关于运行时，conv2d中出现的错误。

[xiao-mb]

extern "C" uint8_t app_recon(uint8_t img) { // input Tensor input; input.set_element((int16_t )img).set_exponent(0).set_shape({64, 64, 3}).set_auto_free(false); ORDER model; dl::tool::Latency latency;

// model forward
latency.start();
model.forward(input);
latency.end();

}

当运行到model.forwad(input);时出现了如下错误： Guru Meditation Error: Core 1 panic'ed (InstrFetchProhibited). Exception was un handled.

Core 1 register dump: PC : 0x00000000 PS : 0x00060e30 A0 : 0x82017c47 A1 : 0x3f c9ee70 A2 : 0x3dddd010 A3 : 0x3fce1dd0 A4 : 0x00000000 A5 : 0x00 000000 A6 : 0x3fce1d28 A7 : 0x00000000 A8 : 0x82017771 A9 : 0x3f c9ee50 A10 : 0x3fce1dd0 A11 : 0x3fce25f0 A12 : 0x3fce1d28 A13 : 0x3f ce25f0 A14 : 0x00000000 A15 : 0x00000042 SAR : 0x00000004 EXCCAUSE: 0x00 000014 EXCVADDR: 0x00000000 LBEG : 0x4037d810 LEND : 0x4037d815 LCOUNT : 0xff ffffff 0x4037d810: dl_tie728_bzero at ??:?

0x4037d815: dl_tie728_bzero at ??:?

Backtrace:0xfffffffd:0x3fc9ee700x42017c44:0x3fc9eeb0 0x42016d0c:0x3fc9eed0 0x420 09f69:0x3fc9ef10 0x42016dfd:0x3fc9eff0 0x4200b134:0x3fc9f010 0x4200935a:0x3fc9f5 70 0x4038215d:0x3fc9fc10 0x42017c44: void dl::base::conv2d<short, short, short, short, int>(void*) at /ho me/yehangyang/Documents/workspace/DL/build/../components/dl/base/dl_base_conv2d. cpp:895

0x42016d0c: dl::nn::conv2d(dl::Tensor&, dl::Tensor&, std::vector<i nt, std::allocator >&, dl::Filter const&, int, int, dl::Bias const, dl::Activation const, std::vector<int, std::allocator > con st&) at /home/yehangyang/Documents/workspace/DL/build/../components/dl/nn/src/dl _nn_conv2d.cpp:40

0x42009f69: dl::layer::Conv2D::call(dl::Tensor&, bool, std::vector <int, std::allocator > const&) at C:\Users\gaochao\eclipse-workspace\ESP32S 3_RTOS_SDK_master\build/../components/esp-dl/include/layer/dl_layer_conv2d.hpp:1 15 (inlined by) ORDER::call(dl::Tensor&) at C:\Users\gaochao\eclipse-worksp ace\ESP32S3_RTOS_SDK_master\build/../main/includes/order_model.hpp:106

0x42016dfd: dl::layer::Model::forward(dl::Tensor&) at /home/yehang yang/Documents/workspace/DL/build/../components/dl/layer/src/dl_layer_model.cpp: 30

0x4200b134: app_recon at C:\Users\gaochao\eclipse-workspace\ESP32S3_RTOS_SDK_mas ter\build/../main/app_order_recon.cpp:31

0x4200935a: task_app_recon_process(void*) at C:\Users\gaochao\eclipse-workspace\ ESP32S3_RTOS_SDK_master\build/../main/app_recon_process.cpp:122

0x4038215d: vPortTaskWrapper at C:/Users/gaochao/Desktop/esp-idf/components/free rtos/port/xtensa/port.c:168

[yehangyang]

Hi @xiao-mb, 请展示一下你调用 nn::conv2d(...) 的代码上下文。

[xiao-mb]

pragma once

include "dl_layer_model.hpp"

include "dl_layer_conv2d.hpp"

include "dl_layer_depthwise_conv2d.hpp"

include "dl_layer_concat2d.hpp"

include "order_coefficient.hpp"

include

using namespace dl; using namespace layer; using namespace order_coefficient;

class ORDER : public Model // Derive the Model class in "dl_layer_model.hpp" { private: // Declare layers as member variables Conv2D l1_conv; // a layer named l1 DepthwiseConv2D l2_dp; // a layer named l2_depth Conv2D l2_exp; // a layer named l2_compress DepthwiseConv2D l3_dp; // a layer named l3_a_depth Conv2D l3_exp; // a layer named l3_a_compress DepthwiseConv2D l4_dp; // a layer named l3_b_depth Conv2D l4_exp; // a layer named l3_b_compress DepthwiseConv2D l5_dp; // a layer named l3_c_depth Conv2D l5_exp; // a layer named l3_c_compress DepthwiseConv2D l6_dp; // a layer named l3_d_depth Conv2D l6_exp; // a layer named l3_d_compress Conv2D l7_conv; // a layer named l3_e_compress

public: Conv2D l8_cp; // a layer named l5_compress

/**
 * @brief Initialize layers in constructor function
 *
 */
ORDER() : l1_conv(Conv2D<int16_t>(-16, get_l1_conv_filter(), NULL, get_l1_conv_activation(), PADDING_VALID, 2, 2, "l1")),
          l2_dp(DepthwiseConv2D<int16_t>(-15, get_l2_dp_filter(), NULL, get_l2_dp_activation(), PADDING_SAME, 1, 1, "l2_depth")),
          l2_exp(Conv2D<int16_t>(-3, get_l2_dp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l2_compress")),
          l3_dp(DepthwiseConv2D<int16_t>(-16, get_l3_dp_filter(), NULL, get_l3_dp_activation(), PADDING_SAME, 1, 1, "l3depth")),
          l3_exp(Conv2D<int16_t>(-13, get_l3_exp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l3_a_compress")),
          l4_dp(DepthwiseConv2D<int16_t>(-15, get_l3_dp_filter(), NULL, get_l3_dp_activation(), PADDING_SAME, 2, 2, "l4_depth")),
          l4_exp(Conv2D<int16_t>(-14, get_l4_exp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l3_b_compress")),
          l5_dp(DepthwiseConv2D<int16_t>(-15, get_l4_dp_filter(), NULL, get_l5_dp_activation(), PADDING_SAME, 1, 1, "l5_depth")),
          l5_exp(Conv2D<int16_t>(-13, get_l5_exp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l5_compress")),
          l6_dp(DepthwiseConv2D<int16_t>(-15, get_l6_dp_filter(), NULL, get_l6_dp_activation(), PADDING_SAME, 2, 2, "l6_depth")),
          l6_exp(Conv2D<int16_t>(-15, get_l6_exp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l6_compress")),
          l7_conv(Conv2D<int16_t>(-14, get_l7_conv_filter(), NULL, get_l7_conv_activation(), PADDING_VALID, 1, 1, "l7_compress")),
          l8_cp(Conv2D<int16_t>(-14, get_l8_full_c_filter(), get_l8_full_c_bias(), NULL, PADDING_VALID, 1, 1, "l8_compress")){}

/**
 * @brief call each layers' build(...) function in sequence
 *
 * @param input
 */
void build(Tensor<int16_t> &input)
{
    this->l1_conv.build(input);
    this->l2_dp.build(this->l1_conv.output);
    this->l2_exp.build(this->l2_dp.output);
    this->l3_dp.build(this->l2_exp.output);
    this->l3_exp.build(this->l3_dp.output);
    this->l4_dp.build(this->l3_exp.output);
    this->l4_exp.build(this->l4_dp.output);
    this->l5_dp.build(this->l4_exp.output);
    this->l5_exp.build(this->l5_dp.output);
    this->l6_dp.build(this->l5_exp.output);
    this->l6_exp.build(this->l6_dp.output);
    this->l7_conv.build(this->l6_exp.output);
    this->l8_cp.build(this->l7_conv.output);
}

/**
 * @brief call each layers' call(...) function in sequence
 *
 * @param input
 */
void call(Tensor<int16_t> &input)
{
    this->l1_conv.call(input);
    input.free_element();

    this->l2_dp.call(this->l1_conv.output);
    this->l1_conv.output.free_element();

    this->l2_exp.call(this->l2_dp.output);
    this->l2_dp.output.free_element();

    this->l3_dp.call(this->l2_exp.output);
    this->l2_exp.output.free_element();

    this->l3_exp.call(this->l3_dp.output);
    this->l3_dp.output.free_element();

    this->l4_dp.call(this->l3_exp.output);
    this->l3_exp.output.free_element();

    this->l4_exp.call(this->l4_dp.output);
    this->l4_dp.output.free_element();

    this->l5_dp.call(this->l4_exp.output);
    this->l4_exp.output.free_element();

    this->l5_exp.call(this->l5_dp.output);
    this->l5_dp.output.free_element();

    this->l6_dp.call(this->l5_exp.output);
    this->l5_exp.output.free_element();

    this->l6_exp.call(this->l6_dp.output);
    this->l6_dp.output.free_element();

    this->l7_conv.call(this->l6_exp.output);
    this->l6_exp.output.free_element();
    this->l8_cp.call(this->l7_conv.output);
    this->l7_conv.output.free_element();
}

};

[yehangyang]

Hi @xiao-mb,

函数 uint8_t app_recon(uint8_t *img) 的 img 是如何得到的？

请务必用 inline void *malloc_aligned(int number, int size, int align = 0) 申请内存。并且设定 align = 16。

[xiao-mb]

之前是用这个函数申请的内存：uint8_t img6464 = (uint8_t )heap_caps_calloc(1, 64 64 3 sizeof(uint8_t), MALLOC_CAP_INTERNAL); 换成：uint8_t img6464 = (uint8_t )dl::tool::malloc_aligned(1, 64 64 3 sizeof(uint8_t), 16);后运行时还是有错误，如下： Guru Meditation Error: Core 1 panic'ed (InstrFetchProhibited). Exception was un handled.

Core 1 register dump: PC : 0x00000000 PS : 0x00060e30 A0 : 0x82017cab A1 : 0x3f c9ee70 A2 : 0x3dddd010 A3 : 0x3fce4e0c A4 : 0x00000000 A5 : 0x00 000000 A6 : 0x3fce4d64 A7 : 0x00000000 A8 : 0x820177d5 A9 : 0x3f c9ee50 A10 : 0x3fce4e0c A11 : 0x3fce5630 A12 : 0x3fce4d64 A13 : 0x3f ce5630 A14 : 0x00000000 A15 : 0x00000042 SAR : 0x00000004 EXCCAUSE: 0x00 000014 EXCVADDR: 0x00000000 LBEG : 0x4037d810 LEND : 0x4037d815 LCOUNT : 0xff ffffff 0x4037d810: dl_tie728_bzero at ??:?

0x4037d815: dl_tie728_bzero at ??:?

Backtrace:0xfffffffd:0x3fc9ee700x42017ca8:0x3fc9eeb0 0x42016d70:0x3fc9eed0 0x420 09fc1:0x3fc9ef10 0x42016e61:0x3fc9eff0 0x4200b198:0x3fc9f010 0x420093b2:0x3fc9f5 70 0x4038215d:0x3fc9fc10 0x42017ca8: void dl::base::conv2d<short, short, short, short, int>(void*) at /ho me/yehangyang/Documents/workspace/DL/build/../components/dl/base/dl_base_conv2d. cpp:895

0x42016d70: dl::nn::conv2d(dl::Tensor&, dl::Tensor&, std::vector<i nt, std::allocator >&, dl::Filter const&, int, int, dl::Bias const, dl::Activation const, std::vector<int, std::allocator > con st&) at /home/yehangyang/Documents/workspace/DL/build/../components/dl/nn/src/dl _nn_conv2d.cpp:40

0x42009fc1: dl::layer::Conv2D::call(dl::Tensor&, bool, std::vector <int, std::allocator > const&) at C:\Users\gaochao\eclipse-workspace\ESP32S 3_RTOS_SDK_master\build/../components/esp-dl/include/layer/dl_layer_conv2d.hpp:1 15 (inlined by) ORDER::call(dl::Tensor&) at C:\Users\gaochao\eclipse-worksp ace\ESP32S3_RTOS_SDK_master\build/../main/includes/order_model.hpp:107

0x42016e61: dl::layer::Model::forward(dl::Tensor&) at /home/yehang yang/Documents/workspace/DL/build/../components/dl/layer/src/dl_layer_model.cpp: 30

0x4200b198: app_recon at C:\Users\gaochao\eclipse-workspace\ESP32S3_RTOS_SDK_mas ter\build/../main/app_order_recon.cpp:31

0x420093b2: task_app_recon_process(void*) at C:\Users\gaochao\eclipse-workspace\ ESP32S3_RTOS_SDK_master\build/../main/app_recon_process.cpp:123

0x4038215d: vPortTaskWrapper at C:/Users/gaochao/Desktop/esp-idf/components/free rtos/port/xtensa/port.c:168

[yehangyang]

@xiao-mb

uint8_t img6464 = (uint8_t )dl::tool::malloc_aligned(64 64 3, sizeof(uint8_t), 16);

[xiao-mb]

修改之后，还是同样的错误

[yehangyang]

(inlined by) ORDER::call(dl::Tensor&) at C:\Users\gaochao\eclipse-worksp
ace\ESP32S3_RTOS_SDK_master\build/../main/includes/order_model.hpp:107

这一行指向的哪一行代码？

[xiao-mb]

this->l3_exp.call(this->l3_dp.output);

[yehangyang]

typo here:

l3_dp(DepthwiseConv2D(-16, get_l3_dp_filter(), NULL, get_l3_dp_activation(), PADDING_SAME, 1, 1, "l3depth")), l3_exp(Conv2D(-13, get_l3_exp_filter(), NULL, NULL, PADDING_SAME, 1, 1, "l3_a_compress")), l4_dp(DepthwiseConv2D(-15, get_l3_dp_filter(), NULL, get_l3_dp_activation(), PADDING_SAME, 2, 2, "l4_depth")),

不止这两个，你再自己查一查。

[yehangyang]

@xiao-mb 如果问题已解决，请关闭 issue。