Merge branch 'master' into add_lite_interface

.github/workflows/daily.yaml

@@ -12,9 +12,9 @@ jobs:
     strategy:
       matrix:
         runner: [mlu370-m8]
-        mlu_ops_version : [1.4.0]
+        mlu_ops_version : [1.4.1]
         cntoolkit_version : [3.15.2]
-        cnnl_version: [1.27.4]
+        cnnl_version: [1.28.0]
     runs-on: ${{matrix.runner}}
     steps:
       - uses: actions/checkout@v3

.github/workflows/mluops_ci.yaml

@@ -39,7 +39,7 @@ jobs:
     strategy:
       matrix:
         runner: [mlu370-m8]
-        mlu_ops_version : [v1.4.0]
+        mlu_ops_version : [v1.4.1]
     runs-on: [yellow]
     steps:
       - uses: actions/checkout@v3

README.md

@@ -49,8 +49,9 @@ MLU-OPS™提供了以下功能：
 ## 依赖条件
 
 - 操作系统：
-  - 支持 x86_64 架构下的 Ubuntu20.04、Centos7.6、Centos8.5、Kylin10
-  - MLU-OPS™ v1.0.0版本后将不再支持 Ubuntu18.04。Ubuntu22.04系统将在后续的版本提供支持。
+  - 支持 x86_64 架构下的 Ubuntu22.04、Centos7.6、Centos8.5、Kylin10
+  - MLU-OPS™ v1.0.0版本后将不再支持 Ubuntu18.04。
+  - MLU-OPS™ v1.4.1版本后将不再支持 Ubuntu20.04。
 - 寒武纪 MLU SDK：
   - 编译和运行时依赖 CNToolkit v3.15.2 或更高版本，CNNL v1.27.4 或者更高版本
 - 寒武纪 MLU 驱动：

build.property

@@ -1,8 +1,8 @@
 {
-  "version": "1.4.0-1",
+  "version": "1.4.1-1",
   "python": "3.6.0",
   "build_requires": {"cntoolkit": ["release","3.15.2-1"],
-                     "cnnl":["release","1.27.4-1"],
+                     "cnnl":["release","1.28.0-1"],
                      "driver": "6.0.3",
                      "eigen3": "3.4.0",
                      "libxml2": "2.9.0",

core/context.cpp

@@ -29,9 +29,9 @@
 #include "core/tool.h"
 #include "kernels/kernel.h"
 
-#define DEP_CHECK_LOG(level)                                              \
+#define DEP_CHECK_LOG(level)                                                \
   mluop::logging::LogMessage(__FILE__, __LINE__, 4, level, "MLU-OPS", true, \
-                             true, true, true)                            \
+                             true, true, true)                              \
       .stream()
 
 namespace mluop {
@@ -46,27 +46,23 @@ static struct deviceName name_list_table[] = {
     // case.
 };
 
-// once cnrtGetDeviceProperties() update and not use
-// device_ordinal, update this funciton.
-mluOpDevType_t convertDeviceName(char *name) {
-  struct deviceName *pName = NULL;
-  int num = sizeof(name_list_table) / sizeof(struct deviceName);
-  if (CONTEXT_DEVICENAME_LEAST_SIZE > strlen(name)) {
-    LOG(ERROR)
-        << "get device name failed. device name too short. device name = "
-        << name << "\n";
-    return MLUOP_UNKNOWN_DEVICE;
-  }
-  for (int i = 0; i < num; i++) {
-    pName = &name_list_table[i];
-    if (0 == strncmp(pName->name, name, strlen(pName->name)) ||
-        (i == num - 1 &&
-         0 >= strncmp(pName->name, name, CONTEXT_DEVICENAME_LEAST_SIZE))) {
-      return pName->type;
+mluOpDevType_t convertDeviceNameFromInt(int device_code) {
+  switch (device_code) {
+    case 372: {
+      return MLUOP_MLU370;
+      break;
+    }
+    case 592: {
+      return MLUOP_MLU590;
+      break;
+    }
+    case 613: {
+      return MLUOP_MTP613;
+      break;
     }
+    default:
+      break;
   }
-  LOG(ERROR) << "get device name failed. return unknown device. device name = "
-             << name << "\n";
   return MLUOP_UNKNOWN_DEVICE;
 }
 }  // namespace mluop
@@ -179,6 +175,7 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
   int32_t persisting_l2cache_maxsize = 0;
   double memory_band_width = 0;
   char device_name[CONTEXT_DEVICENAME_BUFFER_SIZE] = "";
+  int device_code = 0;
   mluOpContext *ctx = new (std::nothrow) mluOpContext();
   CNcontext drv_ctx;
   CNctxConfigParam ctx_conf_param;
@@ -246,6 +243,11 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
           cnDeviceGetAttribute(&persisting_l2cache_maxsize,
                                CN_DEVICE_ATTRIBUTE_MAX_PERSISTING_L2_CACHE_SIZE,
                                mlu_dev));
+  INTERNAL_CHECK(
+      "[mluOpCreate]",
+      CN_SUCCESS == cnDeviceGetAttribute(&device_code,
+                                         CN_DEVICE_ATTRIBUTE_MLU_ISA_VERSION,
+                                         mlu_dev));
   INTERNAL_CHECK(
       "[mluOpCreate]",
       CN_SUCCESS == cnDeviceGetName(device_name, CONTEXT_DEVICENAME_BUFFER_SIZE,
@@ -266,8 +268,7 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
   }
 
   ctx->capability_job_limit = (int32_t)ctx_conf_param.unionLimit;
-  ctx->arch = mluop::convertDeviceName(
-      device_name);  // warning: possible return unknown.
+  ctx->arch = mluop::convertDeviceNameFromInt(device_code);
   ctx->sram_size = sram_size - REM_FOR_STACK;
 
   strncpy(ctx->device_name, device_name, sizeof(device_name));

core/context.h

@@ -57,6 +57,7 @@ typedef enum {
   MLUOP_MLU270 = 270,
   MLUOP_MLU370 = 372,
   MLUOP_MLU590 = 592,
+  MLUOP_MTP613 = 613,
   MLUOP_MLU290 = 290,
 } mluOpDevType_t;
 

docs/api_guide/update.rst

@@ -3,9 +3,17 @@ Update History
 
 This section lists contents that were made for each product release.
 
+* V1.4.1
+
+  **Date:** December 5, 2024
+
+  **Changes:**
+
+  - None.
+
 * V1.4.0
 
-  **Date:** October 21, 2024
+  **Date:** November 29, 2024
 
   **Changes:**
 

docs/design_docs/roi_align_rotated/roi_align_rotated_forward_vector.md

@@ -333,6 +333,8 @@ __mlu_func__ void bilinearInterpolatePosWeight(
         w3[i] += w3[j];
         w4[i] += w4[j];
         w1[j] = -1;
+      } else {
+        break;
       }
     }
     if (unique_num != i) {
@@ -386,22 +388,21 @@ bin_hw_order_num = bin_order_num ^ 2。<br>
 | pos4 | sizeof(uint) * bin_hw_order_num | pos4 坐标 |
 
 
-剩余空间对齐均分为三份 vi, vi_t, val，记空间大小为 max_v_size。<br>
-其中 vi 复用多次，最终的 val_sum 也存储于 vi 中。<br>
-此时 max_once_c = max_v_size / unique_num / sizeof(T)。 <br>
+剩余空间对齐均分为两份 val, v_t，记空间大小为 max_v_size。<br>
+此时 max_once_c = max_v_size / 4 / unique_num / sizeof(T)。 <br>
 以float 类型为例：
-- 若 bin_order_num 为 32，固定的 size 为 53376, max_vi_size 为 113280。
-unique_num 最大可到 bin_hw_order_num(1024)，此时 max_once_c = 27。
-- 若 bin_order_num 为 8，固定的 size 为 7296, max_vi_size 为 128640。
-unique_num 最大可到 bin_hw_order_num(64)，此时 max_once_c = 502。
+- 若 bin_order_num 为 32，固定的 size 为 53376, max_vi_size 为 169920。
+unique_num 最大可到 bin_hw_order_num(1024)，此时 max_once_c = 10。
+- 若 bin_order_num 为 8，固定的 size 为 7296, max_vi_size 为 192960。
+unique_num 最大可到 bin_hw_order_num(64)，此时 max_once_c = 188。
 
 
 ### 3.4 性能优化设计
 1.向量化加速。
 2.减少重复计算，例如:roi_info 计算，bin_h、bin_w 二维序列构建等。
 3.使用 fuse.nram 融合三条以上的乘加法。
 4.双线性插值坐标进行查重，减少 IO 的数量。
-
+5.将周围四个点坐标搬运成连续向量，gather时一次性处理，在有效点较少时能提升 IO 效率。
 
 ### 3.5 可维护性设计
 

docs/release_notes/mlu_ops.rst

@@ -64,6 +64,24 @@ Cambricon MLU-OPS具有以下特点：
    +-----------------------------+------------------------+--------------------------------+
 
 
+v1.4.1
+-----------------
+
+特性变更
+~~~~~~~~~~~~~~~~~~~~~
+
+- 无。
+
+已修复问题
+~~~~~~~~~~~~~~~~~~~~~
+
+- 无。
+
+已知遗留问题
+~~~~~~~~~~~~~~~~~~~~~
+
+- 无。
+
 v1.4.0
 -----------------
 

docs/user_guide/2_update_history/index.rst

@@ -1,6 +1,13 @@
 更新历史
 ========
 
+* **V1.4.1**
+  **更新时间**：2024年12月5日
+
+  **更新内容**：
+
+  -  无算子更新。
+
 * **V1.4.0**
   **更新时间**：2024年11月29日
 

independent_build.sh

@@ -52,6 +52,7 @@ long_args=(
   help
   mlu370 # mlu arch
   mlu590
+  mtp613
   no_prepare
   perf
   prepare
@@ -69,6 +70,9 @@ add_mlu_arch_support () {
     --mlu590)
       bang_arch="mtp_592;"
       ;;
+    --mtp613)
+      bang_arch="mtp_613;"
+      ;;
     *)
       ;;
   esac

installer/centos7.5/SPECS/mluops-independent.spec

@@ -5,7 +5,7 @@
 
 Name: mluops
 Summary: The Machine Lerning Unit OPerators
-Version: 1.4.0
+Version: 1.4.1
 Release: 1%{?dist}
 License: Cambricon Release License
 Vendor: Cambricon Inc.
@@ -64,6 +64,8 @@ cp $RPM_SOURCE_DIR/neuware-env.conf $RPM_BUILD_ROOT/etc/ld.so.conf.d/
 %postun -p /sbin/ldconfig
 
 %changelog
+* Thu Dec 5 2024 Cambricon Software Team <[email protected]>
+- release mluops v1.4.1
 * Thu Nov 29 2024 Cambricon Software Team <[email protected]>
 - release mluops v1.4.0
 * Mon Oct 21 2024 Cambricon Software Team <[email protected]>

installer/independent/debian/changelog

@@ -1,8 +1,14 @@
+mluops (1.4.1-1.ubuntu16.04) xenial; urgency=medium
+
+  * Release mluops v1.4.1
+
+ -- Cambricon <[email protected]>  Thu, 5 Dec 2024 00:00:00 +0100
+
 mluops (1.4.0-1.ubuntu16.04) xenial; urgency=medium
 
   * Release mluops v1.4.0
 
- -- Cambricon <[email protected]>  Thu, 29 Nov 2024 00:00:00 +0100
+ -- Cambricon <[email protected]>  Fri, 29 Nov 2024 00:00:00 +0100
 
 mluops (1.3.2-1.ubuntu16.04) xenial; urgency=medium
 

kernels/box_iou_rotated/box_iou_rotated_utils.h

@@ -24,6 +24,7 @@
 #define KERNELS_BOX_IOU_ROTATED_BOX_IOU_ROTATED_UTILS_H_
 
 #include "kernels/utils/common.h"
+#include "kernels/utils/scatter_gather.h"
 
 #define FIILED_ONES (int)0xffffffff
 #define HALF_FILLED_ONES (int16_t)0xffff
@@ -590,21 +591,22 @@ __mlu_func__ void convexHullGraham(
                       sizeof(T), actual_compute_box_num);
 
     // get the ordered points according to the angle value
-    __gather(ordered_pts_x + (i + 1) * actual_compute_box_num, intersect_pts_x,
-             (unsigned int *)temp_offset, sizeof(T), NRAM2NRAM, sizeof(T),
-             actual_compute_box_num);
-    __gather(ordered_pts_y + (i + 1) * actual_compute_box_num, intersect_pts_y,
-             (unsigned int *)temp_offset, sizeof(T), NRAM2NRAM, sizeof(T),
-             actual_compute_box_num);
-    __gather(temp_long_1 + (i + 1) * actual_compute_box_num, valid_pts,
-             (unsigned int *)temp_offset, sizeof(T), NRAM2NRAM, sizeof(T),
-             actual_compute_box_num);
+    __mluop_gather<T>(ordered_pts_x + (i + 1) * actual_compute_box_num,
+                      intersect_pts_x, (unsigned int *)temp_offset, NULL,
+                      sizeof(T), NRAM2NRAM, sizeof(T), actual_compute_box_num);
+    __mluop_gather<T>(ordered_pts_y + (i + 1) * actual_compute_box_num,
+                      intersect_pts_y, (unsigned int *)temp_offset, NULL,
+                      sizeof(T), NRAM2NRAM, sizeof(T), actual_compute_box_num);
+    __mluop_gather<T>(temp_long_1 + (i + 1) * actual_compute_box_num, valid_pts,
+                      (unsigned int *)temp_offset, NULL, sizeof(T), NRAM2NRAM,
+                      sizeof(T), actual_compute_box_num);
 
     // assign a invalid value to the point which has been get ordered
-    __scatter(temp_long_2, temp1_ram, (unsigned int *)temp_offset, sizeof(T),
-              NRAM2NRAM, sizeof(T), actual_compute_box_num);
-    __scatter(valid_pts, temp2_ram, (unsigned int *)temp_offset, sizeof(T),
-              NRAM2NRAM, sizeof(T), actual_compute_box_num);
+    __mluop_scatter<T>(temp_long_2, temp1_ram, (unsigned int *)temp_offset,
+                       NULL, sizeof(T), NRAM2NRAM, sizeof(T),
+                       actual_compute_box_num);
+    __mluop_scatter<T>(valid_pts, temp2_ram, (unsigned int *)temp_offset, NULL,
+                       sizeof(T), NRAM2NRAM, sizeof(T), actual_compute_box_num);
   }
   __bang_move(valid_pts, temp_long_1, total_points * sizeof(T));
 #else