Merge pull request #174 from aous72/improving_wasm

Adding Support for WASM without SIMD.
aous72 · Feb 2, 2025 · 2ba1b73 · 2ba1b73
2 parents 7cd6186 + 03802f0
commit 2ba1b73
Show file tree

Hide file tree

Showing 10 changed files with 144 additions and 76 deletions.
diff --git a/.github/workflows/emcc.yml b/.github/workflows/emcc.yml
@@ -18,8 +18,30 @@ jobs:
         with:
           actions-cache-folder: 'emsdk-cache'
 
-      - name: Build
+      - name: Build non-SIMD and Debug
         run: |
           cd build
-          emcmake cmake ..
-          emmake make
+          emcmake cmake .. --fresh -DOJPH_DISABLE_SIMD=ON -DCMAKE_BUILD_TYPE=Debug
+          cmake --build . --config Debug --clean-first
+          cd ..
+
+      - name: Build non-SIMD and Release
+        run: |
+          cd build
+          emcmake cmake .. --fresh -DOJPH_DISABLE_SIMD=ON -DCMAKE_BUILD_TYPE=Release
+          cmake --build . --config Release --clean-first
+          cd ..
+
+      - name: Build SIMD and Debug
+        run: |
+          cd build
+          emcmake cmake .. --fresh -DOJPH_DISABLE_SIMD=OFF -DCMAKE_BUILD_TYPE=Debug
+          cmake --build . --config Debug --clean-first
+          cd ..
+
+      - name: Build SIMD and Release
+        run: |
+          cd build
+          emcmake cmake .. --fresh -DOJPH_DISABLE_SIMD=OFF -DCMAKE_BUILD_TYPE=Release
+          cmake --build . --config Release --clean-first
+          cd ..
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -58,16 +58,26 @@ if (DEFINED OJPH_ENABLE_INTEL_AVX512)
 endif()
 
 ## Setting some of the options if EMSCRIPTEN is the compiler
-# WebAssembly SIMD is treated differently.  The SIMD flags above have no effect on the 
-# use of WASM SIMD.  This is because, for WASM, both non-SIMD and SIMD are required,
-# and therefore two sets of binaries are generated. For CPUs, one binary can carry both 
-# non-SIMD and SIMD, and the program, at run-time, can decide which path to follow, 
-# depending on what CPU instructions are available.
+# In previous releases, the cmake script used to produce both non-SIMD and 
+# SIMD builds in one go.  At the time of this writing, all interpreters and 
+# compilers of WASM code, such as web-browser and node, support SIMD, therefore 
+# it is time to make the SIMD build the default.  In other words, this cmake 
+# script builds only WASM SIMD code by default, if desired, a non-SIMD build
+# can be generated using the OJPH_DISABLE_SIMD option (in this case, the 
+# WASM SIMD code is not generated).
+# It is worth remembering that the SIMD/non-SIMD issue arose because it is
+# NOT possible to have multiple execution paths in the code, one for non-SIMD
+# and one for SIMD, as we do for CPUs, letting the program select, at run-time, 
+# the best path to follow.
 if(EMSCRIPTEN)
   set(BUILD_SHARED_LIBS OFF)
   set(OJPH_ENABLE_TIFF_SUPPORT OFF)
   set(OJPH_BUILD_STREAM_EXPAND OFF)
-  set(OJPH_DISABLE_SIMD ON)
+  if (OJPH_DISABLE_SIMD)
+    set(OJPH_ENABLE_WASM_SIMD OFF)
+  else()
+    set(OJPH_ENABLE_WASM_SIMD ON)
+  endif()
 endif()
 
 # This is related to how the timestamp is set for URL downloaded files.
@@ -106,6 +116,12 @@ if (CMAKE_CXX_COMPILER_ID MATCHES "Clang|GNU")
   -Wunused-parameter
   )
 endif()
+if (EMSCRIPTEN)
+  add_compile_options(-fexceptions)
+  if(OJPH_ENABLE_WASM_SIMD)
+    add_compile_options(-DOJPH_ENABLE_WASM_SIMD -msimd128)
+  endif()
+endif()
 
 ## Enhanced instruction options
 if (OJPH_DISABLE_SIMD)

diff --git a/src/apps/CMakeLists.txt b/src/apps/CMakeLists.txt
@@ -19,7 +19,6 @@ endif()
 
 if (EMSCRIPTEN)
   add_link_options(-sWASM=1 -sASSERTIONS=1 -sALLOW_MEMORY_GROWTH=1 -sNODERAWFS=1 -sENVIRONMENT=node -sEXIT_RUNTIME=1 -sEXCEPTION_CATCHING_ALLOWED=['fake'])
-  add_compile_options(-DOJPH_ENABLE_WASM_SIMD -msimd128 -msse4.1 -std=c++11 -O3 -fexceptions)
 endif()
 
 ## Build executables

diff --git a/src/apps/ojph_compress/CMakeLists.txt b/src/apps/ojph_compress/CMakeLists.txt
@@ -14,7 +14,11 @@ source_group("others"      FILES ${OJPH_IMG_IO})
 source_group("common"      FILES ${OJPH_IMG_IO_H})
 
 if(EMSCRIPTEN)
-  source_group("others" FILES ${OJPH_IMG_IO_SSE4})
+  if (OJPH_ENABLE_WASM_SIMD)
+    list(APPEND SOURCES ${OJPH_IMG_IO_SSE4})
+    source_group("others" FILES ${OJPH_IMG_IO_SSE4})
+    set_source_files_properties(${OJPH_IMG_IO_SSE4} PROPERTIES COMPILE_FLAGS -msse4.1)
+  endif()
 else()
   if (NOT OJPH_DISABLE_SIMD)
     if (("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_X86_64") OR ("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_I386"))

diff --git a/src/apps/ojph_expand/CMakeLists.txt b/src/apps/ojph_expand/CMakeLists.txt
@@ -1,7 +1,6 @@
 ## building ojph_expand
 #######################
 
-
 file(GLOB OJPH_EXPAND         "ojph_expand.cpp")
 file(GLOB OJPH_IMG_IO         "../others/ojph_img_io.cpp")
 file(GLOB OJPH_IMG_IO_SSE4    "../others/ojph_img_io_sse41.cpp")
@@ -15,7 +14,11 @@ source_group("others"      FILES ${OJPH_IMG_IO})
 source_group("common"      FILES ${OJPH_IMG_IO_H})
 
 if(EMSCRIPTEN)
-  source_group("others" FILES ${OJPH_IMG_IO_SSE4})
+  if (OJPH_ENABLE_WASM_SIMD)
+    list(APPEND SOURCES ${OJPH_IMG_IO_SSE4})
+    source_group("others" FILES ${OJPH_IMG_IO_SSE4})
+    set_source_files_properties(${OJPH_IMG_IO_SSE4} PROPERTIES COMPILE_FLAGS -msse4.1)
+  endif()
 else()
   if (NOT OJPH_DISABLE_SIMD)
     if (("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_X86_64") OR ("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_I386"))

diff --git a/src/apps/others/ojph_img_io.cpp b/src/apps/others/ojph_img_io.cpp
@@ -458,6 +458,8 @@ namespace ojph {
     samples_per_line = num_components * width;
     bytes_per_line = bytes_per_sample * samples_per_line;
 
+#if !defined(OJPH_ENABLE_WASM_SIMD) || !defined(OJPH_EMSCRIPTEN)
+
     if (bytes_per_sample == 1) {
       if (num_components == 1) 
         converter = gen_cvrt_32b1c_to_8ub1c;
@@ -471,49 +473,66 @@ namespace ojph {
         converter = gen_cvrt_32b3c_to_16ub3c_be;
     }
 
-#ifndef OJPH_DISABLE_SIMD
+  #ifndef OJPH_DISABLE_SIMD
 
-  #if (defined(OJPH_ARCH_X86_64) || defined(OJPH_ARCH_I386))
+    #if (defined(OJPH_ARCH_X86_64) || defined(OJPH_ARCH_I386))
 
-    #ifndef OJPH_DISABLE_SSE4
-      if (get_cpu_ext_level() >= X86_CPU_EXT_LEVEL_SSE41) {
-        if (bytes_per_sample == 1) {
-          if (num_components == 1) 
-            converter = sse41_cvrt_32b1c_to_8ub1c;
-          else
-            converter = sse41_cvrt_32b3c_to_8ub3c;
-        }
-        else {
-          if (num_components == 1) 
-            converter = sse41_cvrt_32b1c_to_16ub1c_be;
-          else
-            converter = sse41_cvrt_32b3c_to_16ub3c_be;
-        }
-      }
-    #endif // !OJPH_DISABLE_SSE4
-
-    #ifndef OJPH_DISABLE_AVX2
-      if (get_cpu_ext_level() >= X86_CPU_EXT_LEVEL_AVX2) {
-        if (bytes_per_sample == 1) {
-          if (num_components == 1) 
-            converter = avx2_cvrt_32b1c_to_8ub1c;
-          else
-            converter = avx2_cvrt_32b3c_to_8ub3c;
+      #ifndef OJPH_DISABLE_SSE4
+        if (get_cpu_ext_level() >= X86_CPU_EXT_LEVEL_SSE41) {
+          if (bytes_per_sample == 1) {
+            if (num_components == 1) 
+              converter = sse41_cvrt_32b1c_to_8ub1c;
+            else
+              converter = sse41_cvrt_32b3c_to_8ub3c;
+          }
+          else {
+            if (num_components == 1) 
+              converter = sse41_cvrt_32b1c_to_16ub1c_be;
+            else
+              converter = sse41_cvrt_32b3c_to_16ub3c_be;
+          }
         }
-        else {
-          if (num_components == 1) 
-            converter = avx2_cvrt_32b1c_to_16ub1c_be;
-          else
-            { } // did not find an implementation better than sse41
+      #endif // !OJPH_DISABLE_SSE4
+
+      #ifndef OJPH_DISABLE_AVX2
+        if (get_cpu_ext_level() >= X86_CPU_EXT_LEVEL_AVX2) {
+          if (bytes_per_sample == 1) {
+            if (num_components == 1) 
+              converter = avx2_cvrt_32b1c_to_8ub1c;
+            else
+              converter = avx2_cvrt_32b3c_to_8ub3c;
+          }
+          else {
+            if (num_components == 1) 
+              converter = avx2_cvrt_32b1c_to_16ub1c_be;
+            else
+              { } // did not find an implementation better than sse41
+          }
         }
-      }
-    #endif // !OJPH_DISABLE_AVX2
+      #endif // !OJPH_DISABLE_AVX2
+
+    #elif defined(OJPH_ARCH_ARM)
 
-  #elif defined(OJPH_ARCH_ARM)
+    #endif // !(defined(OJPH_ARCH_X86_64) || defined(OJPH_ARCH_I386))
 
-  #endif // !(defined(OJPH_ARCH_X86_64) || defined(OJPH_ARCH_I386))
+  #endif // !OJPH_DISABLE_SIMD
 
-#endif // !OJPH_DISABLE_SIMD
+#else // OJPH_ENABLE_WASM_SIMD
+
+    if (bytes_per_sample == 1) {
+      if (num_components == 1) 
+        converter = sse41_cvrt_32b1c_to_8ub1c;
+      else
+        converter = sse41_cvrt_32b3c_to_8ub3c;
+    }
+    else {
+      if (num_components == 1) 
+        converter = sse41_cvrt_32b1c_to_16ub1c_be;
+      else
+        converter = sse41_cvrt_32b3c_to_16ub3c_be;
+    }
+
+#endif // !OJPH_ENABLE_WASM_SIMD
   }
 
   ////////////////////////////////////////////////////////////////////////////

diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
@@ -32,13 +32,12 @@ source_group("others"            FILES ${OTHERS})
 source_group("transform"         FILES ${TRANSFORM})
 
 if(EMSCRIPTEN)
-  add_compile_options(-std=c++11 -O3 -fexceptions -DOJPH_ENABLE_WASM_SIMD -msimd128)
-  list(APPEND SOURCES ${CODESTREAM_WASM} ${CODING_WASM} ${TRANSFORM_WASM})
-
-  source_group("codestream" FILES ${CODESTREAM_WASM})
-  source_group("coding" FILES ${CODING_WASM})
-  source_group("transform" FILES ${TRANSFORM_WASM})
-
+  if (OJPH_ENABLE_WASM_SIMD)
+    list(APPEND SOURCES ${CODESTREAM_WASM} ${CODING_WASM} ${TRANSFORM_WASM})
+    source_group("codestream" FILES ${CODESTREAM_WASM})
+    source_group("coding" FILES ${CODING_WASM})
+    source_group("transform" FILES ${TRANSFORM_WASM})
+  endif()
 else()
   if (NOT OJPH_DISABLE_SIMD)
     if (("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_X86_64") OR ("${OJPH_TARGET_ARCH}" MATCHES "OJPH_ARCH_I386"))

diff --git a/subprojects/js/CMakeLists.txt b/subprojects/js/CMakeLists.txt
@@ -2,18 +2,26 @@ cmake_minimum_required(VERSION 3.10.0)
 
 set(CMAKE_SYSTEM_NAME Generic)
 
-project (openjphwasm DESCRIPTION "Open source implementation of JPH" LANGUAGES CXX)
-
-set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/../html)
-set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/../html)
+project (OpenJPH_WASM DESCRIPTION "Open source implementation of JPH" LANGUAGES CXX)
 
 add_subdirectory("../.." openjph EXCLUDE_FROM_ALL)
 add_executable(libopenjph "src/ojph_wrapper.cpp")
-set_target_properties(libopenjph PROPERTIES SUFFIX ".js" LINK_FLAGS "-O3 -fexceptions -s WASM=1 -s EXPORT_ES6=1 -s MODULARIZE=1 -s ENVIRONMENT=web -s EXPORTED_FUNCTIONS=[_free,_malloc] -s EXPORTED_RUNTIME_METHODS=[ccall,cwrap,writeArrayToMemory] -s NO_EXIT_RUNTIME=1 -s ALLOW_MEMORY_GROWTH=1 -s INITIAL_MEMORY=134217728")
+if (OJPH_DISABLE_SIMD)
+else()
+  target_compile_options(libopenjph PRIVATE -DOJPH_ENABLE_WASM_SIMD -msimd128)
+endif()
+set_target_properties(libopenjph PROPERTIES SUFFIX ".js")
+target_link_options(libopenjph PRIVATE
+  -fexceptions
+  -sWASM=1
+  -sEXPORT_ES6=1
+  -sMODULARIZE=1
+  -sENVIRONMENT=web
+  -sEXPORTED_FUNCTIONS=[_free,_malloc]
+  -sEXPORTED_RUNTIME_METHODS=[ccall,cwrap,writeArrayToMemory]
+  -sNO_EXIT_RUNTIME=1
+  -sALLOW_MEMORY_GROWTH=1
+  -sINITIAL_MEMORY=134217728
+)
 target_link_libraries(libopenjph PRIVATE openjph)
 
-add_executable(libopenjph_simd "src/ojph_wrapper.cpp" )
-target_compile_options(libopenjph_simd PRIVATE -DOJPH_ENABLE_WASM_SIMD -msimd128)
-set_target_properties(libopenjph_simd PROPERTIES SUFFIX ".js" LINK_FLAGS "-O3 -fexceptions -s WASM=1 -s EXPORT_ES6=1 -s MODULARIZE=1 -s ENVIRONMENT=web -s EXPORTED_FUNCTIONS=[_free,_malloc] -s EXPORTED_RUNTIME_METHODS=[ccall,cwrap,writeArrayToMemory] -s NO_EXIT_RUNTIME=1 -s ALLOW_MEMORY_GROWTH=1 -s INITIAL_MEMORY=134217728")
-target_link_libraries(libopenjph_simd PRIVATE openjphsimd)
-
diff --git a/subprojects/js/build.sh b/subprojects/js/build.sh
@@ -1,5 +1,9 @@
-#!/bin/sh
+#!/bin/bash
+
 mkdir -p build
-#(cd build && emcmake cmake -DCMAKE_BUILD_TYPE=Debug ..)
-(cd build && emcmake cmake ..)
-(cd build && emmake make VERBOSE=1 -j)
+cd build
+emcmake cmake .. -DCMAKE_BUILD_TYPE=Release -DOJPH_DISABLE_SIMD=ON && emmake make -j8 && mv libopenjph.* ../html/
+emcmake cmake .. -DCMAKE_BUILD_TYPE=Release -DOJPH_DISABLE_SIMD=OFF && emmake make -j8 && mv libopenjph.wasm ../html/libopenjph_simd.wasm
+cd ..
+sed 's/libopenjph.wasm/libopenjph_simd.wasm/g' build/libopenjph.js > html/libopenjph_simd.js
+rm build/libopenjph.js
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
@@ -69,15 +69,9 @@ else()
     COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE:ojph_compress>" "./"
   )
   if(EMSCRIPTEN)
-    add_custom_command(TARGET test_executables POST_BUILD
-      COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE:ojph_expand_simd>" "./"
-      COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE:ojph_compress_simd>" "./"
-    )
     add_custom_command(TARGET test_executables POST_BUILD
       COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE_DIR:ojph_expand>/ojph_expand.wasm" "./"
       COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE_DIR:ojph_compress>/ojph_compress.wasm" "./"
-      COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE_DIR:ojph_expand_simd>/ojph_expand_simd.wasm" "./"
-      COMMAND ${CMAKE_COMMAND} -E copy "$<TARGET_FILE_DIR:ojph_compress_simd>/ojph_compress_simd.wasm" "./"
     )
   endif(EMSCRIPTEN)
   if(MSYS)