[js/web] Add Wasm Relaxed SIMD support to wasm backend (#22794)

### Description
<!-- Describe your changes. -->
Add Wasm Relaxed SIMD support. 
Use integer dot product instructions for QGemmU8X8.
1. Build with --enable_wasm_relaxed_simd
2. Use env.wasm.relaxedSimd to run it


### Motivation and Context
#22533

---------

Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>

Author: jing-bao

cmake/CMakeLists.txt

@@ -193,6 +193,7 @@ option(onnxruntime_ENABLE_WEBASSEMBLY_DEBUG_INFO "Enable this option to turn on
 option(onnxruntime_ENABLE_WEBASSEMBLY_PROFILING "Enable this option to turn on WebAssembly profiling and preserve function names" OFF)
 option(onnxruntime_ENABLE_WEBASSEMBLY_OUTPUT_OPTIMIZED_MODEL "Enable this option to allow WebAssembly to output optimized model" OFF)
 option(onnxruntime_ENABLE_WEBASSEMBLY_MEMORY64 "Enable this option to allow WebAssembly to use 64bit memory" OFF)
+option(onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD "Enable WebAssembly Relaxed SIMD" OFF)
 
 # Enable bitcode for iOS
 option(onnxruntime_ENABLE_BITCODE "Enable bitcode for iOS only" OFF)

cmake/adjust_global_compile_flags.cmake

@@ -35,7 +35,10 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
     set(CMAKE_CXX_FLAGS_DEBUG "-g2")
   endif()
 
-  if (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
+  if (onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD)
+    string(APPEND CMAKE_C_FLAGS " -msimd128 -mrelaxed-simd")
+    string(APPEND CMAKE_CXX_FLAGS " -msimd128 -mrelaxed-simd")
+  elseif (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
     string(APPEND CMAKE_C_FLAGS " -msimd128")
     string(APPEND CMAKE_CXX_FLAGS " -msimd128")
   endif()

cmake/external/xnnpack.cmake

@@ -143,7 +143,12 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
   list(APPEND wasm_srcs ${XNNPACK_DIR}/src/amalgam/gen/scalar.c)
   list(APPEND wasm_srcs ${XNNPACK_DIR}/src/amalgam/gen/wasm.c)
 
-  if(onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
+  if(onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD)
+    list(APPEND wasm_srcs ${XNNPACK_DIR}/src/amalgam/gen/wasmsimd.c)
+    list(APPEND wasm_srcs ${XNNPACK_DIR}/src/amalgam/gen/wasmrelaxedsimd.c)
+    target_compile_options(XNNPACK PRIVATE "-msimd128")
+    target_compile_options(XNNPACK PRIVATE "-mrelaxed-simd")
+  elseif(onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
     list(APPEND wasm_srcs ${XNNPACK_DIR}/src/amalgam/gen/wasmsimd.c)
     target_compile_options(XNNPACK PRIVATE "-msimd128")
   endif()

cmake/onnxruntime_mlas.cmake

@@ -287,6 +287,12 @@ if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
       ${mlas_platform_srcs}
       ${MLAS_SRC_DIR}/qgemm_kernel_wasmsimd.cpp
     )
+    if (onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD)
+      set(mlas_platform_srcs
+        ${mlas_platform_srcs}
+        ${MLAS_SRC_DIR}/qgemm_kernel_wasmrelaxedsimd.cpp
+      )
+    endif()
   else()
     file(GLOB_RECURSE mlas_platform_srcs
       "${MLAS_SRC_DIR}/scalar/*.cpp"

cmake/onnxruntime_unittests.cmake

@@ -222,8 +222,11 @@ function(AddTest)
       else()
         set(TEST_NODE_FLAGS)
 
+        if (onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD)
+          message(WARNING "Use system `node` to test Wasm relaxed SIMD. Please make sure to install node v21 or newer.")
+          set(NODE_EXECUTABLE node)
         # prefer Node from emsdk so the version is more deterministic
-        if (DEFINED ENV{EMSDK_NODE})
+        elseif (DEFINED ENV{EMSDK_NODE})
           set(NODE_EXECUTABLE $ENV{EMSDK_NODE})
         else()
           message(WARNING "EMSDK_NODE environment variable was not set. Falling back to system `node`.")

cmake/onnxruntime_webassembly.cmake

@@ -485,7 +485,9 @@ jsepDownload:_pp_")
 
   list(APPEND target_name_list  "wasm")
 
-  if (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
+  if (onnxruntime_ENABLE_WEBASSEMBLY_RELAXED_SIMD)
+    list(APPEND target_name_list  "relaxedsimd")
+  elseif (onnxruntime_ENABLE_WEBASSEMBLY_SIMD)
     list(APPEND target_name_list  "simd")
   endif()
 

onnxruntime/core/mlas/inc/mlas.h

@@ -63,7 +63,10 @@ Module Name:
 #endif
 #if defined(__wasm__)
 #define MLAS_TARGET_WASM
-#if defined(__wasm_simd128__)
+#if defined(__wasm_relaxed_simd__)
+#define MLAS_TARGET_WASM_RELAXED_SIMD
+#define MLAS_TARGET_WASM_SIMD
+#elif defined(__wasm_simd128__)
 #define MLAS_TARGET_WASM_SIMD
 #else
 #define MLAS_TARGET_WASM_SCALAR

onnxruntime/core/mlas/lib/mlasi.h

@@ -996,9 +996,14 @@ extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmS8S8DispatchSdot;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8X8DispatchUmmla;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmS8S8DispatchSmmla;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8X8DispatchWasmSimd;
+extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8X8DispatchWasmRelaxedSimd;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmQuantDispatchDefault;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemm8X8DispatchPOWER10;
 
+#if defined(MLAS_TARGET_WASM_RELAXED_SIMD)
+extern bool HasUSDot();
+#endif
+
 //
 // Symmetric quantized qgemm dispatch structure
 //

onnxruntime/core/mlas/lib/qgemm.h

@@ -886,6 +886,14 @@ MlasGemmQuantGetDispatch(
     if(BIsSigned || !AIsSigned) {
         GemmQuantDispatch = &MlasGemmU8X8DispatchNeon;
     }
+#elif defined(MLAS_TARGET_WASM_RELAXED_SIMD)
+    if (!AIsSigned) {
+        if (HasUSDot()) {
+          GemmQuantDispatch = &MlasGemmU8X8DispatchWasmRelaxedSimd;
+        } else {
+          GemmQuantDispatch = &MlasGemmU8X8DispatchWasmSimd;
+        }
+    }
 #elif defined(MLAS_TARGET_WASM_SIMD)
     if (!AIsSigned) {
         GemmQuantDispatch = &MlasGemmU8X8DispatchWasmSimd;