fixed for __riscv_v_intrinsic >= 12000

Author: alex-spacemit

cmake/deps.txt

@@ -22,7 +22,7 @@ dlpack;https://github.com/dmlc/dlpack/archive/refs/tags/v0.6.zip;4d565dd2e5b3132
 # it contains changes on top of 3.4.0 which are required to fix build issues.
 # Until the 3.4.1 release this is the best option we have.
 # Issue link: https://gitlab.com/libeigen/eigen/-/issues/2744
-eigen;https://gitlab.com/libeigen/eigen/-/archive/e7248b26a1ed53fa030c5c459f7ea095dfd276ac/eigen-e7248b26a1ed53fa030c5c459f7ea095dfd276ac.zip;be8be39fdbc6e60e94fa7870b280707069b5b81a
+eigen;https://gitlab.com/libeigen/eigen/-/archive/e7248b26a1ed53fa030c5c459f7ea095dfd276ac/eigen-e7248b26a1ed53fa030c5c459f7ea095dfd276ac.zip;32b145f525a8308d7ab1c09388b2e288312d8eba
 flatbuffers;https://github.com/google/flatbuffers/archive/refs/tags/v23.5.26.zip;59422c3b5e573dd192fead2834d25951f1c1670c
 fp16;https://github.com/Maratyszcza/FP16/archive/0a92994d729ff76a58f692d3028ca1b64b145d91.zip;b985f6985a05a1c03ff1bb71190f66d8f98a1494
 fxdiv;https://github.com/Maratyszcza/FXdiv/archive/63058eff77e11aa15bf531df5dd34395ec3017c8.zip;a5658f4036402dbca7cebee32be57fb8149811e1

cmake/riscv64.spacemit.toolchain.cmake

@@ -17,20 +17,12 @@ else()
     set(CMAKE_CROSSCOMPILING TRUE)
 endif()
 
-if(NOT DEFINED CMAKE_CXX_FLAGS)
-    set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba")
-endif()
-
-if(NOT DEFINED CMAKE_CXX_FLAGS)
-    set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba")
-endif()
-
 set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
 set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
-
-set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba -latomic -mabi=lp64d ${CMAKE_C_FLAGS}")
-set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba -latomic -mabi=lp64d ${CXX_FLAGS}")
+set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CMAKE_C_FLAGS}")
+set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CXX_FLAGS}")
+set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -latomic")
 
 add_definitions(-D__fp16=_Float16)

onnxruntime/core/mlas/lib/q4gemm.h

@@ -126,7 +126,7 @@ MlasQ4GemmOperation(
 
         size_t RowsRemaining = RangeCountM;
         while (RowsRemaining > 0) {
-#if defined(MLAS_TARGET_AMD64_IX86) || defined(MLAS_TARGET_POWER) || defined(MLAS_TARGET_LARCH64) || defined(MLAS_TARGET_RISCV64)
+#if defined(MLAS_TARGET_AMD64_IX86) || defined(MLAS_TARGET_POWER) || defined(MLAS_TARGET_LARCH64)
             auto RowsHandled = GetMlasPlatform().GemmFloatKernel(
                 a_row, dequant_b, c_blk, K, RowsRemaining, CountN, lda, ldc, 1.f, true);
 #else

onnxruntime/core/mlas/lib/quantize.cpp

@@ -923,7 +923,11 @@ MlasQuantizeLinear<int8_t>(
         vfloat32m4_t vec_f32 = __riscv_vle32_v_f32m4(input_ptr, vl);
         vec_f32 = __riscv_vfmul_vf_f32m4(vec_f32, div_scale, vl);
         vec_f32 = __riscv_vfadd_vf_f32m4(vec_f32, ZeroPoint, vl);
+#if __riscv_v_intrinsic <= 11000
         vint8m1_t vec_out = __riscv_vnclip_wx_i8m1(__riscv_vfncvt_x_f_w_i16m2(vec_f32, vl), 0, vl);
+#else
+        vint8m1_t vec_out = __riscv_vnclip_wx_i8m1(__riscv_vfncvt_x_f_w_i16m2(vec_f32, vl), 0, 0, vl);
+#endif
         __riscv_vse8_v_i8m1(output_ptr, vec_out, vl);
     }
 }
@@ -950,7 +954,11 @@ MlasQuantizeLinear<uint8_t>(
         vec_f32 = __riscv_vfadd_vf_f32m4(vec_f32, ZeroPoint, vl);
         vint16m2_t vec_i16 = __riscv_vfncvt_x_f_w_i16m2(vec_f32, vl);
         vec_i16 = __riscv_vmax_vx_i16m2(vec_i16, 0, vl);
+#if __riscv_v_intrinsic <= 11000
         vuint8m1_t vec_out = __riscv_vnclipu_wx_u8m1(__riscv_vreinterpret_v_i16m2_u16m2(vec_i16), 0, vl);
+#else
+        vuint8m1_t vec_out = __riscv_vnclipu_wx_u8m1(__riscv_vreinterpret_v_i16m2_u16m2(vec_i16), 0, 0, vl);
+#endif
         __riscv_vse8_v_u8m1(output_ptr, vec_out, vl);
     }
 }
@@ -2078,11 +2086,19 @@ MlasRequantizeOutputImpl(
 
             if constexpr (std::is_signed<OutputType>::value) {
                 vint16m2_t I16Vector = __riscv_vfncvt_x_f_w_i16m2(FloatVector, vl);
+#if __riscv_v_intrinsic <= 11000
                 vint8m1_t OutVector = __riscv_vnclip_wx_i8m1(I16Vector, 0, vl);
+#else
+                vint8m1_t OutVector = __riscv_vnclip_wx_i8m1(I16Vector, 0, 0, vl);
+#endif
                 __riscv_vse8_v_i8m1(RowOutput, OutVector, vl);
             } else {
                 vuint16m2_t U16Vector = __riscv_vfncvt_xu_f_w_u16m2(FloatVector, vl);
+#if __riscv_v_intrinsic <= 11000
                 vuint8m1_t OutVector = __riscv_vnclipu_wx_u8m1(U16Vector, 0, vl);
+#else
+                vuint8m1_t OutVector = __riscv_vnclipu_wx_u8m1(U16Vector, 0, 0, vl);
+#endif
                 __riscv_vse8_v_u8m1(RowOutput, OutVector, vl);
             }
         }

tools/scripts/build_riscv64.spacemit.sh

@@ -1,8 +1,9 @@
 #!/bin/bash
 # Copyright (c) 2023 SpacemiT. All rights reserved.
-# bash build_riscv64.spacemit.sh <onnxruntime_src_dir> <arch> <config>, 
+# bash build_riscv64.spacemit.sh <onnxruntime_src_dir> <arch> <config>,
 # arch is x86 or rv64, config is Release or RelWithDebInfo or Debug
 # before run this script, please ensure that RISCV_ROOT_PATH is set to the path of riscv toolchain
+# https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v1.1.2.tar.xz
 
 ARCH=$(uname -m)
 MATCH_ARCH=""