Optimized NTT parallel implementation on CPU with Taskflow (#714)

icicle/CMakeLists.txt

@@ -86,6 +86,7 @@ add_library(icicle_device SHARED
   src/config_extension.cpp
 )
 target_link_libraries(icicle_device PUBLIC dl)
+
 include_directories(include)
 
 # Define the install directory (default is /usr/local)

icicle/backend/cpu/CMakeLists.txt

@@ -1,5 +1,25 @@
 cmake_minimum_required(VERSION 3.18)
 
+message(STATUS "Fetching Taskflow v3.8.0 (CPU backend)")
+include(FetchContent)
+FetchContent_Declare(
+  Taskflow
+  GIT_REPOSITORY https://github.com/taskflow/taskflow.git
+  GIT_TAG v3.8.0
+  GIT_SHALLOW TRUE
+)
+# Disable unnecessary components
+set(TF_BUILD_BENCHMARKS OFF CACHE BOOL "Disable Taskflow benchmarks" FORCE)
+set(TF_BUILD_PROFILER OFF CACHE BOOL "Disable Taskflow profiler" FORCE)
+set(TF_BUILD_CUDA OFF CACHE BOOL "Disable Taskflow CUDA support" FORCE)
+set(TF_BUILD_SYCL OFF CACHE BOOL "Disable Taskflow SYCL support" FORCE)
+set(TF_BUILD_TESTS OFF CACHE BOOL "Disable Taskflow tests" FORCE)
+set(TF_BUILD_EXAMPLES OFF CACHE BOOL "Disable Taskflow examples" FORCE)
+
+FetchContent_MakeAvailable(Taskflow)
+# Use icicle_device as interface for TaskFlow headers
+target_include_directories(icicle_device INTERFACE ${Taskflow_SOURCE_DIR})
+
 # CPU backend is built directly into icicle library
 
 target_sources(icicle_device PRIVATE src/cpu_device_api.cpp)

icicle/backend/cpu/include/cpu_ntt_domain.h

@@ -1,22 +1,14 @@
 #pragma once
 #include "icicle/backend/ntt_backend.h"
 #include "icicle/errors.h"
-#include "icicle/runtime.h"
 #include "icicle/utils/log.h"
-#include "icicle/fields/field_config.h"
-#include "icicle/vec_ops.h"
 
-#include <thread>
-#include <vector>
-#include <chrono>
 #include <algorithm>
-#include <iostream>
 #include <cmath>
 #include <cstdint>
 #include <memory>
 #include <mutex>
 
-using namespace field_config;
 using namespace icicle;
 namespace ntt_cpu {
 

icicle/backend/cpu/include/cpu_ntt_main.h

@@ -1,9 +1,6 @@
 #pragma once
-#include "icicle/utils/log.h"
 #include "ntt_cpu.h"
-#include <iostream>
 
-using namespace field_config;
 using namespace icicle;
 
 /**
@@ -14,16 +11,14 @@ using namespace icicle;
  * hierarchy, and memory management for efficient computation.
  *
  * The NTT problem is given at a specific size and is divided into subproblems to enable
- * parallel solving of independent tasks, ensuring that the number of problems solved
- * simultaneously does not exceed cache size. The original problem is divided into hierarchies
+ * parallel solving of independent tasks. The original problem is divided into hierarchies
  * of subproblems. Beyond a certain size, the problem is divided into two layers of sub-NTTs in
  * hierarchy 1. Within hierarchy 1, the problem is further divided into 1-3 layers of sub-NTTs
  * belonging to hierarchy 0. The division into hierarchies and the sizes of the sub-NTTs are
  * determined by the original problem size.
  *
  * The sub-NTTs within hierarchy 0 are the units of work that are assigned to individual threads.
- * The overall computation is executed in a multi-threaded fashion, with the degree of parallelism
- * determined by the number of available hardware cores.
+ * The overall computation is executed in a multi-threaded fashion.
  *
  * @param device The device on which the NTT is being performed.
  * @param input Pointer to the input data.

icicle/backend/cpu/include/ntt_data.h

@@ -1,16 +1,11 @@
 #pragma once
-#include "icicle/utils/log.h"
-#include "tasks_manager.h"
 #include "cpu_ntt_domain.h"
 
-// #include <_types/_uint32_t.h>
-#include <csetjmp>
 #include <sys/types.h>
 #include <deque>
-#include <functional>
 #include <unordered_map>
 
-#define HIERARCHY_1 22
+#define HIERARCHY_1 26
 
 namespace ntt_cpu {
 
@@ -25,19 +20,17 @@ namespace ntt_cpu {
    * layer, 13 for the second, and 0 for the third.
    */
   constexpr uint32_t layers_sub_logn[31][3] = {
-    {0, 0, 0},   {1, 0, 0},   {2, 0, 0},   {3, 0, 0},   {4, 0, 0},   {5, 0, 0},   {3, 3, 0},  {4, 3, 0},
-    {4, 4, 0},   {5, 4, 0},   {5, 5, 0},   {4, 4, 3},   {4, 4, 4},   {5, 4, 4},   {5, 5, 4},  {5, 5, 5},
-    {5, 5, 6},   {5, 5, 7},   {5, 5, 8},   {5, 5, 9},   {5, 5, 10},  {5, 5, 11},  {5, 5, 12}, {12, 11, 0},
-    {12, 12, 0}, {13, 12, 0}, {13, 13, 0}, {14, 13, 0}, {14, 14, 0}, {15, 14, 0}, {15, 15, 0}};
+    {0, 0, 0},  {1, 0, 0},  {2, 0, 0},  {3, 0, 0},   {4, 0, 0},   {5, 0, 0},   {3, 3, 0},  {4, 3, 0},
+    {4, 4, 0},  {5, 4, 0},  {5, 5, 0},  {4, 4, 3},   {4, 4, 4},   {5, 4, 4},   {5, 5, 4},  {5, 5, 5},
+    {5, 5, 6},  {5, 5, 7},  {5, 5, 8},  {5, 5, 9},   {5, 5, 10},  {5, 5, 11},  {5, 5, 12}, {5, 5, 13},
+    {5, 5, 14}, {5, 5, 15}, {5, 5, 16}, {14, 13, 0}, {14, 14, 0}, {15, 14, 0}, {15, 15, 0}};
 
   /**
    * @brief Represents the log sizes of sub-NTTs in the NTT computation hierarchy.
    *
    * This struct stores the log sizes of the sub-NTTs for both hierarchy_0 and hierarchy_1  layers,
    * based on the overall log size (`logn`) of the NTT problem.
    *
-   * @param logn The log size of the entire NTT problem.
-   * @param size The size of the NTT problem, calculated as `1 << logn`.
    * @param hierarchy_0_layers_sub_logn Log sizes of sub-NTTs for hierarchy_0 layers.
    * @param hierarchy_1_layers_sub_logn Log sizes of sub-NTTs for hierarchy_1 layers.
    *
@@ -75,7 +68,7 @@ namespace ntt_cpu {
     const uint32_t logn;                         // log of the original NTT size.
     const uint32_t size;                         // Size of the original NTT problem.
     const NttSubHierarchies ntt_sub_hierarchies; // Log sizes of sub-NTTs based on the original NTT log size.
-    E* const elements;                           // Pointer to the output elements array.
+    E* elements;                                 // Pointer to the elements array.
     const NTTConfig<S>& config;                  // Configuration settings for the NTT computation.
     const NTTDir direction;                      // Direction of the NTT computation (forward or inverse).
     const bool is_parallel;                      // Flag indicating if the NTT computation is parallel.