Add LSX support for S8S8 and S8U8 GEMM kernels

- Add LSX support for S8S8/S8U8 to fix build on loong64:
  ```
  error: struct MLAS_PLATFORM’ has no member named ‘GemmS8S8Dispatch
  ```
- Add new dispatch entries for S8S8 and S8U8 GEMM operations in mlasi.h
- Extend MLAS_PLATFORM struct to include S8S8 and S8U8 dispatch pointers
- Add GemmS8S8Dispatch/GemmS8U8Dispatch to qgemm.h's AIsSigned paths

Signed-off-by: Zhou Qiankang <[email protected]>

Author: wszqkzqk

onnxruntime/core/mlas/lib/mlasi.h

@@ -982,6 +982,8 @@ struct MLAS_GEMM_QUANT_DISPATCH;
 
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8X8DispatchSse;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8X8DispatchLSX;
+extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmS8S8DispatchLSX;
+extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmS8U8DispatchLSX;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8S8DispatchSse41;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8S8DispatchAvx2;
 extern const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8U8DispatchAvx2;
@@ -1150,6 +1152,8 @@ struct MLAS_PLATFORM {
 #if defined(MLAS_TARGET_LARCH64)
     const MLAS_GEMM_QUANT_DISPATCH* GemmU8S8Dispatch;
     const MLAS_GEMM_QUANT_DISPATCH* GemmU8U8Dispatch;
+    const MLAS_GEMM_QUANT_DISPATCH* GemmS8S8Dispatch;
+    const MLAS_GEMM_QUANT_DISPATCH* GemmS8U8Dispatch;
     MLAS_GEMM_FLOAT_KERNEL* GemmFloatKernel;
     MLAS_GEMM_DOUBLE_KERNEL* GemmDoubleKernel;
     MLAS_CONV_FLOAT_KERNEL* ConvNchwFloatKernel;

onnxruntime/core/mlas/lib/platform.cpp

@@ -658,10 +658,14 @@ Return Value:
 
         this->GemmU8S8Dispatch = &MlasGemmU8X8DispatchLSX;
         this->GemmU8U8Dispatch = &MlasGemmU8X8DispatchLSX;
+        this->GemmS8S8Dispatch = &MlasGemmS8S8DispatchLSX;
+        this->GemmS8U8Dispatch = &MlasGemmS8U8DispatchLSX;
     }else if( cap_lsx ){
         this->GemmFloatKernel = MlasGemmFloatKernelLSX;
         this->GemmU8S8Dispatch = &MlasGemmU8X8DispatchLSX;
         this->GemmU8U8Dispatch = &MlasGemmU8X8DispatchLSX;
+        this->GemmS8S8Dispatch = &MlasGemmS8S8DispatchLSX;
+        this->GemmS8U8Dispatch = &MlasGemmS8U8DispatchLSX;
         this->TransposePackB16x4Routine = MlasSgemmTransposePackB16x4LSX;
         this->GemmDoubleKernel = MlasGemmDoubleKernelLSX;
         this->ConvNchwFloatKernel = MlasConvNchwFloatKernelLSX;

onnxruntime/core/mlas/lib/qgemm.h

@@ -905,7 +905,10 @@ MlasGemmQuantGetDispatch(
         GemmQuantDispatch = GetMlasPlatform().GemmU8X8Dispatch;
     }
 #elif defined(MLAS_TARGET_LARCH64)
-    if (!AIsSigned) {
+    if (AIsSigned) {
+        GemmQuantDispatch =
+            BIsSigned ? GetMlasPlatform().GemmS8S8Dispatch : GetMlasPlatform().GemmS8U8Dispatch;
+    } else { // !AIsSigned
         GemmQuantDispatch =
             BIsSigned ? GetMlasPlatform().GemmU8S8Dispatch : GetMlasPlatform().GemmU8U8Dispatch;
     }