[Feature](mlu-ops): Add mtp_617 compile options. (Cambricon#1288)

Co-authored-by: dongchengwei <dongchengwei@cambricon.com>

Author: Unireverse

CMakeLists.txt

@@ -140,7 +140,8 @@ endif()
 if (NOT MLUOP_MLU_ARCH_LIST)
   message(STATUS "build all arch")
   set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_592"
-                                           "--bang-mlu-arch=mtp_613")
+                                           "--bang-mlu-arch=mtp_613"
+                                           "--bang-mlu-arch=mtp_617")
 else()
   foreach (arch ${MLUOP_MLU_ARCH_LIST})
     set(CNCC_FLAGS_ARCH ${CNCC_FLAGS_ARCH} "--bang-mlu-arch=${arch}" )

independent_build.sh

@@ -72,6 +72,9 @@ add_mlu_arch_support () {
     --mtp613)
       bang_arch="mtp_613;"
       ;;
+    --mtp617)
+      bang_arch="mtp_617;"
+      ;;
     *)
       ;;
   esac

kernels/fft/common/fft_common_kernels.mlu

@@ -107,10 +107,12 @@ __mlu_func__ void selectVec(float *src_addr, int32_t *offset_int_addr,
                             float *dst_addr, int deal_size) {
 #if __BANG_ARCH__ >= 372 && __BANG_ARCH__ != 520
   __asm__ volatile(
-      "gather.clean.nram.nram.nram.b32.u32 "
-      "[%[dst]], [%[src]], [%[offset]], %[data_num];\n\t" ::[dst] "r"(dst_addr),
-      [src] "r"(src_addr), [offset] "r"(offset_int_addr),
-      [data_num] "r"(deal_size));
+      "gather.vector.nram.nram.nram.u32 "
+      "[%[dst]], [%[src]], [%[offset]], %[transfer_size], %[transfer_num], %[stride];\n\t"
+      ::[dst] "r"(dst_addr),
+        [src] "r"(src_addr), [offset] "r"(offset_int_addr),
+        [transfer_size] "i"(4), [transfer_num] "r"(deal_size),
+        [stride] "i"(4));
 #else
   for (auto i = 0; i < deal_size; i++) {
     dst_addr[i] = src_addr[offset_int_addr[i]];

samples/mlu-ops/abs_sample/CMakeLists.txt

@@ -46,8 +46,9 @@ endif()
 # setup cncc flags
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -fPIC -Wall -Werror -std=c++11 -pthread")
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -O3")
-set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_372"
-                                         "--bang-mlu-arch=mtp_592"
+set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_592"
+                                         "--bang-mlu-arch=mtp_613"
+                                         "--bang-mlu-arch=mtp_617"
 )
 
 # build project

samples/mlu-ops/fault_sample/CMakeLists.txt

@@ -45,8 +45,9 @@ endif()
 # setup cncc flags
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -fPIC -Wall -Werror -std=c++11 -pthread")
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -O3")
-set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_372"
-                                         "--bang-mlu-arch=mtp_592"
+set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_592"
+                                         "--bang-mlu-arch=mtp_613"
+                                         "--bang-mlu-arch=mtp_617"
 )
 
 # build project

samples/mlu-ops/poly_nms_sample/CMakeLists.txt

@@ -45,8 +45,9 @@ endif()
 # setup cncc flags
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -fPIC -Wall -Werror -std=c++11 -pthread")
 set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -O3")
-set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_372"
-                                         "--bang-mlu-arch=mtp_592"
+set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_592"
+                                         "--bang-mlu-arch=mtp_613"
+                                         "--bang-mlu-arch=mtp_617"
 )
 
 # build project