks, threads, 0, stream>>>(codes.data_handle(),
+//   list_data.data_handle(),
+//   n_rows,
+//   dim,
+//   4);
+//   RAFT_CUDA_TRY(cudaPeekAtLastError());
+// }
+
 }  // namespace raft::neighbors::ivf_flat::detail

cpp/include/raft/neighbors/ivf_flat_helpers.cuh

@@ -0,0 +1,103 @@
+/*
+ * Copyright (c) 2023, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <raft/core/resource/cuda_stream.hpp>
+#include <raft/neighbors/detail/ivf_flat_build.cuh>
+#include <raft/neighbors/ivf_flat_types.hpp>
+
+#include <raft/core/device_mdspan.hpp>
+#include <raft/core/resources.hpp>
+
+// #include <omp.h>
+
+namespace raft::neighbors::ivf_flat::helpers {
+/**
+ * @defgroup ivf_flat_helpers Helper functions for manipulationg IVF Flat Index
+ * @{
+ */
+
+namespace codepacker {
+/**
+ * Write one flat code into a block by the given offset. The offset indicates the id of the record in the list. This function interleaves the code and is intended to later copy the interleaved codes over to the IVF list on device.
+ * NB: no memory allocation happens here; the block must fit the record (offset + 1).
+ *
+ * @tparam T
+ *
+ * @param[in] flat_code input flat code
+ * @param[out] block block of memory to write interleaved codes to
+ * @param[in] dim dimension of the flat code
+ * @param[in] veclen size of interleaved data chunks
+ * @param[in] offset how many records to skip before writing the data into the list
+ */
+template <typename T>
+__host__ __device__ void pack_1_interleaved(
+        const T* flat_code,
+        T* block,
+        uint32_t dim,
+        uint32_t veclen,
+        uint32_t offset) {    
+  // The data is written in interleaved groups of `index::kGroupSize` vectors
+  using interleaved_group = Pow2<kIndexGroupSize>;
+
+  // Interleave dimensions of the source vector while recording it.
+  // NB: such `veclen` is selected, that `dim % veclen == 0`
+    auto group_offset = interleaved_group::roundDown(offset);
+    auto ingroup_id = interleaved_group::mod(offset) * veclen;
+
+    for (uint32_t l = 0; l < dim; l += veclen) {
+      for (uint32_t j = 0; j < veclen; j++) {
+        block[group_offset * dim + l * kIndexGroupSize + ingroup_id + j] = flat_code[l + j];
+      }
+    }
+}
+
+/**
+ * Unpack 1 record of a single list (cluster) in the index to fetch the flat code. The offset indicates the id of the record. This function fetches one flat code from an interleaved code.
+ *
+ * @tparam T
+ *
+ * @param[in] block interleaved block. The block can be thought of as the whole inverted list in interleaved format.
+ * @param[out] flat_code output flat code
+ * @param[in] dim dimension of the flat code
+ * @param[in] veclen size of interleaved data chunks
+ * @param[in] offset fetch the flat code by the given offset
+ */
+template <typename T>
+__host__ __device__ void unpack_1_interleaved(
+        const T* block,
+        T* flat_code,
+        uint32_t dim,
+        uint32_t veclen,
+        uint32_t offset) {
+
+  // The data is written in interleaved groups of `index::kGroupSize` vectors
+  using interleaved_group = Pow2<kIndexGroupSize>;
+
+  // NB: such `veclen` is selected, that `dim % veclen == 0`
+  auto group_offset = interleaved_group::roundDown(offset);
+  auto ingroup_id = interleaved_group::mod(offset) * veclen;
+
+    for (uint32_t l = 0; l < dim; l += veclen) {
+      for (uint32_t j = 0; j < veclen; j++) {
+          flat_code[l + j] = block[group_offset * dim + l * kIndexGroupSize + ingroup_id + j];
+      }
+    }
+}
+} // namespace codepacker
+/** @} */
+}  // namespace raft::neighbors::ivf_flat::helpers

cpp/include/raft/spatial/knn/detail/ann_quantized.cuh

@@ -18,9 +18,14 @@
 
 #include "../ann_common.h"
 #include "../ivf_flat.cuh"
+#include <cstring>
 #include <raft/core/resource/cuda_stream.hpp>
 
 #include "processing.cuh"
+#include "raft/core/host_mdarray.hpp"
+#include "raft/neighbors/ivf_flat_types.hpp"
+#include "raft/neighbors/ivf_flat_helpers.cuh"
+#include "raft/util/pow2_utils.cuh"
 #include <raft/core/operators.hpp>
 #include <raft/util/cuda_utils.cuh>
 #include <raft/util/cudart_utils.hpp>
@@ -73,6 +78,91 @@ void approx_knn_build_index(raft::resources const& handle,
     auto new_params               = from_legacy_index_params(*ivf_ft_pams, metric, metricArg);
     index->ivf_flat<T, int64_t>() = std::make_unique<const ivf_flat::index<T, int64_t>>(
       ivf_flat::build(handle, new_params, index_array, int64_t(n), D));
+
+    // raft::resource::sync_stream(handle);
+
+    // auto old_list = index->ivf_flat<T, int64_t>()->lists()[0];
+    // uint32_t n_rows   = old_list->size.load();
+    // uint32_t roundup = Pow2<raft::neighbors::ivf_flat::kIndexGroupSize>::roundUp(n_rows);
+
+    // RAFT_LOG_INFO("roundup %d, n_rows %d", roundup, n_rows);
+
+    // if (n_rows == 0) { return; }
+
+    // auto dim = index->ivf_flat<T, int64_t>()->dim();
+    // auto veclen = index -> ivf_flat<T, int64_t>()->veclen();
+    // RAFT_LOG_INFO("roundup %d, n_rows %d, veclen %d, dim %d", roundup, n_rows, veclen, dim);
+    // auto codes   = make_host_matrix<T>(roundup, dim);
+    // auto block = make_host_matrix<T>(roundup, dim);
+
+    // T* firstArray;
+    // cudaMemcpy(&firstArray, index->ivf_flat<T, int64_t>()->data_ptrs().data_handle(), sizeof(float*), cudaMemcpyDeviceToHost);  // Copy the pointer to the first array from device to host
+
+    // raft::print_device_vector("codes_gpu", firstArray, 1, std::cout);
+    // raft::update_host(codes.data_handle(), firstArray, (size_t)(roundup * dim), stream);
+    // raft::resource::sync_stream(handle);
+    // raft::neighbors::ivf_flat::helpers::pack_host_interleaved(
+    //     codes.data_handle(),
+    //     block.data_handle(),
+    //     n_rows,
+    //     dim,
+    //     veclen);
+
+    // RAFT_LOG_INFO("veclen %d", veclen);
+    // raft::print_host_vector("codes", codes.data_handle(), roundup * dim, std::cout);
+    // raft::print_host_vector("block", block.data_handle(), roundup * dim, std::cout);
+    // // auto indices = make_device_vector<IdxT>(handle_, n_rows);
+    // copy(indices.data_handle(), old_list->indices.data_handle(), n_rows, stream_);
+
+    // ivf_flat::helpers::pack_list_data(handle_, *index, codes.view(), label, 0);
+    // ivf_pq::helpers::erase_list(handle_, index, label);
+    // ivf_pq::helpers::extend_list_with_codes<IdxT>(
+    //   handle_, index, codes.view(), indices.view(), label);
+
+    // auto& new_list = index->lists()[label];
+    // ASSERT_NE(old_list.get(), new_list.get())
+    //   << "The old list should have been shared and retained after ivf_pq index has erased the "
+    //      "corresponding cluster.";
+    // auto list_data_size = (n_rows / ivf_pq::kIndexGroupSize) * new_list->data.extent(1) *
+    //                       new_list->data.extent(2) * new_list->data.extent(3);
+
+    // ASSERT_TRUE(old_list->data.size() >= list_data_size);
+    // ASSERT_TRUE(new_list->data.size() >= list_data_size);
+    // ASSERT_TRUE(devArrMatch(old_list->data.data_handle(),
+    //                         new_list->data.data_handle(),
+    //                         list_data_size,
+    //                         Compare<uint8_t>{}));
+
+    // // Pack a few vectors back to the list.
+    // int row_offset = 9;
+    // int n_vec      = 3;
+    // ASSERT_TRUE(row_offset + n_vec < n_rows);
+    // size_t offset      = row_offset * index->pq_dim();
+    // auto codes_to_pack = make_device_matrix_view<const uint8_t, uint32_t>(
+    //   codes.data_handle() + offset, n_vec, index->pq_dim());
+    // ivf_pq::helpers::pack_list_data(handle_, index, codes_to_pack, label, row_offset);
+    // ASSERT_TRUE(devArrMatch(old_list->data.data_handle(),
+    //                         new_list->data.data_handle(),
+    //                         list_data_size,
+    //                         Compare<uint8_t>{}));
+
+    // Another test with the API that take list_data directly
+    // auto list_data  = index->lists()[label]->data.view();
+    // uint32_t n_take = 4;
+    // ASSERT_TRUE(row_offset + n_take < n_rows);
+    // auto codes2 = raft::make_device_matrix<uint8_t>(handle_, n_take, index->pq_dim());
+    // ivf_pq::helpers::codepacker::unpack(
+    //   handle_, list_data, index->pq_bits(), row_offset, codes2.view());
+
+    // // Write it back
+    // ivf_pq::helpers::codepacker::pack(
+    //   handle_, make_const_mdspan(codes2.view()), index->pq_bits(), row_offset, list_data);
+    // ASSERT_TRUE(devArrMatch(old_list->data.data_handle(),
+    //                         new_list->data.data_handle(),
+    //                         list_data_size,
+    //                         Compare<uint8_t>{}));
+  // }
+
   } else if (ivf_pq_pams) {
     neighbors::ivf_pq::index_params params;
     params.metric     = metric;