{"id":96004,"date":"2026-08-27T10:16:29","date_gmt":"2026-08-27T02:16:29","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/96004.html"},"modified":"2026-08-27T10:16:29","modified_gmt":"2026-08-27T02:16:29","slug":"%e8%bd%bb%e6%9d%be%e5%ad%a6%e4%b9%a0tflm_day8","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/96004.html","title":{"rendered":"\u8f7b\u677e\u5b66\u4e60TFLM_day8"},"content":{"rendered":"

\u6458\u8981：\u672c\u6587\u628a\u524d 7 \u5929\u5b66\u5230\u7684 TFLM \u77e5\u8bc6\u6574\u7406\u6210\u4e00\u4efd\u53ef\u6267\u884c\u7684\u5de5\u7a0b\u8ba1\u5212，\u6307\u5bfc\u5982\u4f55\u8ba9\u4e00\u9897\u65b0\u7684 MCU\u3001DSP\u3001NPU \u6216 AI accelerator \u7a33\u5b9a\u8fd0\u884c TFLM \u5e76\u9010\u6b65\u83b7\u5f97\u53ef\u91cf\u5316\u7684\u6027\u80fd\u6536\u76ca\u3002\u6838\u5fc3\u539f\u5219\u662f\u300c\u5148\u8dd1\u901a reference，\u518d\u63a5\u5165\u4f18\u5316，\u6700\u540e\u7528\u6a21\u578b\u548c benchmark \u8bc1\u660e\u6536\u76ca\u300d\u3002\u6587\u7ae0\u6309 8 \u4e2a\u9636\u6bb5\u5c55\u5f00：\u4ece\u9700\u6c42\u4e0e baseline\u3001\u5e73\u53f0 bring-up\u3001reference \u6a21\u578b\u95ed\u73af\u3001NN library \u8fb9\u754c，\u5230\u7b2c\u4e00\u4e2a optimized kernel\u3001\u6269\u5c55\u70ed\u70b9 op\u3001\u5185\u5b58\u4e0e\u7a33\u5b9a\u6027，\u518d\u5230 CI \u4e0e\u7248\u672c\u7ef4\u62a4，\u5e76\u4e3a\u6bcf\u4e2a\u9636\u6bb5\u7ed9\u51fa\u76ee\u6807\u3001\u5de5\u4f5c\u9879\u3001\u4ea4\u4ed8\u7269\u548c\u5b8c\u6210\u6761\u4ef6，\u6700\u540e\u8865\u5145\u89d2\u8272\u5206\u5de5\u3001\u98ce\u9669\u5e94\u5bf9\u3001\u9a8c\u6536\u6807\u51c6\u4e0e\u6700\u5c0f\u53ef\u6267\u884c\u8def\u7ebf\u56fe\u3002<\/p>\n

\u8f7b\u677e\u5b66\u4e60 TFLM Day 8：AI \u82af\u7247\u63a5\u5165 TFLM \u7684\u5de5\u7a0b\u8ba1\u5212<\/h2>\n

Day 1 \u5230 Day 7，\u6211\u4eec\u4f9d\u6b21\u5b66\u4e60\u4e86 TFLM \u7684\u8fd0\u884c\u65f6\u3001\u8c03\u7528\u94fe\u3001\u5185\u5b58\u3001kernel\u3001\u540e\u7aef\u76ee\u5f55\u3001\u6a21\u578b op \u5206\u6790，\u4ee5\u53ca\u628a .tflite \u6a21\u578b\u63a5\u5165\u793a\u4f8b\u3002<\/p>\n

\u4eca\u5929\u628a\u8fd9\u4e9b\u77e5\u8bc6\u6574\u7406\u6210\u4e00\u4efd\u53ef\u4ee5\u6267\u884c\u7684\u5de5\u7a0b\u8ba1\u5212：<\/p>\n

\u5982\u4f55\u8ba9\u4e00\u9897\u65b0\u7684 MCU\u3001DSP\u3001NPU \u6216 AI accelerator \u7a33\u5b9a\u5730\u8fd0\u884c TFLM，\u5e76\u9010\u6b65\u83b7\u5f97\u53ef\u91cf\u5316\u7684\u6027\u80fd\u6536\u76ca？<\/p>\n

\u6838\u5fc3\u539f\u5219\u53ea\u6709\u4e00\u53e5\u8bdd：<\/p>\n

\u5148\u8dd1\u901a reference，\u518d\u63a5\u5165\u4f18\u5316，\u6700\u540e\u7528\u6a21\u578b\u548c benchmark \u8bc1\u660e\u6536\u76ca\u3002<\/p>\n

1. \u9879\u76ee\u76ee\u6807\u548c\u8fb9\u754c<\/h3>\n

1.1 \u6700\u7ec8\u76ee\u6807<\/h4>\n

\u5b8c\u6210\u540e，\u76ee\u6807\u5e73\u53f0\u5e94\u80fd\u591f：<\/p>\n

  • \u7f16\u8bd1 TFLM \u9759\u6001\u5e93\u548c\u76ee\u6807\u5e73\u53f0\u5e94\u7528\u3002<\/li>\n
  • \u52a0\u8f7d\u81f3\u5c11\u4e00\u4e2a .tflite \u6a21\u578b\u3002<\/li>\n
  • \u4f7f\u7528\u9759\u6001 tensor_arena \u5b8c\u6210 AllocateTensors()\u3002<\/li>\n
  • \u8c03\u7528 Invoke() \u83b7\u5f97\u6b63\u786e\u8f93\u51fa\u3002<\/li>\n
  • \u901a\u8fc7 UART\u3001RTT \u6216\u5e73\u53f0\u65e5\u5fd7\u89c2\u5bdf\u8fd0\u884c\u72b6\u6001\u3002<\/li>\n
  • \u6d4b\u91cf\u5355\u7b97\u5b50\u548c\u6574\u6a21\u578b latency\u3002<\/li>\n
  • \u6309\u9700\u5c06\u70ed\u70b9 op \u66ff\u6362\u4e3a\u82af\u7247\u4f18\u5316\u5b9e\u73b0\u3002<\/li>\n
  • \u5728 reference \u4e0e optimized \u4e4b\u95f4\u8fdb\u884c\u6b63\u786e\u6027\u548c\u6027\u80fd\u5bf9\u6bd4\u3002<\/li>\n

    1.2 \u4e0d\u5c5e\u4e8e\u7b2c\u4e00\u9636\u6bb5\u7684\u5185\u5bb9<\/h4>\n

    \u4ee5\u4e0b\u5185\u5bb9\u4e0d\u8981\u4f5c\u4e3a\u9879\u76ee\u7b2c\u4e00\u6b65：<\/p>\n

      \n
    • \u4e00\u5f00\u59cb\u5c31\u652f\u6301\u6240\u6709 TFLM op\u3002<\/li>\n
    • \u4e00\u5f00\u59cb\u5c31\u5b9e\u73b0\u5b8c\u6574 graph compiler\u3002<\/li>\n
    • \u4e00\u5f00\u59cb\u5c31\u91cd\u5199 TFLM reference kernel\u3002<\/li>\n
    • \u6ca1\u6709 baseline \u5c31\u58f0\u79f0\u786c\u4ef6\u52a0\u901f\u6709\u6548\u3002<\/li>\n
    • \u53ea\u6d4b\u8bd5\u4e00\u4e2a\u8f93\u5165\u5c31\u5ba3\u5e03\u6a21\u578b\u517c\u5bb9\u3002<\/li>\n<\/ul>\n

      TFLM \u7684 reference kernel \u662f\u6b63\u786e\u6027\u57fa\u7ebf\u3002\u82af\u7247\u4f18\u5316\u5e94\u653e\u5728\u76ee\u6807\u5e73\u53f0 kernel \u5b50\u76ee\u5f55\u4e2d，\u800c\u4e0d\u662f\u7834\u574f\u5171\u4eab\u7684 reference \u5b9e\u73b0\u3002<\/p>\n

      2. \u63a8\u8350\u7684\u603b\u4f53\u67b6\u6784<\/h3>\n

      \u7528\u6237\u5e94\u7528
      \n |
      \n v
      \nTFLM MicroInterpreter
      \n |
      \n v
      \nMicroMutableOpResolver
      \n |
      \n v
      \nyour_chip optimized kernels
      \n |
      \n v
      \nyour_chip NN library \/ driver
      \n |
      \n v
      \nNPU \/ DSP \/ AI accelerator<\/p>\n

      \u5efa\u8bae\u628a\u5de5\u7a0b\u62c6\u6210\u4e09\u4e2a\u72ec\u7acb\u6a21\u5757：<\/p>\n\n\u6a21\u5757\u4e3b\u8981\u804c\u8d23<\/tr>\n\n\n\n\n
      \u5e73\u53f0\u9002\u914d<\/td>\n\u5de5\u5177\u94fe\u3001\u542f\u52a8\u3001\u65e5\u5fd7\u3001\u8ba1\u65f6\u3001\u7cfb\u7edf\u521d\u59cb\u5316\u3001\u94fe\u63a5\u811a\u672c\u3002<\/td>\n<\/tr>\n
      TFLM \u540e\u7aef<\/td>\nconv.cc\u3001fully_connected.cc \u7b49\u8584 wrapper\u3002<\/td>\n<\/tr>\n
      \u82af\u7247\u8f6f\u4ef6\u6808<\/td>\nNN library\u3001driver\u3001firmware\u3001DMA\u3001cache \u548c\u786c\u4ef6\u547d\u4ee4\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      NN library \u4e0d\u5fc5\u5b8c\u5168\u9075\u5faa TFLM \u7684\u4ee3\u7801\u98ce\u683c，\u4e5f\u53ef\u4ee5\u72ec\u7acb\u53d1\u5e03\u548c\u6d4b\u8bd5\u3002TFLM wrapper \u53ea\u8d1f\u8d23 tensor\u3001shape\u3001\u91cf\u5316\u53c2\u6570\u548c\u72b6\u6001\u7801\u4e4b\u95f4\u7684\u8f6c\u6362\u3002<\/p>\n

      3. \u9879\u76ee\u9636\u6bb5\u603b\u89c8<\/h3>\n

      \u9636\u6bb5 0 \u9700\u6c42\u548c\u57fa\u7ebf
      \n |
      \n\u9636\u6bb5 1 \u5e73\u53f0 bring-up
      \n |
      \n\u9636\u6bb5 2 reference \u6a21\u578b\u95ed\u73af
      \n |
      \n\u9636\u6bb5 3 \u82af\u7247 NN library \u8fb9\u754c
      \n |
      \n\u9636\u6bb5 4 \u7b2c\u4e00\u4e2a\u4f18\u5316 kernel
      \n |
      \n\u9636\u6bb5 5 \u6269\u5c55\u70ed\u70b9 op
      \n |
      \n\u9636\u6bb5 6 \u6027\u80fd\u3001\u5185\u5b58\u548c\u7a33\u5b9a\u6027
      \n |
      \n\u9636\u6bb5 7 CI\u3001\u53d1\u5e03\u548c\u7248\u672c\u7ef4\u62a4<\/p>\n

      \u6bcf\u4e2a\u9636\u6bb5\u90fd\u5fc5\u987b\u6709\u201c\u5b8c\u6210\u6761\u4ef6\u201d\u3002\u6ca1\u6709\u901a\u8fc7\u5f53\u524d\u9636\u6bb5\u7684\u9a8c\u8bc1，\u5c31\u4e0d\u8981\u628a\u95ee\u9898\u5e26\u5230\u4e0b\u4e00\u9636\u6bb5\u3002<\/p>\n

      4. \u9636\u6bb5 0：\u9700\u6c42\u3001\u6a21\u578b\u548c baseline<\/h3>\n

      \u76ee\u6807<\/h4>\n

      \u786e\u5b9a\u82af\u7247\u3001\u8f6f\u4ef6\u73af\u5883\u3001\u76ee\u6807\u6a21\u578b\u548c\u53ef\u590d\u73b0\u7684 reference \u6570\u636e\u3002<\/p>\n

      \u5de5\u4f5c\u9879<\/h4>\n
    • \u786e\u5b9a\u76ee\u6807\u82af\u7247\u578b\u53f7\u3001\u6838\u5fc3\u3001\u65f6\u949f\u548c\u5185\u5b58\u5e03\u5c40\u3002<\/li>\n
    • \u786e\u8ba4 C++17 \u5de5\u5177\u94fe\u3001SDK\u3001IDE \u548c\u94fe\u63a5\u5668\u7248\u672c\u3002<\/li>\n
    • \u9009\u62e9\u4e00\u5230\u4e09\u4e2a\u4ee3\u8868\u6027 .tflite \u6a21\u578b\u3002<\/li>\n
    • \u7528 Day 6 \u7684\u5de5\u5177\u627e\u51fa\u6bcf\u4e2a\u6a21\u578b\u7684\u552f\u4e00 op\u3002<\/li>\n
    • \u8bb0\u5f55\u6a21\u578b\u8f93\u5165\u3001\u8f93\u51fa\u3001\u7c7b\u578b\u3001shape \u548c\u91cf\u5316\u53c2\u6570\u3002<\/li>\n
    • \u5728 host \u4e0a\u4fdd\u5b58 reference \u8f93\u51fa\u548c\u6d4b\u8bd5\u8f93\u5165\u3002<\/li>\n
    • \u5b9a\u4e49 latency\u3001arena\u3001\u4ee3\u7801\u4f53\u79ef\u548c\u529f\u8017\u7684\u6d4b\u91cf\u65b9\u6cd5\u3002<\/li>\n

      \u4ea4\u4ed8\u7269<\/h4>\n

      requirements.md
      \nmodels\/
      \n model_a.tflite
      \n model_b.tflite
      \nbaseline\/
      \n input_data.*
      \n reference_output.*
      \nbenchmark_definition.md<\/p>\n

      \u5b8c\u6210\u6761\u4ef6<\/h4>\n

      [ ] \u6bcf\u4e2a\u76ee\u6807\u6a21\u578b\u90fd\u6709\u660e\u786e\u8f93\u5165\u548c\u8f93\u51fa
      \n[ ] \u6bcf\u4e2a\u6a21\u578b\u7684 op \u96c6\u5408\u5df2\u7ecf\u786e\u8ba4
      \n[ ] host reference \u8f93\u51fa\u53ef\u4ee5\u91cd\u590d\u5f97\u5230
      \n[ ] latency \u548c\u5185\u5b58\u7684\u6d4b\u91cf\u65b9\u6cd5\u5df2\u7ecf\u786e\u5b9a<\/p>\n

      \u4e0d\u8981\u53ea\u9009\u62e9\u4e00\u4e2a\u201c\u6700\u5bb9\u6613\u8dd1\u201d\u7684\u6a21\u578b\u3002\u81f3\u5c11\u5e94\u5305\u542b\u4e00\u4e2a\u5c0f\u6a21\u578b\u7528\u4e8e bring-up，\u4ee5\u53ca\u4e00\u4e2a\u80fd\u4ee3\u8868\u5b9e\u9645\u4e1a\u52a1\u8d1f\u8f7d\u7684\u6a21\u578b\u7528\u4e8e\u6027\u80fd\u8bc4\u4f30\u3002<\/p>\n

      5. \u9636\u6bb5 1：\u5e73\u53f0 bring-up<\/h3>\n

      TFLM \u5b98\u65b9\u5efa\u8bae\u5148\u4e0d\u63a5\u786c\u4ef6\u4f18\u5316，\u4f7f\u7528 reference kernel \u5efa\u7acb\u5e73\u53f0\u8fd0\u884c\u73af\u5883\u3002<\/p>\n

      \u76ee\u6807<\/h4>\n

      \u8ba9 TFLM \u5728\u76ee\u6807\u82af\u7247\u4e0a\u5b8c\u6210\u6700\u5c0f\u793a\u4f8b\u8fd0\u884c\u3002<\/p>\n

      \u4e3b\u8981\u5de5\u4f5c<\/h4>\n

      5.1 \u5de5\u5177\u94fe\u548c\u94fe\u63a5<\/h5>\n

      \u786e\u8ba4：<\/p>\n

        \n
      • \u7f16\u8bd1\u5668\u652f\u6301 C++17\u3002<\/li>\n
      • include \u8def\u5f84\u548c\u5b8f\u5b9a\u4e49\u6b63\u786e\u3002<\/li>\n
      • \u94fe\u63a5\u811a\u672c\u80fd\u653e\u4e0b\u4ee3\u7801\u3001\u53ea\u8bfb\u6a21\u578b\u548c tensor arena\u3002<\/li>\n
      • C\/C++ ABI \u548c\u6d6e\u70b9 ABI \u8bbe\u7f6e\u4e00\u81f4\u3002<\/li>\n
      • \u6808\u3001\u5168\u5c40\u533a\u3001\u5806\u7b56\u7565\u7b26\u5408\u82af\u7247 SDK\u3002<\/li>\n<\/ul>\n
        5.2 \u5e73\u53f0\u6587\u4ef6<\/h5>\n

        \u51c6\u5907\u76ee\u6807\u5e73\u53f0\u7248\u672c\u7684：<\/p>\n

        tensorflow\/lite\/micro\/debug_log.cc
        \ntensorflow\/lite\/micro\/micro_time.cc
        \ntensorflow\/lite\/micro\/system_setup.cc<\/p>\n

        \u5b83\u4eec\u5206\u522b\u8d1f\u8d23\u65e5\u5fd7\u8f93\u51fa\u3001\u65f6\u95f4\u6d4b\u91cf\u548c\u7cfb\u7edf\u521d\u59cb\u5316\u3002\u5b9e\u73b0\u53ef\u4ee5\u653e\u5728\u5e73\u53f0\u81ea\u5df1\u7684\u76ee\u5f55，\u53ea\u8981\u6700\u7ec8\u94fe\u63a5\u65f6\u80fd\u63d0\u4f9b\u5bf9\u5e94\u63a5\u53e3\u3002<\/p>\n

        5.3 \u9879\u76ee\u88c1\u526a<\/h5>\n

        \u53ef\u4ee5\u4f7f\u7528\u9879\u76ee\u751f\u6210\u811a\u672c\u521b\u5efa\u53ea\u5305\u542b\u6240\u9700\u793a\u4f8b\u7684\u6e90\u7801\u6811：<\/p>\n

        python3 tensorflow\/lite\/micro\/tools\/project_generation\/create_tflm_tree.py \\\\<\/span>
        \n -e<\/span> hello_world \\\\<\/span>
        \n -e<\/span> person_detection \\\\<\/span>
        \n \/tmp\/tflm-tree<\/p>\n

        \u7136\u540e\u7528\u82af\u7247\u81ea\u5df1\u7684\u6784\u5efa\u7cfb\u7edf\u7f16\u8bd1 TFLM \u9759\u6001\u5e93，\u4f8b\u5982：<\/p>\n

        libtensorflow-microlite.a<\/p>\n

        \u4ea4\u4ed8\u7269<\/h4>\n

        tensorflow\/lite\/micro\/your_chip\/
        \n README.md
        \n debug_log.cc
        \n micro_time.cc
        \n system_setup.cc
        \nplatform_build\/
        \n toolchain file
        \n linker script
        \n startup code<\/p>\n

        \u5b8c\u6210\u6761\u4ef6<\/h4>\n

        [ ] \u7f16\u8bd1\u901a\u8fc7
        \n[ ] \u65e5\u5fd7\u53ef\u4ee5\u8f93\u51fa
        \n[ ] \u8ba1\u65f6\u63a5\u53e3\u8fd4\u56de\u6709\u6548\u503c
        \n[ ] \u7cfb\u7edf\u521d\u59cb\u5316\u5b8c\u6210
        \n[ ] \u76ee\u6807\u677f\u80fd\u8fd0\u884c\u6700\u5c0f\u7a0b\u5e8f<\/p>\n

        6. \u9636\u6bb5 2：reference \u6a21\u578b\u95ed\u73af<\/h3>\n

        \u76ee\u6807<\/h4>\n

        \u5728\u4e0d\u4f7f\u7528\u82af\u7247 optimized kernel \u7684\u60c5\u51b5\u4e0b，\u8ba9\u4e00\u4e2a\u771f\u5b9e\u6a21\u578b\u5b8c\u6210\u6b63\u786e\u63a8\u7406\u3002<\/p>\n

        \u63a8\u8350\u987a\u5e8f<\/h4>\n

        hello_world
        \n -> \u4e00\u4e2a\u5c0f\u578b int8 \u6a21\u578b
        \n -> person_detection \u6216 micro_speech
        \n -> \u5b9e\u9645\u4e1a\u52a1\u6a21\u578b<\/p>\n

        \u5de5\u4f5c\u9879<\/h4>\n
      • \u628a .tflite \u8f6c\u6210 C \u6570\u7ec4，\u6216\u7531\u6784\u5efa\u7cfb\u7edf\u751f\u6210\u6570\u7ec4\u3002<\/li>\n
      • \u5c06\u6a21\u578b .cc \u548c .h \u52a0\u5165\u76ee\u6807\u5de5\u7a0b\u3002<\/li>\n
      • \u6839\u636e\u6a21\u578b op \u521b\u5efa\u6700\u5c0f resolver\u3002<\/li>\n
      • \u5206\u914d\u9759\u6001 tensor arena\u3002<\/li>\n
      • \u68c0\u67e5\u6a21\u578b schema \u7248\u672c\u3002<\/li>\n
      • \u68c0\u67e5\u8f93\u5165 tensor \u7684 type\u3001shape \u548c bytes\u3002<\/li>\n
      • \u8fdb\u884c\u8f93\u5165\u9884\u5904\u7406\u548c\u91cf\u5316\u3002<\/li>\n
      • \u8c03\u7528 AllocateTensors()\u3002<\/li>\n
      • \u8c03\u7528 Invoke()\u3002<\/li>\n
      • \u5bf9\u8f93\u51fa\u8fdb\u884c\u53cd\u91cf\u5316\u5e76\u548c host reference \u6bd4\u8f83\u3002<\/li>\n

        \u5178\u578b\u4ee3\u7801\u7ed3\u6784<\/h4>\n

        const<\/span> tflite::<\/span>Model*<\/span> model =<\/span> tflite::<\/span>GetModel<\/span>(<\/span>g_model_data)<\/span>;<\/span>
        \nTFLITE_CHECK_EQ<\/span>(<\/span>model-><\/span>version<\/span>(<\/span>)<\/span>,<\/span> TFLITE_SCHEMA_VERSION)<\/span>;<\/span><\/p>\n

        using<\/span> Resolver =<\/span> tflite::<\/span>MicroMutableOpResolver<<\/span>3<\/span>><\/span>;<\/span>
        \nResolver resolver;<\/span>
        \nresolver.<\/span>AddConv2D<\/span>(<\/span>)<\/span>;<\/span>
        \nresolver.<\/span>AddFullyConnected<\/span>(<\/span>)<\/span>;<\/span>
        \nresolver.<\/span>AddSoftmax<\/span>(<\/span>)<\/span>;<\/span><\/p>\n

        uint8_t<\/span> tensor_arena[<\/span>128<\/span> *<\/span> 1024<\/span>]<\/span>;<\/span>
        \ntflite::<\/span>MicroInterpreter interpreter<\/span>(<\/span>
        \n model,<\/span> resolver,<\/span> tensor_arena,<\/span> sizeof<\/span>(<\/span>tensor_arena)<\/span>)<\/span>;<\/span><\/p>\n

        TF_LITE_ENSURE_STATUS<\/span>(<\/span>interpreter.<\/span>AllocateTensors<\/span>(<\/span>)<\/span>)<\/span>;<\/span>
        \nTF_LITE_ENSURE_STATUS<\/span>(<\/span>interpreter.<\/span>Invoke<\/span>(<\/span>)<\/span>)<\/span>;<\/span><\/p>\n

        \u4e0b\u9762\u662f\u4e00\u4e2a\u5b8c\u6574\u7684 main.cc \u793a\u4f8b，\u5c55\u793a\u4ece\u6a21\u578b\u6570\u636e\u52a0\u8f7d\u3001\u8f93\u5165\u9884\u5904\u7406\u3001\u8c03\u7528 AllocateTensors \u548c Invoke，\u5230\u8f93\u51fa\u53cd\u91cf\u5316\u5e76\u4e0e host reference \u6bd4\u8f83\u7684\u5b8c\u6574\u6d41\u7a0b：<\/p>\n

        \/\/ examples\/your_chip_model\/main.cc<\/span>
        \n#<\/span>include<\/span> "tensorflow\/lite\/micro\/micro_interpreter.h"<\/span><\/span>
        \n#<\/span>include<\/span> "tensorflow\/lite\/micro\/micro_mutable_op_resolver.h"<\/span><\/span>
        \n#<\/span>include<\/span> "tensorflow\/lite\/micro\/system_setup.h"<\/span><\/span>
        \n#<\/span>include<\/span> "tensorflow\/lite\/schema\/schema_generated.h"<\/span><\/span><\/p>\n

        #<\/span>include<\/span> "model_data.h"<\/span> \/\/ \u7531 .tflite \u8f6c\u6362\u751f\u6210\u7684\u6a21\u578b\u6570\u7ec4<\/span><\/span>
        \n#<\/span>include<\/span> "input_adapter.h"<\/span> \/\/ \u8f93\u5165\u9884\u5904\u7406：\u539f\u59cb\u6570\u636e -> int8 \u91cf\u5316\u8f93\u5165<\/span><\/span>
        \n#<\/span>include<\/span> "output_adapter.h"<\/span> \/\/ \u8f93\u51fa\u89e3\u91ca：int8 \u8f93\u51fa -> \u53cd\u91cf\u5316\u6d6e\u70b9\u503c<\/span><\/span><\/p>\n

        \/\/ \u9759\u6001 tensor arena：\u5927\u5c0f\u9700\u6839\u636e\u6a21\u578b\u5b9e\u9645\u9700\u6c42\u8c03\u6574<\/span>
        \nconstexpr<\/span> int<\/span> kTensorArenaSize =<\/span> 128<\/span> *<\/span> 1024<\/span>;<\/span>
        \nuint8_t<\/span> tensor_arena[<\/span>kTensorArenaSize]<\/span>;<\/span><\/p>\n

        \/\/ \u4ece host \u7aef\u4fdd\u5b58\u7684 reference \u8f93\u51fa（\u6d6e\u70b9）\u4e2d\u8bfb\u53d6\u671f\u671b\u503c<\/span>
        \nextern<\/span> const<\/span> float<\/span> kReferenceOutput[<\/span>]<\/span>;<\/span> \/\/ \u5b9a\u4e49\u5728 baseline\/reference_output.cc<\/span>
        \nextern<\/span> const<\/span> int<\/span> kReferenceOutputSize;<\/span><\/p>\n

        \/\/ \u53cd\u91cf\u5316：int8 \u539f\u59cb\u503c -> \u6d6e\u70b9\u771f\u5b9e\u503c<\/span>
        \nfloat<\/span> Dequantize<\/span>(<\/span>int8_t<\/span> value,<\/span> float<\/span> scale,<\/span> int32_t<\/span> zero_point)<\/span> {<\/span>
        \n return<\/span> (<\/span>static_cast<\/span><<\/span>float<\/span>><\/span><\/span><\/span>(<\/span>value)<\/span> –<\/span> zero_point)<\/span> *<\/span> scale;<\/span>
        \n}<\/span><\/p>\n

        \/\/ \u6bd4\u8f83\u76ee\u6807\u677f\u8f93\u51fa\u4e0e host reference \u8f93\u51fa<\/span>
        \nbool<\/span> CompareWithReference<\/span>(<\/span>const<\/span> float<\/span>*<\/span> actual,<\/span> int<\/span> size,<\/span>
        \n float<\/span> tolerance =<\/span> 0.01f<\/span>)<\/span> {<\/span>
        \n for<\/span> (<\/span>int<\/span> i =<\/span> 0<\/span>;<\/span> i <<\/span> size;<\/span> ++<\/span>i)<\/span> {<\/span>
        \n float<\/span> diff =<\/span> actual[<\/span>i]<\/span> –<\/span> kReferenceOutput[<\/span>i]<\/span>;<\/span>
        \n if<\/span> (<\/span>diff <<\/span> –<\/span>tolerance ||<\/span> diff ><\/span> tolerance)<\/span> {<\/span>
        \n MicroPrintf<\/span>(<\/span>"Mismatch at [%d]: actual=%f ref=%f"<\/span>,<\/span> i,<\/span> actual[<\/span>i]<\/span>,<\/span>
        \n kReferenceOutput[<\/span>i]<\/span>)<\/span>;<\/span>
        \n return<\/span> false<\/span>;<\/span>
        \n }<\/span>
        \n }<\/span>
        \n return<\/span> true<\/span>;<\/span>
        \n}<\/span><\/p>\n

        int<\/span> main<\/span>(<\/span>)<\/span> {<\/span>
        \n \/\/ 1. \u5e73\u53f0\u521d\u59cb\u5316：\u65e5\u5fd7\u3001\u8ba1\u65f6\u3001\u7cfb\u7edf\u65f6\u949f\u7b49<\/span>
        \n tflite::<\/span>InitializeTarget<\/span>(<\/span>)<\/span>;<\/span><\/p>\n

        \/\/ 2. \u52a0\u8f7d\u6a21\u578b：\u4ece C \u6570\u7ec4\u89e3\u6790 FlatBuffer<\/span>
        \n const<\/span> tflite::<\/span>Model*<\/span> model =<\/span> tflite::<\/span>GetModel<\/span>(<\/span>g_model_data)<\/span>;<\/span>
        \n TFLITE_CHECK_EQ<\/span>(<\/span>model-><\/span>version<\/span>(<\/span>)<\/span>,<\/span> TFLITE_SCHEMA_VERSION)<\/span>;<\/span><\/p>\n

        \/\/ 3. \u521b\u5efa\u6700\u5c0f resolver，\u53ea\u6ce8\u518c\u6a21\u578b\u5b9e\u9645\u7528\u5230\u7684 op<\/span>
        \n using<\/span> Resolver =<\/span> tflite::<\/span>MicroMutableOpResolver<<\/span>3<\/span>><\/span>;<\/span>
        \n Resolver resolver;<\/span>
        \n resolver.<\/span>AddConv2D<\/span>(<\/span>)<\/span>;<\/span>
        \n resolver.<\/span>AddFullyConnected<\/span>(<\/span>)<\/span>;<\/span>
        \n resolver.<\/span>AddSoftmax<\/span>(<\/span>)<\/span>;<\/span><\/p>\n

        \/\/ 4. \u6784\u5efa\u89e3\u91ca\u5668，\u7ed1\u5b9a\u9759\u6001 arena<\/span>
        \n tflite::<\/span>MicroInterpreter interpreter<\/span>(<\/span>model,<\/span> resolver,<\/span> tensor_arena,<\/span>
        \n kTensorArenaSize)<\/span>;<\/span><\/p>\n

        \/\/ 5. \u5206\u914d tensor \u5185\u5b58；\u5931\u8d25\u901a\u5e38\u610f\u5473\u7740 arena \u592a\u5c0f<\/span>
        \n TfLiteStatus allocate_status =<\/span> interpreter.<\/span>AllocateTensors<\/span>(<\/span>)<\/span>;<\/span>
        \n if<\/span> (<\/span>allocate_status !=<\/span> kTfLiteOk)<\/span> {<\/span>
        \n MicroPrintf<\/span>(<\/span>"AllocateTensors failed: arena too small?"<\/span>)<\/span>;<\/span>
        \n return<\/span> –<\/span>1<\/span>;<\/span>
        \n }<\/span><\/p>\n

        \/\/ 6. \u83b7\u53d6\u8f93\u5165 tensor，\u8fdb\u884c\u9884\u5904\u7406\u548c\u91cf\u5316<\/span>
        \n TfLiteTensor*<\/span> input =<\/span> interpreter.<\/span>input<\/span>(<\/span>0<\/span>)<\/span>;<\/span>
        \n TFLITE_CHECK_EQ<\/span>(<\/span>input-><\/span>type,<\/span> kTfLiteInt8)<\/span>;<\/span><\/p>\n

        \/\/ \u4ece input_adapter \u8bfb\u53d6\u539f\u59cb\u6570\u636e\u5e76\u91cf\u5316\u4e3a int8<\/span>
        \n const<\/span> float<\/span>*<\/span> raw_input =<\/span> GetRawInput<\/span>(<\/span>)<\/span>;<\/span> \/\/ \u539f\u59cb\u6d6e\u70b9\u8f93\u5165<\/span>
        \n const<\/span> float<\/span> input_scale =<\/span> input-><\/span>params.<\/span>scale;<\/span>
        \n const<\/span> int32_t<\/span> input_zero_point =<\/span> input-><\/span>params.<\/span>zero_point;<\/span><\/p>\n

        for<\/span> (<\/span>int<\/span> i =<\/span> 0<\/span>;<\/span> i <<\/span> input-><\/span>dims-><\/span>data[<\/span>1<\/span>]<\/span> *<\/span> input-><\/span>dims-><\/span>data[<\/span>2<\/span>]<\/span> *<\/span>
        \n input-><\/span>dims-><\/span>data[<\/span>3<\/span>]<\/span>;<\/span>
        \n ++<\/span>i)<\/span> {<\/span>
        \n \/\/ \u91cf\u5316\u516c\u5f0f：q = round(raw \/ scale) + zero_point<\/span>
        \n float<\/span> scaled =<\/span> raw_input[<\/span>i]<\/span> \/<\/span> input_scale;<\/span>
        \n int32_t<\/span> quantized =<\/span> static_cast<\/span><<\/span>int32_t<\/span>><\/span><\/span><\/span>(<\/span>scaled +<\/span> input_zero_point)<\/span>;<\/span>
        \n input-><\/span>data.<\/span>int8[<\/span>i]<\/span> =<\/span> static_cast<\/span><<\/span>int8_t<\/span>><\/span><\/span><\/span>(<\/span>quantized)<\/span>;<\/span>
        \n }<\/span><\/p>\n

        \/\/ 7. \u6267\u884c\u63a8\u7406<\/span>
        \n TfLiteStatus invoke_status =<\/span> interpreter.<\/span>Invoke<\/span>(<\/span>)<\/span>;<\/span>
        \n if<\/span> (<\/span>invoke_status !=<\/span> kTfLiteOk)<\/span> {<\/span>
        \n MicroPrintf<\/span>(<\/span>"Invoke failed"<\/span>)<\/span>;<\/span>
        \n return<\/span> –<\/span>1<\/span>;<\/span>
        \n }<\/span><\/p>\n

        \/\/ 8. \u83b7\u53d6\u8f93\u51fa tensor，\u53cd\u91cf\u5316\u5e76\u4e0e host reference \u6bd4\u8f83<\/span>
        \n TfLiteTensor*<\/span> output =<\/span> interpreter.<\/span>output<\/span>(<\/span>0<\/span>)<\/span>;<\/span>
        \n TFLITE_CHECK_EQ<\/span>(<\/span>output-><\/span>type,<\/span> kTfLiteInt8)<\/span>;<\/span><\/p>\n

        const<\/span> float<\/span> output_scale =<\/span> output-><\/span>params.<\/span>scale;<\/span>
        \n const<\/span> int32_t<\/span> output_zero_point =<\/span> output-><\/span>params.<\/span>zero_point;<\/span>
        \n const<\/span> int<\/span> output_size =<\/span> output-><\/span>dims-><\/span>data[<\/span>1<\/span>]<\/span> *<\/span> output-><\/span>dims-><\/span>data[<\/span>2<\/span>]<\/span> *<\/span>
        \n output-><\/span>dims-><\/span>data[<\/span>3<\/span>]<\/span>;<\/span><\/p>\n

        \/\/ \u53cd\u91cf\u5316\u5230\u6d6e\u70b9\u6570\u7ec4<\/span>
        \n float<\/span> actual_output[<\/span>64<\/span>]<\/span>;<\/span> \/\/ \u5927\u5c0f\u9700\u6839\u636e\u6a21\u578b\u8f93\u51fa\u8c03\u6574<\/span>
        \n for<\/span> (<\/span>int<\/span> i =<\/span> 0<\/span>;<\/span> i <<\/span> output_size;<\/span> ++<\/span>i)<\/span> {<\/span>
        \n actual_output[<\/span>i]<\/span> =<\/span> Dequantize<\/span>(<\/span>output-><\/span>data.<\/span>int8[<\/span>i]<\/span>,<\/span> output_scale,<\/span>
        \n output_zero_point)<\/span>;<\/span>
        \n }<\/span><\/p>\n

        \/\/ 9. \u4e0e host reference \u6bd4\u8f83，\u9a8c\u8bc1\u6b63\u786e\u6027<\/span>
        \n if<\/span> (<\/span>!<\/span>CompareWithReference<\/span>(<\/span>actual_output,<\/span> output_size)<\/span>)<\/span> {<\/span>
        \n MicroPrintf<\/span>(<\/span>"Output mismatch with host reference"<\/span>)<\/span>;<\/span>
        \n return<\/span> –<\/span>1<\/span>;<\/span>
        \n }<\/span><\/p>\n

        MicroPrintf<\/span>(<\/span>"Inference OK: output matches reference"<\/span>)<\/span>;<\/span>
        \n return<\/span> 0<\/span>;<\/span>
        \n}<\/span><\/p>\n

        \u5173\u952e\u70b9\u8bf4\u660e：<\/p>\n

          \n
        • \u6a21\u578b\u52a0\u8f7d：g_model_data \u7531 model_data.cc \u63d0\u4f9b，\u662f .tflite \u8f6c\u6362\u540e\u7684 C \u6570\u7ec4；GetModel \u89e3\u6790 FlatBuffer \u5e76\u6821\u9a8c schema \u7248\u672c\u3002<\/li>\n
        • \u8f93\u5165\u9884\u5904\u7406：\u539f\u59cb\u6d6e\u70b9\u8f93\u5165\u901a\u8fc7\u91cf\u5316\u516c\u5f0f q = round(raw \/ scale) + zero_point \u8f6c\u6210 int8，\u5199\u5165\u8f93\u5165 tensor\u3002<\/li>\n
        • \u5185\u5b58\u5206\u914d：AllocateTensors \u5728\u9759\u6001 tensor_arena \u4e0a\u5b8c\u6210\u6240\u6709 tensor \u5206\u914d，\u5931\u8d25\u901a\u5e38\u610f\u5473\u7740 arena \u592a\u5c0f，\u9700\u8981\u589e\u5927\u6216\u6539\u7528 RecordingMicroAllocator \u5206\u6790\u3002<\/li>\n
        • \u63a8\u7406\u6267\u884c：Invoke \u9a71\u52a8\u6574\u4e2a\u56fe\u6267\u884c；\u8fd4\u56de\u975e kTfLiteOk \u65f6\u5148\u68c0\u67e5\u6a21\u578b\u3001\u8f93\u5165\u548c arena，\u4e0d\u8981\u5f52\u56e0\u4e8e\u201c\u82af\u7247\u8fd8\u6ca1\u52a0\u901f\u201d\u3002<\/li>\n
        • \u8f93\u51fa\u9a8c\u8bc1：int8 \u8f93\u51fa\u901a\u8fc7 Dequantize \u8fd8\u539f\u4e3a\u6d6e\u70b9，\u518d\u4e0e host \u7aef\u4fdd\u5b58\u7684 kReferenceOutput \u9010\u5143\u7d20\u6bd4\u8f83，\u8bef\u5dee\u5728\u5bb9\u5dee\u5185\u624d\u7b97\u95ed\u73af\u6210\u529f\u3002<\/li>\n<\/ul>\n

          \u4ea4\u4ed8\u7269<\/h4>\n

          examples\/your_chip_model\/
          \n model_data.cc
          \n model_data.h
          \n main.cc
          \n input_adapter.cc
          \n output_adapter.cc
          \n README.md<\/p>\n

          \u5b8c\u6210\u6761\u4ef6<\/h4>\n

          [ ] target board \u4e0a AllocateTensors \u6210\u529f
          \n[ ] target board \u4e0a Invoke \u6210\u529f
          \n[ ] \u8f93\u51fa\u5728\u660e\u786e\u8bef\u5dee\u8303\u56f4\u5185
          \n[ ] \u8fde\u7eed\u8fd0\u884c\u4e0d\u4f1a\u7834\u574f arena \u6216\u72b6\u6001
          \n[ ] \u8bb0\u5f55\u4e86 peak arena \u4f7f\u7528\u91cf<\/p>\n

          \u5982\u679c\u8fd9\u4e00\u9636\u6bb5\u5931\u8d25，\u5148\u4fee\u590d\u6a21\u578b\u3001\u8f93\u5165\u3001arena \u6216\u5e73\u53f0\u95ee\u9898，\u4e0d\u8981\u628a\u5931\u8d25\u5f52\u56e0\u4e8e\u201c\u82af\u7247\u8fd8\u6ca1\u6709\u52a0\u901f\u201d\u3002<\/p>\n

          7. \u9636\u6bb5 3：\u5efa\u7acb NN library \u8fb9\u754c<\/h3>\n

          \u76ee\u6807<\/h4>\n

          \u8ba9\u786c\u4ef6\u5b9e\u73b0\u4e0e TFLM \u89e3\u8026，\u5f62\u6210\u53ef\u72ec\u7acb\u6d4b\u8bd5\u7684 NN library API\u3002<\/p>\n

          \u63a8\u8350\u76ee\u5f55<\/h4>\n

          your_nnlib\/
          \n include\/
          \n your_nnlib.h
          \n src\/
          \n conv_int8.c
          \n depthwise_conv_int8.c
          \n fully_connected_int8.c
          \n tests\/
          \n conv_test.c
          \n fully_connected_test.c<\/p>\n

          tensorflow\/lite\/micro\/kernels\/your_chip\/
          \n README.md
          \n your_chip_common.h
          \n your_chip_common.cc
          \n conv.cc
          \n fully_connected.cc<\/p>\n

          NN library API \u5e94\u660e\u786e<\/h4>\n
            \n
          • \u652f\u6301\u7684\u6570\u636e\u7c7b\u578b\u3002<\/li>\n
          • tensor layout\u3002<\/li>\n
          • padding \u548c stride \u8868\u793a\u6cd5\u3002<\/li>\n
          • activation \u8303\u56f4\u3002<\/li>\n
          • per-tensor\/per-channel quantization\u3002<\/li>\n
          • \u8f93\u5165\u8f93\u51fa alignment \u8981\u6c42\u3002<\/li>\n
          • scratch \u5927\u5c0f\u548c\u751f\u547d\u5468\u671f\u3002<\/li>\n
          • \u540c\u6b65\u6216\u5f02\u6b65\u6267\u884c\u65b9\u5f0f\u3002<\/li>\n
          • \u9519\u8bef\u7801\u548c\u8d85\u65f6\u884c\u4e3a\u3002<\/li>\n<\/ul>\n

            \u4f8b\u5982 wrapper \u4e0d\u5e94\u731c\u6d4b\u786c\u4ef6\u5e93\u7684\u91cf\u5316\u7ea6\u5b9a，\u800c\u5e94\u660e\u786e\u8f6c\u6362：<\/p>\n

            TFLM zero_point \/ multiplier \/ shift
            \n -> your_nnlib quantization parameters<\/p>\n

            \u5b8c\u6210\u6761\u4ef6<\/h4>\n

            [ ] NN library \u53ef\u8131\u79bb TFLM \u5355\u72ec\u7f16\u8bd1
            \n[ ] NN library \u6709\u72ec\u7acb\u8f93\u5165\u8f93\u51fa\u6d4b\u8bd5
            \n[ ] API \u80fd\u8868\u8fbe\u76ee\u6807 kernel \u7684\u9650\u5236
            \n[ ] TFLM wrapper \u4e0d\u9700\u8981\u66b4\u9732\u786c\u4ef6\u5bc4\u5b58\u5668\u7ec6\u8282<\/p>\n

            8. \u9636\u6bb5 4：\u5b9e\u73b0\u7b2c\u4e00\u4e2a optimized kernel<\/h3>\n

            \u76ee\u6807<\/h4>\n

            \u9009\u62e9\u4e00\u4e2a\u9ad8\u6536\u76ca\u3001\u8fb9\u754c\u6e05\u6670\u7684 op，\u5b8c\u6210 reference \u5230\u786c\u4ef6\u5b9e\u73b0\u7684\u66ff\u6362\u3002<\/p>\n

            \u4f18\u5148\u9009\u62e9：<\/p>\n

            CONV_2D
            \nDEPTHWISE_CONV_2D
            \nFULLY_CONNECTED<\/p>\n

            \u9009\u62e9\u4f9d\u636e\u5e94\u662f benchmark，\u800c\u4e0d\u662f\u7b97\u5b50\u540d\u79f0\u7684\u6d41\u884c\u7a0b\u5ea6\u3002<\/p>\n

            wrapper \u7684\u804c\u8d23<\/h4>\n

            TFLM TfLiteEvalTensor
            \n |
            \n \u251c\u2500\u2500 \u68c0\u67e5 type \/ shape \/ quantization
            \n \u251c\u2500\u2500 \u8ba1\u7b97\u6216\u8bfb\u53d6 OpData
            \n \u251c\u2500\u2500 \u7533\u8bf7\u548c\u53d6\u5f97 scratch
            \n \u251c\u2500\u2500 \u8f6c\u6362 padding \/ stride \/ activation
            \n \u251c\u2500\u2500 \u8c03\u7528 your_nnlib
            \n \u2514\u2500\u2500 \u5c06\u9519\u8bef\u7801\u8f6c\u6210 TfLiteStatus<\/p>\n

            Prepare \u8d1f\u8d23\u68c0\u67e5\u80fd\u529b\u548c\u7533\u8bf7\u8d44\u6e90，Eval \u6216 Invoke \u8d1f\u8d23\u63d0\u4ea4\u8ba1\u7b97\u3002\u4e0d\u8981\u5728\u6bcf\u6b21\u63a8\u7406\u4e2d\u91cd\u590d\u505a\u53ef\u4ee5\u63d0\u524d\u5b8c\u6210\u7684\u53c2\u6570\u8ba1\u7b97\u3002<\/p>\n

            \u4e0b\u9762\u662f\u4e00\u4e2a\u5b8c\u6574\u7684 CONV_2D wrapper \u793a\u4f8b，\u5c55\u793a Prepare \u548c Eval \u7684\u9aa8\u67b6\u3001OpData \u7ed3\u6784\u5b9a\u4e49\u3001\u91cf\u5316\u53c2\u6570\u8f6c\u6362，\u4ee5\u53ca\u8c03\u7528 your_nnlib \u7684\u63a5\u53e3：<\/p>\n

            \/\/ tensorflow\/lite\/micro\/kernels\/your_chip\/conv.cc<\/span>
            \n#<\/span>include<\/span> "tensorflow\/lite\/micro\/kernels\/your_chip\/your_chip_common.h"<\/span><\/span>
            \n#<\/span>include<\/span> "your_nnlib.h"<\/span> \/\/ \u82af\u7247 NN library \u5934\u6587\u4ef6<\/span><\/span><\/p>\n

            namespace<\/span> tflite {<\/span><\/p>\n

            \/\/ OpData：\u5728 Prepare \u9636\u6bb5\u8ba1\u7b97\u5e76\u6301\u4e45\u5316，\u907f\u514d\u6bcf\u6b21\u63a8\u7406\u91cd\u590d\u8ba1\u7b97<\/span>
            \nstruct<\/span> OpData<\/span> {<\/span>
            \n \/\/ \u91cf\u5316\u53c2\u6570（TFLM \u683c\u5f0f）<\/span>
            \n int32_t<\/span> output_multiplier;<\/span>
            \n int<\/span> output_shift;<\/span>
            \n int32_t<\/span> output_zero_point;<\/span>
            \n int32_t<\/span> input_zero_point;<\/span>
            \n int32_t<\/span> filter_zero_point;<\/span><\/p>\n

            \/\/ \u8f6c\u6362\u540e\u7684 your_nnlib \u91cf\u5316\u53c2\u6570<\/span>
            \n your_nnlib_quant_params nn_quant;<\/span><\/p>\n

            \/\/ \u5377\u79ef\u53c2\u6570<\/span>
            \n int32_t<\/span> padding_h;<\/span>
            \n int32_t<\/span> padding_w;<\/span>
            \n int32_t<\/span> stride_h;<\/span>
            \n int32_t<\/span> stride_w;<\/span><\/p>\n

            \/\/ scratch buffer \u7d22\u5f15<\/span>
            \n int<\/span> scratch_index;<\/span>
            \n}<\/span>;<\/span><\/p>\n

            \/\/ \u5c06 TFLM \u91cf\u5316\u53c2\u6570\u8f6c\u6362\u4e3a your_nnlib \u683c\u5f0f<\/span>
            \nstatic<\/span> void<\/span> ConvertQuantParams<\/span>(<\/span>const<\/span> TfLiteAffineQuantization*<\/span> quant,<\/span>
            \n OpData*<\/span> data)<\/span> {<\/span>
            \n \/\/ \u8bfb\u53d6 TFLM \u7684 multiplier \/ shift \/ zero_point<\/span>
            \n data-><\/span>output_multiplier =<\/span> quant-><\/span>multiplier-><\/span>data[<\/span>0<\/span>]<\/span>;<\/span>
            \n data-><\/span>output_shift =<\/span> quant-><\/span>shift-><\/span>data[<\/span>0<\/span>]<\/span>;<\/span>
            \n data-><\/span>output_zero_point =<\/span> quant-><\/span>zero_point-><\/span>data[<\/span>0<\/span>]<\/span>;<\/span><\/p>\n

            \/\/ \u8f6c\u6362\u4e3a your_nnlib \u7684\u91cf\u5316\u53c2\u6570\u7ed3\u6784<\/span>
            \n data-><\/span>nn_quant.<\/span>multiplier =<\/span> data-><\/span>output_multiplier;<\/span>
            \n data-><\/span>nn_quant.<\/span>shift =<\/span> data-><\/span>output_shift;<\/span>
            \n data-><\/span>nn_quant.<\/span>zero_point =<\/span> data-><\/span>output_zero_point;<\/span>
            \n}<\/span><\/p>\n

            \/\/ Prepare：\u68c0\u67e5\u80fd\u529b\u3001\u8ba1\u7b97 OpData\u3001\u7533\u8bf7 scratch<\/span>
            \nTfLiteStatus ConvPrepare<\/span>(<\/span>TfLiteContext*<\/span> context,<\/span> TfLiteNode*<\/span> node)<\/span> {<\/span>
            \n TF_LITE_ENSURE_EQ<\/span>(<\/span>context,<\/span> NumInputs<\/span>(<\/span>node)<\/span>,<\/span> 3<\/span>)<\/span>;<\/span> \/\/ input, filter, bias<\/span><\/p>\n

            \/\/ \u68c0\u67e5\u662f\u5426\u652f\u6301 int8（\u672c\u793a\u4f8b\u53ea\u652f\u6301 int8）<\/span>
            \n const<\/span> TfLiteEvalTensor*<\/span> input =<\/span>
            \n tflite::<\/span>micro::<\/span>GetEvalInput<\/span>(<\/span>context,<\/span> node,<\/span> 0<\/span>)<\/span>;<\/span>
            \n TF_LITE_ENSURE_EQ<\/span>(<\/span>context,<\/span> input-><\/span>type,<\/span> kTfLiteInt8)<\/span>;<\/span><\/p>\n

            \/\/ \u5206\u914d\u5e76\u521d\u59cb\u5316 OpData（persistent，\u8de8 Invoke \u4fdd\u7559）<\/span>
            \n OpData*<\/span> data =<\/span> reinterpret_cast<\/span><<\/span>OpData*<\/span>><\/span><\/span><\/span>(<\/span>
            \n context-><\/span>AllocatePersistentBuffer<\/span>(<\/span>context,<\/span> sizeof<\/span>(<\/span>OpData)<\/span>)<\/span>)<\/span>;<\/span>
            \n node-><\/span>user_data =<\/span> data;<\/span><\/p>\n

            \/\/ \u8bfb\u53d6\u5377\u79ef\u53c2\u6570（padding \/ stride）<\/span>
            \n auto<\/span>*<\/span> params =<\/span> reinterpret_cast<\/span><<\/span>TfLiteConvParams*<\/span>><\/span><\/span><\/span>(<\/span>node-><\/span>builtin_data)<\/span>;<\/span>
            \n data-><\/span>stride_h =<\/span> params-><\/span>stride_height;<\/span>
            \n data-><\/span>stride_w =<\/span> params-><\/span>stride_width;<\/span>
            \n data-><\/span>padding_h =<\/span> params-><\/span>padding ==<\/span> kTfLitePaddingSame ?<\/span> 1<\/span> :<\/span> 0<\/span>;<\/span><\/p>\n

            \/\/ \u8f6c\u6362\u91cf\u5316\u53c2\u6570<\/span>
            \n const<\/span> auto<\/span>*<\/span> quant =<\/span> reinterpret_cast<\/span><<\/span>TfLiteAffineQuantization*<\/span>><\/span><\/span><\/span>(<\/span>
            \n input-><\/span>quantization.<\/span>params)<\/span>;<\/span>
            \n ConvertQuantParams<\/span>(<\/span>quant,<\/span> data)<\/span>;<\/span><\/p>\n

            \/\/ \u7533\u8bf7 scratch buffer（\u4f8b\u5982\u7528\u4e8e DMA \u5bf9\u9f50\u7684\u4e2d\u95f4\u7f13\u51b2）<\/span>
            \n TF_LITE_ENSURE_STATUS<\/span>(<\/span>context-><\/span>RequestScratchBufferInArena<\/span>(<\/span>
            \n context,<\/span> input-><\/span>dims-><\/span>data[<\/span>1<\/span>]<\/span> *<\/span> input-><\/span>dims-><\/span>data[<\/span>2<\/span>]<\/span> *<\/span> 4<\/span>,<\/span>
            \n &<\/span>data-><\/span>scratch_index)<\/span>)<\/span>;<\/span><\/p>\n

            return<\/span> kTfLiteOk;<\/span>
            \n}<\/span><\/p>\n

            \/\/ Eval：\u63d0\u4ea4\u8ba1\u7b97\u5230 your_nnlib<\/span>
            \nTfLiteStatus ConvEval<\/span>(<\/span>TfLiteContext*<\/span> context,<\/span> TfLiteNode*<\/span> node)<\/span> {<\/span>
            \n OpData*<\/span> data =<\/span> reinterpret_cast<\/span><<\/span>OpData*<\/span>><\/span><\/span><\/span>(<\/span>node-><\/span>user_data)<\/span>;<\/span><\/p>\n

            \/\/ \u83b7\u53d6\u8f93\u5165\u8f93\u51fa tensor<\/span>
            \n const<\/span> TfLiteEvalTensor*<\/span> input =<\/span>
            \n tflite::<\/span>micro::<\/span>GetEvalInput<\/span>(<\/span>context,<\/span> node,<\/span> 0<\/span>)<\/span>;<\/span>
            \n const<\/span> TfLiteEvalTensor*<\/span> filter =<\/span>
            \n tflite::<\/span>micro::<\/span>GetEvalInput<\/span>(<\/span>context,<\/span> node,<\/span> 1<\/span>)<\/span>;<\/span>
            \n TfLiteEvalTensor*<\/span> output =<\/span>
            \n tflite::<\/span>micro::<\/span>GetEvalOutput<\/span>(<\/span>context,<\/span> node,<\/span> 0<\/span>)<\/span>;<\/span><\/p>\n

            \/\/ \u83b7\u53d6 scratch buffer<\/span>
            \n void<\/span>*<\/span> scratch =<\/span> context-><\/span>GetScratchBuffer<\/span>(<\/span>context,<\/span> data-><\/span>scratch_index)<\/span>;<\/span><\/p>\n

            \/\/ \u6784\u9020 your_nnlib \u7684\u8f93\u5165\u63cf\u8ff0<\/span>
            \n your_nnlib_conv_params nn_params;<\/span>
            \n nn_params.<\/span>input =<\/span> static_cast<\/span><<\/span>const<\/span> int8_t<\/span>*<\/span>><\/span><\/span><\/span>(<\/span>input-><\/span>data.<\/span>data)<\/span>;<\/span>
            \n nn_params.<\/span>filter =<\/span> static_cast<\/span><<\/span>const<\/span> int8_t<\/span>*<\/span>><\/span><\/span><\/span>(<\/span>filter-><\/span>data.<\/span>data)<\/span>;<\/span>
            \n nn_params.<\/span>output =<\/span> static_cast<\/span><<\/span>int8_t<\/span>*<\/span>><\/span><\/span><\/span>(<\/span>output-><\/span>data.<\/span>data)<\/span>;<\/span>
            \n nn_params.<\/span>quant =<\/span> data-><\/span>nn_quant;<\/span>
            \n nn_params.<\/span>stride_h =<\/span> data-><\/span>stride_h;<\/span>
            \n nn_params.<\/span>stride_w =<\/span> data-><\/span>stride_w;<\/span>
            \n nn_params.<\/span>padding_h =<\/span> data-><\/span>padding_h;<\/span>
            \n nn_params.<\/span>padding_w =<\/span> data-><\/span>padding_w;<\/span>
            \n nn_params.<\/span>scratch =<\/span> scratch;<\/span><\/p>\n

            \/\/ \u8c03\u7528\u82af\u7247 NN library，\u5e76\u8f6c\u6362\u9519\u8bef\u7801<\/span>
            \n int<\/span> ret =<\/span> your_nnlib_conv2d_int8<\/span>(<\/span>&<\/span>nn_params)<\/span>;<\/span>
            \n if<\/span> (<\/span>ret !=<\/span> YOUR_NNLIB_OK)<\/span> {<\/span>
            \n return<\/span> kTfLiteError;<\/span> \/\/ \u9519\u8bef\u7801\u8f6c\u6210 TfLiteStatus<\/span>
            \n }<\/span><\/p>\n

            return<\/span> kTfLiteOk;<\/span>
            \n}<\/span><\/p>\n

            \/\/ \u6ce8\u518c\u5230 TFLM kernel \u6ce8\u518c\u8868<\/span>
            \nTFLMRegistration Register_CONV_2D<\/span>(<\/span>)<\/span> {<\/span>
            \n return<\/span> tflite::<\/span>micro::<\/span>RegisterOp<\/span>(<\/span>ConvInit,<\/span> ConvPrepare,<\/span> ConvEval)<\/span>;<\/span>
            \n}<\/span><\/p>\n

            }<\/span> \/\/ namespace tflite<\/span><\/p>\n

            \u5173\u952e\u70b9\u8bf4\u660e：<\/p>\n

              \n
            • Prepare \u53ea\u505a\u4e00\u6b21：\u68c0\u67e5 type\/shape\u3001\u8ba1\u7b97 OpData\u3001\u7533\u8bf7 scratch，\u907f\u514d\u5728\u6bcf\u6b21 Eval \u4e2d\u91cd\u590d\u8ba1\u7b97\u91cf\u5316\u53c2\u6570\u3002<\/li>\n
            • OpData \u4f7f\u7528 AllocatePersistentBuffer \u5206\u914d，\u8de8\u591a\u6b21 Invoke \u4fdd\u7559\u3002<\/li>\n
            • \u91cf\u5316\u53c2\u6570\u5728 Prepare \u4e2d\u4ece TFLM \u683c\u5f0f\u8f6c\u6362\u4e3a your_nnlib \u683c\u5f0f，Eval \u76f4\u63a5\u4f7f\u7528\u3002<\/li>\n
            • Eval \u53ea\u8d1f\u8d23\u628a tensor \u6307\u9488\u3001\u91cf\u5316\u53c2\u6570\u548c\u5377\u79ef\u53c2\u6570\u6253\u5305\u4f20\u7ed9 your_nnlib，\u5e76\u628a\u9519\u8bef\u7801\u8f6c\u6210 TfLiteStatus\u3002<\/li>\n
            • \u4e0d\u652f\u6301\u7684\u7ec4\u5408（\u5982 float32）\u5728 Prepare \u4e2d\u901a\u8fc7 TF_LITE_ENSURE_EQ \u660e\u786e\u62d2\u7edd，\u907f\u514d\u9759\u9ed8\u4ea7\u751f\u9519\u8bef\u7ed3\u679c\u3002<\/li>\n<\/ul>\n

              fallback \u7b56\u7565<\/h4>\n

              \u5bf9\u4e0d\u652f\u6301\u7684\u7ec4\u5408\u5fc5\u987b\u660e\u786e\u5904\u7406：<\/p>\n

              \u652f\u6301 -> \u8c03\u7528 optimized kernel
              \n\u4e0d\u652f\u6301 -> \u8c03\u7528 reference kernel \u6216\u8fd4\u56de\u9519\u8bef<\/p>\n

              \u4e0d\u8981\u9759\u9ed8\u4ea7\u751f\u9519\u8bef\u7ed3\u679c\u3002\u5e38\u89c1\u9650\u5236\u5305\u62ec：<\/p>\n

                \n
              • \u53ea\u652f\u6301 int8，\u4e0d\u652f\u6301 float32\u3002<\/li>\n
              • \u53ea\u652f\u6301\u7279\u5b9a stride\u3001padding \u6216 dilation\u3002<\/li>\n
              • \u53ea\u652f\u6301 per-tensor quantization\u3002<\/li>\n
              • \u53ea\u652f\u6301\u7279\u5b9a channel \u5bf9\u9f50\u3002<\/li>\n
              • DMA buffer \u9700\u8981\u7279\u5b9a alignment\u3002<\/li>\n<\/ul>\n

                \u4e0b\u9762\u662f\u5b8c\u6574\u7684 fallback \u51b3\u7b56\u6d41\u7a0b：<\/p>\n

                #mermaid-svg-HVD7dsufoOHD2OA7{font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-HVD7dsufoOHD2OA7 .error-icon{fill:#552222;}#mermaid-svg-HVD7dsufoOHD2OA7 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-HVD7dsufoOHD2OA7 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-HVD7dsufoOHD2OA7 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-HVD7dsufoOHD2OA7 .marker.cross{stroke:#333333;}#mermaid-svg-HVD7dsufoOHD2OA7 svg{font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-HVD7dsufoOHD2OA7 p{margin:0;}#mermaid-svg-HVD7dsufoOHD2OA7 .label{font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster-label text{fill:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster-label span{color:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster-label span p{background-color:transparent;}#mermaid-svg-HVD7dsufoOHD2OA7 .label text,#mermaid-svg-HVD7dsufoOHD2OA7 span{fill:#333;color:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 .node rect,#mermaid-svg-HVD7dsufoOHD2OA7 .node circle,#mermaid-svg-HVD7dsufoOHD2OA7 .node ellipse,#mermaid-svg-HVD7dsufoOHD2OA7 .node polygon,#mermaid-svg-HVD7dsufoOHD2OA7 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-HVD7dsufoOHD2OA7 .rough-node .label text,#mermaid-svg-HVD7dsufoOHD2OA7 .node .label text,#mermaid-svg-HVD7dsufoOHD2OA7 .image-shape .label,#mermaid-svg-HVD7dsufoOHD2OA7 .icon-shape .label{text-anchor:middle;}#mermaid-svg-HVD7dsufoOHD2OA7 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-HVD7dsufoOHD2OA7 .rough-node .label,#mermaid-svg-HVD7dsufoOHD2OA7 .node .label,#mermaid-svg-HVD7dsufoOHD2OA7 .image-shape .label,#mermaid-svg-HVD7dsufoOHD2OA7 .icon-shape .label{text-align:center;}#mermaid-svg-HVD7dsufoOHD2OA7 .node.clickable{cursor:pointer;}#mermaid-svg-HVD7dsufoOHD2OA7 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-HVD7dsufoOHD2OA7 .arrowheadPath{fill:#333333;}#mermaid-svg-HVD7dsufoOHD2OA7 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-HVD7dsufoOHD2OA7 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-HVD7dsufoOHD2OA7 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-HVD7dsufoOHD2OA7 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-HVD7dsufoOHD2OA7 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-HVD7dsufoOHD2OA7 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster text{fill:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 .cluster span{color:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-HVD7dsufoOHD2OA7 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-HVD7dsufoOHD2OA7 rect.text{fill:none;stroke-width:0;}#mermaid-svg-HVD7dsufoOHD2OA7 .icon-shape,#mermaid-svg-HVD7dsufoOHD2OA7 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-HVD7dsufoOHD2OA7 .icon-shape p,#mermaid-svg-HVD7dsufoOHD2OA7 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-HVD7dsufoOHD2OA7 .icon-shape .label rect,#mermaid-svg-HVD7dsufoOHD2OA7 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-HVD7dsufoOHD2OA7 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-HVD7dsufoOHD2OA7 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-HVD7dsufoOHD2OA7 :root{–mermaid-font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;}<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u652f\u6301<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u652f\u6301<\/p>\n

                <\/span><\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u6210\u529f<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u5931\u8d25<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u4e0d\u652f\u6301<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u4e0d\u652f\u6301<\/p>\n

                <\/span><\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                kernel \u88ab\u8c03\u7528（Eval\/Invoke）<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u68c0\u67e5 type \/ shape \/ \u91cf\u5316\u53c2\u6570<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u68c0\u67e5 stride \/ padding \/ dilation \/ channel \u5bf9\u9f50<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u8c03\u7528 optimized kernel（your_nnlib）<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u6267\u884c\u7ed3\u679c<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u8fd4\u56de kTfLiteOk<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u8fd4\u56de kTfLiteError<\/p>\n

                <\/span><\/p>\n

                <\/p>\n

                \u8c03\u7528 reference kernel<\/p>\n

                <\/span><\/p>\n

                \u51b3\u7b56\u8981\u70b9：<\/p>\n

                  \n
                • \u5728 Prepare \u9636\u6bb5\u5c31\u5b8c\u6210\u80fd\u529b\u68c0\u67e5，\u628a\u4e0d\u652f\u6301\u7684\u7ec4\u5408\u5c3d\u65e9\u62d2\u7edd，\u907f\u514d\u5728\u6bcf\u6b21 Eval \u4e2d\u91cd\u590d\u5224\u65ad\u3002<\/li>\n
                • \u53ea\u6709\u6240\u6709\u6761\u4ef6\u90fd\u6ee1\u8db3\u65f6\u624d\u8d70 optimized \u8def\u5f84；\u4efb\u4f55\u4e00\u9879\u4e0d\u6ee1\u8db3\u90fd\u56de\u9000\u5230 reference kernel，\u4fdd\u8bc1\u6b63\u786e\u6027\u4f18\u5148\u3002<\/li>\n
                • \u786c\u4ef6\u6267\u884c\u5931\u8d25（\u5982 DMA \u8d85\u65f6\u3001\u547d\u4ee4\u63d0\u4ea4\u9519\u8bef）\u4e0d\u80fd\u9759\u9ed8\u541e\u6389，\u5fc5\u987b\u8f6c\u6210 kTfLiteError \u8ba9\u4e0a\u5c42\u53ef\u89c1\u3002<\/li>\n<\/ul>\n

                  \u4e0b\u9762\u662f\u5e38\u89c1\u4e0d\u652f\u6301\u7ec4\u5408\u5728 optimized \u4e0e reference \u4e24\u6761\u8def\u5f84\u4e0b\u7684\u884c\u4e3a\u5bf9\u6bd4：<\/p>\n\n\u4e0d\u652f\u6301\u7ec4\u5408optimized \u8def\u5f84\u884c\u4e3areference \u8def\u5f84\u884c\u4e3a\u9519\u8bef\u5904\u7406\u65b9\u5f0f<\/tr>\n\n\n\n\n\n\n\n\n
                  float32 \u8f93\u5165<\/td>\nPrepare \u9636\u6bb5\u901a\u8fc7 TF_LITE_ENSURE_EQ \u76f4\u63a5\u62d2\u7edd，\u4e0d\u8fdb\u5165 Eval<\/td>\n\u6b63\u5e38\u6267\u884c\u6d6e\u70b9 reference kernel，\u8f93\u51fa\u6b63\u786e\u7ed3\u679c<\/td>\noptimized \u8fd4\u56de kTfLiteError \u5e76\u6253\u5370\u65e5\u5fd7；reference \u6b63\u5e38\u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  \u7279\u5b9a stride（\u5982 stride=3）<\/td>\nPrepare \u68c0\u67e5 stride \u4e0d\u5728\u652f\u6301\u96c6\u5408\u5185，\u56de\u9000\u5230 reference<\/td>\n\u652f\u6301\u4efb\u610f stride，\u6309\u53c2\u6570\u6b63\u5e38\u8ba1\u7b97<\/td>\noptimized \u4e0d\u62a5\u9519，\u9759\u9ed8\u56de\u9000\u5230 reference；reference \u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  per-channel quantization<\/td>\nPrepare \u68c0\u6d4b\u5230 per-channel \u91cf\u5316\u53c2\u6570，\u56de\u9000\u5230 reference<\/td>\n\u652f\u6301 per-channel scale\/zero_point，\u9010\u901a\u9053\u53cd\u91cf\u5316<\/td>\noptimized \u4e0d\u62a5\u9519，\u9759\u9ed8\u56de\u9000\u5230 reference；reference \u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  \u7279\u5b9a dilation（\u5982 dilation=2）<\/td>\nPrepare \u68c0\u67e5 dilation \u8d85\u51fa\u652f\u6301\u8303\u56f4，\u56de\u9000\u5230 reference<\/td>\n\u652f\u6301\u4efb\u610f dilation，\u6309\u53c2\u6570\u6b63\u5e38\u8ba1\u7b97<\/td>\noptimized \u4e0d\u62a5\u9519，\u9759\u9ed8\u56de\u9000\u5230 reference；reference \u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  channel \u6570\u4e0d\u6ee1\u8db3\u5bf9\u9f50\u8981\u6c42<\/td>\nPrepare \u68c0\u67e5 channel \u5bf9\u9f50，\u4e0d\u6ee1\u8db3\u5219\u56de\u9000\u5230 reference<\/td>\n\u65e0\u5bf9\u9f50\u8981\u6c42，\u4efb\u610f channel \u5747\u53ef\u8ba1\u7b97<\/td>\noptimized \u4e0d\u62a5\u9519，\u9759\u9ed8\u56de\u9000\u5230 reference；reference \u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  DMA buffer \u672a\u5bf9\u9f50<\/td>\nEval \u9636\u6bb5\u68c0\u67e5\u5730\u5740 alignment，\u4e0d\u6ee1\u8db3\u5219\u56de\u9000\u5230 reference<\/td>\n\u65e0 DMA \u4f9d\u8d56，\u666e\u901a\u5185\u5b58\u8bbf\u95ee\u5373\u53ef<\/td>\noptimized \u4e0d\u62a5\u9519，\u9759\u9ed8\u56de\u9000\u5230 reference；reference \u8fd4\u56de kTfLiteOk<\/td>\n<\/tr>\n
                  \u786c\u4ef6\u6267\u884c\u5931\u8d25（DMA \u8d85\u65f6\/\u547d\u4ee4\u9519\u8bef）<\/td>\nEval \u8c03\u7528 your_nnlib \u8fd4\u56de\u975e YOUR_NNLIB_OK<\/td>\n\u4e0d\u6d89\u53ca\u786c\u4ef6\u8c03\u7528，\u4e0d\u5b58\u5728\u8be5\u5931\u8d25\u573a\u666f<\/td>\noptimized \u8fd4\u56de kTfLiteError \u5e76\u6253\u5370\u9519\u8bef\u7801；reference \u4e0d\u9002\u7528<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                  \u5bf9\u6bd4\u8981\u70b9：<\/p>\n

                    \n
                  • \u80fd\u529b\u68c0\u67e5\u7c7b（float32\u3001stride\u3001dilation\u3001per-channel\u3001channel \u5bf9\u9f50）\u5728 Prepare \u9636\u6bb5\u5b8c\u6210，\u80fd\u63d0\u524d\u5224\u65ad\u5c31\u63d0\u524d\u5224\u65ad，\u907f\u514d\u6bcf\u6b21 Eval \u91cd\u590d\u68c0\u67e5\u3002<\/li>\n
                  • \u8fd0\u884c\u65f6\u5931\u8d25\u7c7b（DMA \u8d85\u65f6\u3001\u547d\u4ee4\u63d0\u4ea4\u9519\u8bef）\u53ea\u80fd\u5728 Eval \u9636\u6bb5\u53d1\u73b0，\u5fc5\u987b\u8f6c\u6210 kTfLiteError \u8ba9\u4e0a\u5c42\u53ef\u89c1，\u4e0d\u80fd\u9759\u9ed8\u541e\u6389\u3002<\/li>\n
                  • \u56de\u9000\u5230 reference \u662f\u300c\u6b63\u786e\u6027\u4f18\u5148\u300d\u7684\u9ed8\u8ba4\u7b56\u7565：\u53ea\u8981 optimized \u4e0d\u652f\u6301，\u5c31\u4fdd\u8bc1 reference \u80fd\u7ed9\u51fa\u6b63\u786e\u7ed3\u679c；\u53ea\u6709\u786c\u4ef6\u771f\u6b63\u6267\u884c\u5931\u8d25\u65f6\u624d\u8fd4\u56de\u9519\u8bef\u3002<\/li>\n<\/ul>\n

                    \u5b8c\u6210\u6761\u4ef6<\/h4>\n

                    [ ] \u5355\u7b97\u5b50\u6d4b\u8bd5\u901a\u8fc7
                    \n[ ] \u4e0d\u652f\u6301\u7684 case \u6709\u660e\u786e\u884c\u4e3a
                    \n[ ] reference \u548c optimized \u8f93\u51fa\u4e00\u81f4\u6216\u8bef\u5dee\u6709\u5b9a\u4e49
                    \n[ ] kernel \u4e0d\u5f15\u5165\u4e0d\u5fc5\u8981\u7684\u6570\u636e\u590d\u5236
                    \n[ ] benchmark \u8bc1\u660e\u6709\u6027\u80fd\u6536\u76ca<\/p>\n

                    9. \u9636\u6bb5 5：\u6269\u5c55\u70ed\u70b9 op<\/h3>\n

                    \u76ee\u6807<\/h4>\n

                    \u7528 profiling \u7ed3\u679c\u51b3\u5b9a\u4e0b\u4e00\u6279\u4f18\u5316，\u800c\u4e0d\u662f\u4e00\u6b21\u6027\u5b9e\u73b0\u5168\u90e8 kernel\u3002<\/p>\n

                    \u63a8\u8350\u5de5\u4f5c\u5faa\u73af：<\/p>\n

                    \u8dd1\u6574\u6a21\u578b profiler
                    \n -> \u627e\u51fa\u70ed\u70b9 op
                    \n -> \u9009\u62e9\u4e00\u4e2a kernel
                    \n -> \u52a0\u6d4b\u8bd5
                    \n -> \u52a0 optimized wrapper
                    \n -> \u91cd\u65b0\u8dd1\u6574\u6a21\u578b
                    \n -> \u8bb0\u5f55\u6536\u76ca<\/p>\n

                    \u6bcf\u4e2a kernel \u5355\u72ec\u63d0\u4ea4\u3001\u5355\u72ec\u6d4b\u91cf，\u4fbf\u4e8e\u5b9a\u4f4d\u56de\u5f52\u3002<\/p>\n

                    \u8bc4\u4f30\u8868<\/h4>\n\n\u6307\u6807referenceoptimized\u76ee\u6807<\/tr>\n\n\n\n\n\n\n\n
                    \u5355\u6b21 latency<\/td>\n\u8bb0\u5f55<\/td>\n\u8bb0\u5f55<\/td>\n\u964d\u4f4e<\/td>\n<\/tr>\n
                    \u6574\u6a21\u578b latency<\/td>\n\u8bb0\u5f55<\/td>\n\u8bb0\u5f55<\/td>\n\u964d\u4f4e<\/td>\n<\/tr>\n
                    peak arena<\/td>\n\u8bb0\u5f55<\/td>\n\u8bb0\u5f55<\/td>\n\u4e0d\u6076\u5316\u6216\u53ef\u89e3\u91ca<\/td>\n<\/tr>\n
                    \u56fa\u4ef6\u4ee3\u7801\u4f53\u79ef<\/td>\n\u8bb0\u5f55<\/td>\n\u8bb0\u5f55<\/td>\n\u5728\u9884\u7b97\u5185<\/td>\n<\/tr>\n
                    \u529f\u8017<\/td>\n\u8bb0\u5f55<\/td>\n\u8bb0\u5f55<\/td>\n\u5728\u9884\u7b97\u5185<\/td>\n<\/tr>\n
                    \u8f93\u51fa\u8bef\u5dee<\/td>\n\u57fa\u7ebf<\/td>\n\u5bf9\u6bd4<\/td>\n\u6ee1\u8db3\u7cbe\u5ea6\u8981\u6c42<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                    \u5982\u679c\u67d0\u4e2a kernel \u52a0\u901f\u5f88\u5feb，\u4f46\u5f15\u5165\u5927\u91cf scratch \u6216\u590d\u5236\u5bfc\u81f4\u6574\u6a21\u578b\u53d8\u6162，\u5c31\u4e0d\u80fd\u53ea\u770b\u8be5 kernel \u7684\u5c40\u90e8\u6570\u5b57\u3002<\/p>\n

                    10. \u9636\u6bb5 6：\u5185\u5b58\u3001\u5e76\u53d1\u548c\u7a33\u5b9a\u6027<\/h3>\n

                    10.1 Tensor arena<\/h4>\n

                    \u7528 RecordingMicroInterpreter \u6216 RecordingMicroAllocator \u8bb0\u5f55：<\/p>\n

                      \n
                    • head \u4f7f\u7528\u91cf\u3002<\/li>\n
                    • tail \u4f7f\u7528\u91cf\u3002<\/li>\n
                    • temporary \u4f7f\u7528\u91cf\u3002<\/li>\n
                    • \u6bcf\u4e2a op \u7684 persistent allocation\u3002<\/li>\n
                    • scratch buffer \u603b\u91cf\u3002<\/li>\n<\/ul>\n

                      10.2 NPU\/DSP \u5185\u5b58<\/h4>\n

                      \u660e\u786e\u533a\u5206：<\/p>\n

                      TFLM tensor arena
                      \n\u82af\u7247\u672c\u5730 SRAM
                      \nDMA \u53ef\u8bbf\u95ee\u5185\u5b58
                      \nNPU command buffer
                      \ndriver \u5de5\u4f5c\u533a
                      \ncacheable \/ non-cacheable \u533a\u57df<\/p>\n

                      \u5c3d\u91cf\u8ba9\u786c\u4ef6\u76f4\u63a5\u4f7f\u7528 TFLM tensor buffer，\u907f\u514d input\/output copy；\u4f46\u5982\u679c\u786c\u4ef6\u6709 alignment\u3001cache \u6216\u5730\u5740\u7a7a\u95f4\u8981\u6c42，\u5fc5\u987b\u5728\u63a5\u53e3\u5c42\u660e\u786e\u5904\u7406\u3002<\/p>\n

                      10.3 \u5f02\u6b65\u6267\u884c<\/h4>\n

                      \u5982\u679c accelerator \u662f\u5f02\u6b65\u7684，\u5fc5\u987b\u5b9a\u4e49：<\/p>\n

                        \n
                      • command \u63d0\u4ea4\u63a5\u53e3\u3002<\/li>\n
                      • \u5b8c\u6210\u901a\u77e5\u6216\u8f6e\u8be2\u63a5\u53e3\u3002<\/li>\n
                      • timeout\u3002<\/li>\n
                      • \u9519\u8bef\u6062\u590d\u3002<\/li>\n
                      • Invoke() \u8fd4\u56de\u524d\u7684\u540c\u6b65\u4fdd\u8bc1\u3002<\/li>\n
                      • \u591a\u6b21 Invoke() \u7684\u72b6\u6001\u6e05\u7406\u3002<\/li>\n<\/ul>\n

                        TFLM \u7684 kernel Invoke \u8fd4\u56de\u540e，\u76f8\u5173 output tensor \u5fc5\u987b\u5df2\u7ecf\u6ee1\u8db3\u5e94\u7528\u8bfb\u53d6\u7ea6\u5b9a\u3002<\/p>\n

                        \u5b8c\u6210\u6761\u4ef6<\/h4>\n

                        [ ] \u8fde\u7eed Invoke \u7a33\u5b9a
                        \n[ ] \u590d\u4f4d\u540e\u53ef\u4ee5\u518d\u6b21\u8fd0\u884c
                        \n[ ] \u8d85\u65f6\u548c\u786c\u4ef6\u9519\u8bef\u53ef\u89c2\u5bdf
                        \n[ ] cache\/DMA \u4e00\u81f4\u6027\u6709\u6d4b\u8bd5
                        \n[ ] \u5cf0\u503c\u5185\u5b58\u6ca1\u6709\u8d85\u51fa\u82af\u7247\u9884\u7b97<\/p>\n

                        11. \u9636\u6bb5 7：CI\u3001\u53d1\u5e03\u548c\u7248\u672c\u7ef4\u62a4<\/h3>\n

                        CI \u6700\u5c0f\u77e9\u9635<\/h4>\n

                        host reference kernel tests
                        \nhost optimized kernel tests
                        \n\u76ee\u6807\u5e73\u53f0\u7f16\u8bd1
                        \n\u76ee\u6807\u5e73\u53f0 smoke test
                        \n\u4ee3\u8868\u6a21\u578b\u63a8\u7406\u6d4b\u8bd5
                        \nbenchmark \/ size profiling<\/p>\n

                        \u4ed3\u5e93\u5df2\u6709\u7684\u6784\u5efa\u548c\u6d4b\u8bd5\u5165\u53e3\u5305\u62ec：<\/p>\n

                        bazel test<\/span> \/\/tensorflow\/lite\/micro\/..<\/span>.<\/p>\n

                        make<\/span> -f<\/span> tensorflow\/lite\/micro\/tools\/make\/Makefile \\\\<\/span>
                        \n TARGET<\/span>=<\/span>your_chip \\\\<\/span>
                        \n OPTIMIZED_KERNEL_DIR<\/span>=<\/span>your_chip \\\\<\/span>
                        \n microlite<\/p>\n

                        \u5b9e\u9645 target \u540d\u79f0\u9700\u8981\u6839\u636e\u5e73\u53f0\u6784\u5efa\u6587\u4ef6\u52a0\u5165\u7684\u4f4d\u7f6e\u8c03\u6574\u3002<\/p>\n

                        \u4f18\u5316\u540e\u7aef\u5efa\u8bae\u63d0\u4f9b：<\/p>\n

                        tensorflow\/lite\/micro\/tools\/make\/targets\/your_chip_makefile.inc
                        \ntensorflow\/lite\/micro\/tools\/make\/ext_libs\/your_chip_nnlib.inc
                        \ntensorflow\/lite\/micro\/tools\/make\/ext_libs\/your_chip_nnlib_download.sh<\/p>\n

                        \u5982\u679c\u4f7f\u7528 Bazel，\u4e5f\u5e94\u63d0\u4f9b\u5bf9\u5e94\u7684 BUILD target \u548c\u5916\u90e8\u4f9d\u8d56\u58f0\u660e\u3002<\/p>\n

                        \u7248\u672c\u7b56\u7565<\/h4>\n

                        \u8bb0\u5f55\u4ee5\u4e0b\u7248\u672c：<\/p>\n

                          \n
                        • TFLM commit \u6216 release\u3002<\/li>\n
                        • schema \u7248\u672c\u3002<\/li>\n
                        • \u82af\u7247 SDK \u7248\u672c\u3002<\/li>\n
                        • NN library \u7248\u672c\u3002<\/li>\n
                        • compiler\/toolchain \u7248\u672c\u3002<\/li>\n
                        • firmware\/driver \u7248\u672c\u3002<\/li>\n
                        • \u6a21\u578b\u8f6c\u6362\u5de5\u5177\u7248\u672c\u3002<\/li>\n<\/ul>\n

                          TFLM\u3001\u6a21\u578b\u8f6c\u6362\u5668\u548c NN library \u4efb\u4e00\u5347\u7ea7，\u90fd\u5e94\u91cd\u65b0\u8dd1\u6a21\u578b\u517c\u5bb9\u6027\u6d4b\u8bd5\u3002<\/p>\n

                          12. \u89d2\u8272\u5206\u5de5\u5efa\u8bae<\/h3>\n\n\u89d2\u8272\u8d1f\u8d23\u5185\u5bb9<\/tr>\n\n\n\n\n\n\n\n
                          \u5e73\u53f0\u5de5\u7a0b\u5e08<\/td>\n\u5de5\u5177\u94fe\u3001\u542f\u52a8\u3001\u94fe\u63a5\u3001\u65e5\u5fd7\u3001\u8ba1\u65f6\u3001\u5185\u5b58\u5e03\u5c40\u3002<\/td>\n<\/tr>\n
                          TFLM \u5de5\u7a0b\u5e08<\/td>\nresolver\u3001\u6a21\u578b\u63a5\u5165\u3001kernel wrapper\u3001reference \u5bf9\u9f50\u3002<\/td>\n<\/tr>\n
                          NN library \u5de5\u7a0b\u5e08<\/td>\n\u7b97\u6cd5\u5b9e\u73b0\u3001SIMD\/NPU API\u3001scratch \u548c\u6027\u80fd\u3002<\/td>\n<\/tr>\n
                          driver\/firmware \u5de5\u7a0b\u5e08<\/td>\n\u547d\u4ee4\u63d0\u4ea4\u3001DMA\u3001cache\u3001\u540c\u6b65\u3001\u9519\u8bef\u6062\u590d\u3002<\/td>\n<\/tr>\n
                          \u6a21\u578b\u5de5\u7a0b\u5e08<\/td>\n\u6a21\u578b\u8f6c\u6362\u3001\u91cf\u5316\u3001\u8f93\u5165\u9884\u5904\u7406\u3001\u8f93\u51fa\u89e3\u91ca\u3002<\/td>\n<\/tr>\n
                          QA\/\u6027\u80fd\u5de5\u7a0b\u5e08<\/td>\n\u6b63\u786e\u6027\u3001\u56de\u5f52\u3001benchmark\u3001size \u548c\u529f\u8017\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                          \u4e00\u4e2a kernel \u7684 owner \u5e94\u540c\u65f6\u62e5\u6709：\u63a5\u53e3\u7ea6\u5b9a\u3001\u6d4b\u8bd5\u6570\u636e\u3001benchmark \u548c\u5931\u8d25 case，\u800c\u4e0d\u662f\u53ea\u7ef4\u62a4\u4e00\u4efd .cc \u6587\u4ef6\u3002<\/p>\n

                          13. \u98ce\u9669\u4e0e\u5e94\u5bf9<\/h3>\n\n\u98ce\u9669\u65e9\u671f\u4fe1\u53f7\u5e94\u5bf9<\/tr>\n\n\n\n\n\n\n\n\n\n
                          \u6a21\u578b op \u672a\u8986\u76d6<\/td>\nMissing registration<\/td>\n\u5148\u5206\u6790 operator_codes，\u518d\u8865 resolver\u3002<\/td>\n<\/tr>\n
                          arena \u4e0d\u591f<\/td>\nAllocateTensors() \u5931\u8d25<\/td>\nrecording allocator \u5206\u6790 head\/tail\/scratch\u3002<\/td>\n<\/tr>\n
                          \u91cf\u5316\u4e0d\u4e00\u81f4<\/td>\n\u8f93\u51fa\u504f\u5dee\u5927<\/td>\n\u5bf9\u7167 scale\u3001zero point\u3001multiplier\u3001shift\u3002<\/td>\n<\/tr>\n
                          optimized \u4e0d\u7a33\u5b9a<\/td>\n\u53ea\u5728\u67d0\u4e9b shape \u5931\u8d25<\/td>\n\u5efa\u7acb shape\/type\/padding \u53c2\u6570\u77e9\u9635\u6d4b\u8bd5\u3002<\/td>\n<\/tr>\n
                          \u52a0\u901f\u6ca1\u6709\u6536\u76ca<\/td>\n\u5355 op \u5feb\u4f46\u6574\u6a21\u578b\u4e0d\u5feb<\/td>\n\u6d4b\u91cf copy\u3001\u540c\u6b65\u548c\u8c03\u5ea6\u5f00\u9500\u3002<\/td>\n<\/tr>\n
                          DMA\/cache \u9519\u8bef<\/td>\n\u5076\u53d1\u9519\u8bef\u6216\u91cd\u590d\u8fd0\u884c\u5931\u8d25<\/td>\n\u589e\u52a0 alignment\u3001flush\/invalidate \u548c\u538b\u529b\u6d4b\u8bd5\u3002<\/td>\n<\/tr>\n
                          SDK \u5347\u7ea7\u7834\u574f\u6784\u5efa<\/td>\nCI \u7f16\u8bd1\u5931\u8d25<\/td>\n\u56fa\u5b9a\u7248\u672c\u5e76\u4fdd\u5b58\u6784\u5efa manifest\u3002<\/td>\n<\/tr>\n
                          \u53ea\u9a8c\u8bc1 demo<\/td>\n\u4e1a\u52a1\u6a21\u578b\u5931\u8d25<\/td>\n\u5c3d\u65e9\u52a0\u5165\u771f\u5b9e\u4ee3\u8868\u6a21\u578b\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

                          14. \u5de5\u7a0b\u9a8c\u6536\u6807\u51c6<\/h3>\n

                          \u529f\u80fd\u9a8c\u6536<\/h4>\n

                          [ ] \u81f3\u5c11\u4e00\u4e2a float \u6216 int8 \u6a21\u578b\u53ef\u4ee5\u8fd0\u884c
                          \n[ ] \u76ee\u6807\u6a21\u578b\u6240\u6709 op \u90fd\u6709 registration
                          \n[ ] AllocateTensors \u548c Invoke \u7a33\u5b9a\u6210\u529f
                          \n[ ] \u8f93\u51fa\u4e0e reference \u5728\u7ea6\u5b9a\u8bef\u5dee\u5185
                          \n[ ] custom op \u6709\u660e\u786e options \u89e3\u6790<\/p>\n

                          \u6027\u80fd\u9a8c\u6536<\/h4>\n

                          [ ] \u6709 reference baseline
                          \n[ ] \u6709 optimized latency
                          \n[ ] \u6709\u6574\u6a21\u578b\u6536\u76ca\u6570\u636e
                          \n[ ] \u6709 peak arena \u6570\u636e
                          \n[ ] \u6709\u4ee3\u7801\u4f53\u79ef\u548c\u529f\u8017\u6570\u636e
                          \n[ ] benchmark \u547d\u4ee4\u53ef\u590d\u73b0<\/p>\n

                          \u5de5\u7a0b\u9a8c\u6536<\/h4>\n

                          [ ] \u5e73\u53f0\u4ee3\u7801\u548c TFLM \u540e\u7aef\u8fb9\u754c\u6e05\u6670
                          \n[ ] NN library \u53ef\u72ec\u7acb\u6d4b\u8bd5
                          \n[ ] \u4e0d\u652f\u6301\u7684\u8f93\u5165\u7ec4\u5408\u6709\u660e\u786e\u884c\u4e3a
                          \n[ ] CI \u80fd\u963b\u6b62\u529f\u80fd\u548c\u6027\u80fd\u56de\u5f52
                          \n[ ] \u7248\u672c\u3001\u5de5\u5177\u94fe\u548c\u6a21\u578b\u6765\u6e90\u6709\u8bb0\u5f55<\/p>\n

                          15. \u6700\u5c0f\u53ef\u6267\u884c\u8def\u7ebf\u56fe<\/h3>\n

                          \u5982\u679c\u56e2\u961f\u8d44\u6e90\u6709\u9650，\u53ef\u4ee5\u6309\u4ee5\u4e0b\u987a\u5e8f\u843d\u5730：<\/p>\n

                          \u7b2c 1 \u5468：\u5de5\u5177\u94fe\u3001\u94fe\u63a5\u811a\u672c\u3001\u65e5\u5fd7\u3001\u8ba1\u65f6\u3001hello_world
                          \n\u7b2c 2 \u5468：\u4e00\u4e2a\u771f\u5b9e\u6a21\u578b reference \u63a8\u7406\u548c\u8f93\u51fa\u6bd4\u5bf9
                          \n\u7b2c 3 \u5468：NN library API\u3001\u72ec\u7acb\u6d4b\u8bd5\u3001\u6027\u80fd\u57fa\u7ebf
                          \n\u7b2c 4 \u5468：CONV_2D \u6216 FULLY_CONNECTED \u7b2c\u4e00\u4e2a optimized kernel
                          \n\u7b2c 5 \u5468：\u6574\u6a21\u578b benchmark\u3001arena \u548c\u9519\u8bef\u5904\u7406
                          \n\u7b2c 6 \u5468：\u7b2c\u4e8c\u4e2a\u70ed\u70b9 kernel\u3001CI \u548c\u7248\u672c\u6587\u6863<\/p>\n

                          \u5468\u6570\u53ea\u662f\u7ec4\u7ec7\u65b9\u5f0f，\u4e0d\u80fd\u66ff\u4ee3\u5b9e\u9645\u9a8c\u6536\u6761\u4ef6\u3002\u6a21\u578b\u590d\u6742\u5ea6\u3001\u82af\u7247 SDK \u5b8c\u6574\u5ea6\u548c NPU driver \u72b6\u6001\u90fd\u53ef\u80fd\u6539\u53d8\u5468\u671f\u3002<\/p>\n

                          16. \u4eca\u5929\u7684\u7cbe\u9ad3<\/h3>\n

                          \u9700\u6c42\u548c\u6a21\u578b
                          \n -> reference bring-up
                          \n -> \u5e73\u53f0\u9002\u914d
                          \n -> \u6a21\u578b\u6b63\u786e\u6027\u95ed\u73af
                          \n -> NN library \u8fb9\u754c
                          \n -> \u7b2c\u4e00\u4e2a optimized kernel
                          \n -> profiling \u9a71\u52a8\u6269\u5c55
                          \n -> CI \u548c\u7248\u672c\u7ef4\u62a4<\/p>\n

                          \u4e00\u53e5\u8bdd\u603b\u7ed3：<\/p>\n

                          AI \u82af\u7247\u63a5\u5165 TFLM \u4e0d\u662f\u628a\u51e0\u4e2a\u7b97\u5b50\u6587\u4ef6\u590d\u5236\u8fdb\u5de5\u7a0b，\u800c\u662f\u5efa\u7acb\u4e00\u6761\u53ef\u9a8c\u8bc1\u7684\u94fe\u8def：\u5e73\u53f0\u80fd\u542f\u52a8，\u6a21\u578b\u80fd\u6b63\u786e\u8fd0\u884c，\u786c\u4ef6\u5b9e\u73b0\u6709\u6e05\u6670\u8fb9\u754c，\u6027\u80fd\u6536\u76ca\u53ef\u91cd\u590d\u6d4b\u91cf，\u7248\u672c\u5347\u7ea7\u4e0d\u4f1a\u6084\u6084\u7834\u574f\u7ed3\u679c\u3002<\/p>\n

                          17. \u76f8\u5173\u6e90\u7801\u548c\u6587\u6863<\/h3>\n
                            \n
                          • tensorflow\/lite\/micro\/docs\/new_platform_support.md<\/li>\n
                          • tensorflow\/lite\/micro\/docs\/optimized_kernel_implementations.md<\/li>\n
                          • tensorflow\/lite\/micro\/examples\/hello_world\/hello_world_test.cc<\/li>\n
                          • tensorflow\/lite\/micro\/examples\/hello_world\/BUILD<\/li>\n
                          • tensorflow\/lite\/micro\/micro_interpreter.cc<\/li>\n
                          • tensorflow\/lite\/micro\/micro_mutable_op_resolver.h<\/li>\n
                          • tensorflow\/lite\/micro\/kernels\/cmsis_nn\/<\/li>\n
                          • tensorflow\/lite\/micro\/tools\/project_generation\/create_tflm_tree.py<\/li>\n
                          • tensorflow\/lite\/micro\/tools\/make\/targets\/<\/li>\n
                          • tensorflow\/lite\/micro\/tools\/make\/ext_libs\/<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

                            \u6458\u8981：\u672c\u6587\u628a\u524d 7 \u5929\u5b66\u5230\u7684 TFLM \u77e5\u8bc6\u6574\u7406\u6210\u4e00\u4efd\u53ef\u6267\u884c\u7684\u5de5\u7a0b\u8ba1\u5212，\u6307\u5bfc\u5982\u4f55\u8ba9\u4e00\u9897\u65b0\u7684 MCU\u3001DSP\u3001NPU \u6216 AI accelerator \u7a33\u5b9a\u8fd0\u884c TFLM \u5e76\u9010\u6b65\u83b7\u5f97\u53ef\u91cf\u5316\u7684\u6027\u80fd\u6536\u76ca\u3002\u6838\u5fc3\u539f\u5219\u662f\u300c\u5148\u8dd1\u901a reference，\u518d\u63a5\u5165\u4f18\u5316，\u6700\u540e\u7528\u6a21\u578b\u548c benchmark \u8bc1\u660e\u6536\u76ca\u300d\u3002\u6587\u7ae0\u6309 8 \u4e2a\u9636\u6bb5\u5c55\u5f00：\u4ece\u9700\u6c42\u4e0e baseline\u3001\u5e73\u53f0 bring-up\u3001reference \u6a21\u578b\u95ed\u73af\u3001NN l<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[997,50,78],"topic":[],"class_list":["post-96004","post","type-post","status-publish","format-standard","hentry","category-server","tag-langchain","tag-50","tag-78"],"yoast_head":"\n\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/96004.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6458\u8981：\u672c\u6587\u628a\u524d 7 \u5929\u5b66\u5230\u7684 TFLM \u77e5\u8bc6\u6574\u7406\u6210\u4e00\u4efd\u53ef\u6267\u884c\u7684\u5de5\u7a0b\u8ba1\u5212，\u6307\u5bfc\u5982\u4f55\u8ba9\u4e00\u9897\u65b0\u7684 MCU\u3001DSP\u3001NPU \u6216 AI accelerator \u7a33\u5b9a\u8fd0\u884c TFLM \u5e76\u9010\u6b65\u83b7\u5f97\u53ef\u91cf\u5316\u7684\u6027\u80fd\u6536\u76ca\u3002\u6838\u5fc3\u539f\u5219\u662f\u300c\u5148\u8dd1\u901a reference，\u518d\u63a5\u5165\u4f18\u5316，\u6700\u540e\u7528\u6a21\u578b\u548c benchmark \u8bc1\u660e\u6536\u76ca\u300d\u3002\u6587\u7ae0\u6309 8 \u4e2a\u9636\u6bb5\u5c55\u5f00：\u4ece\u9700\u6c42\u4e0e baseline\u3001\u5e73\u53f0 bring-up\u3001reference \u6a21\u578b\u95ed\u73af\u3001NN l\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/96004.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-27T02:16:29+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"14 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/96004.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/96004.html\",\"name\":\"\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-27T02:16:29+00:00\",\"dateModified\":\"2026-08-27T02:16:29+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/96004.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/96004.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/96004.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8f7b\u677e\u5b66\u4e60TFLM_day8\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/96004.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6458\u8981：\u672c\u6587\u628a\u524d 7 \u5929\u5b66\u5230\u7684 TFLM \u77e5\u8bc6\u6574\u7406\u6210\u4e00\u4efd\u53ef\u6267\u884c\u7684\u5de5\u7a0b\u8ba1\u5212，\u6307\u5bfc\u5982\u4f55\u8ba9\u4e00\u9897\u65b0\u7684 MCU\u3001DSP\u3001NPU \u6216 AI accelerator \u7a33\u5b9a\u8fd0\u884c TFLM \u5e76\u9010\u6b65\u83b7\u5f97\u53ef\u91cf\u5316\u7684\u6027\u80fd\u6536\u76ca\u3002\u6838\u5fc3\u539f\u5219\u662f\u300c\u5148\u8dd1\u901a reference，\u518d\u63a5\u5165\u4f18\u5316，\u6700\u540e\u7528\u6a21\u578b\u548c benchmark \u8bc1\u660e\u6536\u76ca\u300d\u3002\u6587\u7ae0\u6309 8 \u4e2a\u9636\u6bb5\u5c55\u5f00：\u4ece\u9700\u6c42\u4e0e baseline\u3001\u5e73\u53f0 bring-up\u3001reference \u6a21\u578b\u95ed\u73af\u3001NN l","og_url":"https:\/\/www.wsisp.com\/helps\/96004.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-27T02:16:29+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"14 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/96004.html","url":"https:\/\/www.wsisp.com\/helps\/96004.html","name":"\u8f7b\u677e\u5b66\u4e60TFLM_day8 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-27T02:16:29+00:00","dateModified":"2026-08-27T02:16:29+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/96004.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/96004.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/96004.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8f7b\u677e\u5b66\u4e60TFLM_day8"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/96004","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=96004"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/96004\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=96004"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=96004"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=96004"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=96004"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}