Remove fxdiv usages from XNNPACK, keeping it only for pthreadpool

BUILD.bazel

@@ -528,6 +528,7 @@ xnnpack_cc_library(
         ":microparams_h",
         ":operator_h",
         ":xnnpack_h",
+        "@FXdiv",
     ],
 )
 

BUILD.gn

@@ -918,6 +918,7 @@ xnnpack_source_set("scalar_microkernels") {
   deps = [
     ":microkernel_defs",
     ":microkernel_headers",
+    "//third_party/fxdiv",
   ]
   sources = ALL_SCALAR_MICROKERNEL_SRCS
 }
@@ -965,6 +966,7 @@ xnnpack_source_set("xnnpack") {
     ":scalar_microkernels",
     ":subgraph",
     ":table",
+    "//third_party/fxdiv",
   ]
   if (xnnpack_enable_arm_kleidiai) {
     deps += [ "//third_party/kleidiai" ]

CMakeLists.txt

@@ -445,7 +445,16 @@ IF(NOT XNNPACK_USE_SYSTEM_LIBS)
     EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" --build .
       WORKING_DIRECTORY "${CMAKE_BINARY_DIR}/cpuinfo-download")
     SET(CPUINFO_SOURCE_DIR "${CMAKE_BINARY_DIR}/cpuinfo-source" CACHE STRING "cpuinfo source directory")
+  ENDIF()
 
+  IF(NOT DEFINED FXDIV_SOURCE_DIR)
+    MESSAGE(STATUS "Downloading FXdiv to ${CMAKE_BINARY_DIR}/FXdiv-source (define FXDIV_SOURCE_DIR to avoid it)")
+    CONFIGURE_FILE(cmake/DownloadFXdiv.cmake "${CMAKE_BINARY_DIR}/FXdiv-download/CMakeLists.txt")
+    EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
+      WORKING_DIRECTORY "${CMAKE_BINARY_DIR}/FXdiv-download")
+    EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" --build .
+      WORKING_DIRECTORY "${CMAKE_BINARY_DIR}/FXdiv-download")
+    SET(FXDIV_SOURCE_DIR "${CMAKE_BINARY_DIR}/FXdiv-source" CACHE STRING "FXdiv source directory")
   ENDIF()
 
   IF(NOT DEFINED PTHREADPOOL_SOURCE_DIR)
@@ -1300,7 +1309,32 @@ IF(NOT TARGET pthreadpool)
 ENDIF()
 TARGET_LINK_LIBRARIES(xnnpack-base INTERFACE pthreadpool)
 
-
+# ---[ Configure FXdiv
+IF(NOT TARGET fxdiv)
+  IF(NOT XNNPACK_USE_SYSTEM_LIBS)
+    SET(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+    SET(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+    ADD_SUBDIRECTORY(
+      "${FXDIV_SOURCE_DIR}"
+      "${CMAKE_BINARY_DIR}/FXdiv")
+  ELSE()
+    FIND_FILE(FXDIV_HDR fxdiv.h PATH_SUFFIXES include PATHS "${FXDIV_SOURCE_DIR}")
+    IF(NOT FXDIV_HDR)
+      MESSAGE(FATAL_ERROR "Cannot find fxdiv")
+    ENDIF()
+    ADD_LIBRARY(fxdiv STATIC "${FXDIV_HDR}")
+    TARGET_INCLUDE_DIRECTORIES(fxdiv INTERFACE "${FXDIV_SOURCE_DIR}/include")
+    SET_PROPERTY(TARGET fxdiv PROPERTY LINKER_LANGUAGE C)
+  ENDIF()
+ENDIF()
+IF(XNNPACK_BUILD_ALL_MICROKERNELS)
+  TARGET_LINK_LIBRARIES(xnnpack-microkernels-all PRIVATE fxdiv)
+ENDIF()
+TARGET_LINK_LIBRARIES(xnnpack-microkernels-prod PRIVATE fxdiv)
+TARGET_LINK_LIBRARIES(xnnpack-indirection PRIVATE fxdiv)
+IF(XNNPACK_BUILD_LIBRARY)
+  TARGET_LINK_LIBRARIES(XNNPACK PRIVATE fxdiv)
+ENDIF()
 
 IF(XNNPACK_BUILD_LIBRARY)
   INSTALL(TARGETS XNNPACK xnnpack-microkernels-prod

MODULE.bazel

@@ -64,15 +64,16 @@ http_archive(
     urls = ["https://github.com/google/benchmark/archive/7da00e8f6763d6e8c284d172c9cfcc5ae0ce9b7a.zip"],
 )
 # LINT.ThenChange(cmake/DownloadGoogleBenchmark.cmake)
+
+# LINT.IfChange(FXdiv)
 # FXdiv library, used for repeated integer division by the same factor
 http_archive(
     name = "FXdiv",
     sha256 = "ab7dfb08829bee33dca38405d647868fb214ac685e379ec7ef2bebcd234cd44d",
     strip_prefix = "FXdiv-b408327ac2a15ec3e43352421954f5b1967701d1",
     urls = ["https://github.com/Maratyszcza/FXdiv/archive/b408327ac2a15ec3e43352421954f5b1967701d1.zip"],
 )
-
-
+# LINT.ThenChange(cmake/DownloadFXdiv.cmake)
 
 # LINT.IfChange(pthreadpool)
 # pthreadpool library, used for parallelization

build_params.bzl

@@ -212,6 +212,7 @@ XNNPACK_PARAMS_FOR_ARCH = {
         ],
         extra_deps = [
             "//src/configs:config_hdrs",
+            "@FXdiv",
         ],
     ),
 

cmake/DownloadFXdiv.cmake

@@ -0,0 +1,30 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# Copyright 2019 Google LLC
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+CMAKE_MINIMUM_REQUIRED(VERSION 3.5 FATAL_ERROR)
+
+PROJECT(fxdiv-download NONE)
+
+# Set file timestamps to the time of extraction.
+IF(POLICY CMP0135)
+  CMAKE_POLICY(SET CMP0135 NEW)
+ENDIF()
+
+# LINT.IfChange
+INCLUDE(ExternalProject)
+ExternalProject_Add(fxdiv
+  URL https://github.com/Maratyszcza/FXdiv/archive/b408327ac2a15ec3e43352421954f5b1967701d1.zip
+  URL_HASH SHA256=ab7dfb08829bee33dca38405d647868fb214ac685e379ec7ef2bebcd234cd44d
+  SOURCE_DIR "${CMAKE_BINARY_DIR}/FXdiv-source"
+  BINARY_DIR "${CMAKE_BINARY_DIR}/FXdiv"
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND ""
+  INSTALL_COMMAND ""
+  TEST_COMMAND ""
+)
+# LINT.ThenChange(../MODULE.bazel:FXdiv)

src/indirection.c

@@ -9,6 +9,7 @@
 #include "src/xnnpack/indirection.h"
 
 #include <assert.h>
+#include <fxdiv.h>
 #include <math.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -40,25 +41,17 @@ void xnn_indirection_init_conv2d(
   size_t input_padding_top,
   size_t input_padding_left)
 {
-  assert(output_height != 0);
-  assert(output_width != 0);
   const size_t output_size = output_height * output_width;
   const size_t kernel_size = kernel_height * kernel_width;
 
-  size_t cur_oy = output_start / output_width;
-  size_t cur_ox = output_start % output_width;
-  const size_t clamp_oy = (output_size - 1) / output_width;
-  const size_t clamp_ox = (output_size - 1) % output_width;
+  const struct fxdiv_divisor_size_t output_width_divisor = fxdiv_init_size_t(output_width);
 
   for (size_t output_tile_start = output_start; output_tile_start < output_end; output_tile_start += output_tile_size) {
-    size_t oy = cur_oy;
-    size_t ox = cur_ox;
     for (size_t output_tile_offset = 0; output_tile_offset < output_tile_size; output_tile_offset++) {
-      const size_t output_index = output_tile_start + output_tile_offset;
-      const bool clamped = output_index >= output_size - 1;
-      const size_t output_x = clamped ? clamp_ox : ox;
-      const size_t output_y = clamped ? clamp_oy : oy;
-
+      const size_t output_index = min(output_tile_start + output_tile_offset, output_size - 1);
+      const struct fxdiv_result_size_t output_y_x = fxdiv_divide_size_t(output_index, output_width_divisor);
+      const size_t output_x = output_y_x.remainder;
+      const size_t output_y = output_y_x.quotient;
       for (size_t kernel_y = 0; kernel_y < kernel_height; kernel_y++) {
         const size_t input_y = output_y * stride_height + kernel_y * dilation_height - input_padding_top;
         if (input_y < input_height) {
@@ -81,17 +74,7 @@ void xnn_indirection_init_conv2d(
           }
         }
       }
-
-      if (output_index < output_size - 1) {
-        ox++;
-        if (ox == output_width) {
-          ox = 0;
-          oy++;
-        }
-      }
     }
-    cur_oy = oy;
-    cur_ox = ox;
   }
 }
 
@@ -118,26 +101,22 @@ void xnn_indirection_init_deconv2d(
   const size_t tiled_output_size = round_up(output_size, output_tile_size);
   const size_t kernel_size = kernel_height * kernel_width;
 
-  size_t cur_oy = 0;
-  size_t cur_ox = 0;
-  const size_t clamp_oy = (output_size - 1) / output_width;
-  const size_t clamp_ox = (output_size - 1) % output_width;
+  const struct fxdiv_divisor_size_t output_width_divisor = fxdiv_init_size_t(output_width);
+  const struct fxdiv_divisor_size_t stride_height_divisor = fxdiv_init_size_t(stride_height);
+  const struct fxdiv_divisor_size_t stride_width_divisor = fxdiv_init_size_t(stride_width);
 
   for (size_t output_tile_start = 0; output_tile_start < tiled_output_size; output_tile_start += output_tile_size) {
-    size_t oy = cur_oy;
-    size_t ox = cur_ox;
     for (size_t output_tile_offset = 0; output_tile_offset < output_tile_size; output_tile_offset++) {
-      const size_t output_index = output_tile_start + output_tile_offset;
-      const bool clamped = output_index >= output_size - 1;
-      const size_t output_x = clamped ? clamp_ox : ox;
-      const size_t output_y = clamped ? clamp_oy : oy;
-
+      const size_t output_index = min(output_tile_start + output_tile_offset, output_size - 1);
+      const struct fxdiv_result_size_t output_y_x = fxdiv_divide_size_t(output_index, output_width_divisor);
+      const size_t output_x = output_y_x.remainder;
+      const size_t output_y = output_y_x.quotient;
       for (size_t kernel_y = 0; kernel_y < kernel_height; kernel_y++) {
         const size_t y = output_y + padding_top - kernel_y * dilation_height;
-        const size_t input_y = y / stride_height;
+        const size_t input_y = fxdiv_quotient_size_t(y, stride_height_divisor);
         for (size_t kernel_x = 0; kernel_x < kernel_width; kernel_x++) {
           const size_t x = output_x + padding_left - kernel_x * dilation_width;
-          const size_t input_x = x / stride_width;
+          const size_t input_x = fxdiv_quotient_size_t(x, stride_width_divisor);
           const size_t kernel_index = kernel_y * kernel_width + kernel_x;
           const size_t index = output_tile_start * kernel_size + kernel_index * output_tile_size + output_tile_offset;
           if (input_y * stride_height == y && input_y < input_height && input_x * stride_width == x && input_x < input_width) {
@@ -147,17 +126,7 @@ void xnn_indirection_init_deconv2d(
           }
         }
       }
-
-      if (output_index < output_size - 1) {
-        ox++;
-        if (ox == output_width) {
-          ox = 0;
-          oy++;
-        }
-      }
     }
-    cur_oy = oy;
-    cur_ox = ox;
   }
 }
 

src/u8-lut32norm/u8-lut32norm-scalar.c

@@ -8,6 +8,8 @@
 
 #include <assert.h>
 
+#include <fxdiv.h>
+
 #include "src/xnnpack/lut.h"
 
 
@@ -37,11 +39,12 @@ void xnn_u8_lut32norm_ukernel__scalar(
   const uint32_t vsum = compute_sum(n, x, t);
   assert(vsum != 0);
 
+  struct fxdiv_divisor_uint32_t vsum_divisor = fxdiv_init_uint32_t(vsum);
   const uint32_t vrounding = (vsum >> 1);
   do {
     const size_t vx = *x++;
     const uint32_t vt = t[vx];
-    const uint32_t vq = ((vt << 8) + vrounding) / vsum;
+    const uint32_t vq = fxdiv_quotient_uint32_t((vt << 8) + vrounding, vsum_divisor);
     const uint8_t vy = vq > 255 ? UINT8_C(255) : (uint8_t) vq;
     *y++ = vy;
   } while (--n != 0);