RK3588: NPU: Update to 0.9.6 version

external/config/boards/orangepi5pro.conf

@@ -12,4 +12,4 @@ SKIP_BOOTSPLASH="yes" # Skip boot splash patch, conflicts with CONFIG_VT=yes
 DISTRIB_TYPE_LEGACY="bullseye bookworm bionic focal jammy"
 DISTRIB_TYPE_CURRENT="bullseye bookworm jammy"
 BOOTFS_TYPE="fat"
-REVISION="1.0.2"
+REVISION="1.0.4"

external/config/kernel/linux-rockchip-rk3588-current.config

@@ -4378,7 +4378,7 @@ CONFIG_VIDEO_GC8034=y
 # CONFIG_VIDEO_HI847 is not set
 # CONFIG_VIDEO_IMX208 is not set
 # CONFIG_VIDEO_IMX214 is not set
-# CONFIG_VIDEO_IMX219 is not set
+CONFIG_VIDEO_IMX219=y
 # CONFIG_VIDEO_IMX258 is not set
 # CONFIG_VIDEO_IMX274 is not set
 # CONFIG_VIDEO_IMX290 is not set
@@ -4447,7 +4447,7 @@ CONFIG_VIDEO_OV4689=y
 CONFIG_VIDEO_OV50C40=y
 # CONFIG_VIDEO_OV5640 is not set
 # CONFIG_VIDEO_OV5645 is not set
-# CONFIG_VIDEO_OV5647 is not set
+CONFIG_VIDEO_OV5647=y
 # CONFIG_VIDEO_OV5648 is not set
 # CONFIG_VIDEO_OV5670 is not set
 # CONFIG_VIDEO_OV5675 is not set

external/packages/bsp/common/usr/lib/orangepi/orangepi-hardware-optimization

@@ -207,6 +207,7 @@ prepare_board() {
 			elif [[ $BOARD =~ orangepi5max|orangepi5ultra ]]; then
 
 				rfkill unblock all
+				sleep 2
 				brcm_patchram_plus --bd_addr_rand --enable_hci --no2bytes --use_baudrate_for_download --tosleep 200000 \
 				        --baudrate 1500000 --patchram /lib/firmware/SYN43711A0.hcd /dev/ttyS7 &
 			fi

external/packages/bsp/rk3588/usr/include/rkllm.h

@@ -5,45 +5,115 @@
 extern "C" {
 #endif
 
-typedef void* LLMHandle;
+typedef void* LLMHandle;        /* Handle for an instance of a language model. */
 
+/**
+ * @brief Structure for possible states of an inference call.
+ * 
+ */
 typedef enum {
-    LLM_RUN_NORMAL = 0,  /*推理状态正常，推理尚未结束*/
+    LLM_RUN_NORMAL = 0,         /* Inference status is normal and inference has not yet finished. */
-    LLM_RUN_FINISH = 1,  /*推理状态正常，推理结束*/
+    LLM_RUN_FINISH = 1,         /* Inference status is normal and inference has finished. */
-    LLM_RUN_ERROR = 2    /*推理状态异常*/
+    LLM_RUN_ERROR = 2           /* Inference status is abnormal. */
 } LLMCallState;
 
+/**
+ * @brief Structure for setting up parameters for the language model
+ * 
+ */
 typedef struct {
-    const char* modelPath;          /*模型文件的存放路径*/
+    const char* model_path;     /* Path where the model file is located. */
-    const char* target_platform;    /*模型运行的硬件平台*/
+    int32_t num_npu_core;       /* Number of NPU cores used for model inference. */
-    int32_t num_npu_core;           /*模型推理时使用的 NPU 核心数量*/
+    int32_t max_context_len;    /* Maximum size of the context. */
-    int32_t max_context_len;        /*设置提示上下文的大小*/
+    int32_t max_new_tokens;     /* Maximum number of tokens to generate during model inference. */
-    int32_t max_new_tokens;         /*用于设置模型推理时生成 Token 的数量上限*/
+    int32_t top_k;              /* The number of highest probability tokens to consider for generation. */
-
+    float top_p;                /* Nucleus sampling: cumulative probability cutoff to use for token selection. */
-    int32_t top_k;                  
+    float temperature;          /* Hyperparameter to control the randomness of predictions by scaling the logits before applying softmax. */
-    float top_p;
+    float repeat_penalty;       /* Penalty applied to the logits of previously generated tokens, helps prevent repetitive or monotonic text. */
-    float temperature;
+    float frequency_penalty;    /* Penalty for repeating the same word or phrase, reducing the likelihood of repeated content. */
-    float repeat_penalty;
+    float presence_penalty;     /* Penalty or reward for introducing new tokens into the generated text. */
-    float frequency_penalty;
+    int32_t mirostat;           /* Enables mirostat algorithm, where 0 = off, 1 = use mirostat algorithm, 2 = use mirostat 2.0 algorithm. */
-    float presence_penalty;
+    float mirostat_tau;         /* Target entropy (perplexity) for mirostat algorithm, setting the desired complexity of the generated text. */
-    int32_t mirostat;
+    float mirostat_eta;         /* Learning rate for the mirostat algorithm. */
-    float mirostat_tau;
+    bool logprobs;              /* Whether to return the log probabilities for each output token along with their token ids. */
-    float mirostat_eta;
+    int32_t top_logprobs;       /* The number of top tokens for which to return log probabilities, along with their token ids. */
-
+    bool use_gpu;               /* Flag to indicate whether to use GPU for inference. */
 } RKLLMParam;
 
-typedef void(*LLMResultCallback)(const char* result, void* userdata, LLMCallState state);
+/**
+ * @brief Structure representing a token with its associated log probability.
+ * 
+ */
+typedef struct {
+    float logprob;              /* Log probability corresponding to the token ID. */
+    int id;                     /* Token ID. */
+} Token;
 
-RKLLMParam rkllm_createDefaultParam();                                              /*初始化RKLLMParam并设置默认参数*/
+/**
+ * @brief Structure to hold the results from the language model inference, including text and token details.
+ * 
+ */
+typedef struct {
+    const char* text;           /* Decoded text from the inference output. */
+    Token* tokens;              /* Array of Token structures, each containing a log probability and a token ID. */
+    int num;                    /* Number of top tokens returned, typically those with the highest probabilities. */
+} RKLLMResult;
 
-int rkllm_init(LLMHandle* handle, RKLLMParam param, LLMResultCallback callback);    /*模型初始化*/
+/**
+ * @brief Callback function for handling inference results.
+ * 
+ * @param result A pointer to an RKLLMResult struct containing the inference results.
+ * @param userdata A pointer to user-defined function or null if no user function was provided.
+ * @param state The state of the inference process, indicating success, failure, or completion.
+ */
+typedef void(*LLMResultCallback)(RKLLMResult* result, void* userdata, LLMCallState state);
 
-int rkllm_run(LLMHandle handle, const char* prompt, void* userdata);                /*模型推理*/
+/**
+ * @brief Initializes RKLLMParam with default settings.
+ * 
+ * @return RKLLMParam An RKLLMParam struct with default values set.
+ */
+RKLLMParam rkllm_createDefaultParam();
 
-int rkllm_destroy(LLMHandle handle);                                                /*模型释放*/
+/**
+ * @brief Initializes the model with specified parameters.
+ * 
+ * @param handle Pointer to a handle for the language model, which will be initialized by this function.
+ * @param param An RKLLMParam struct containing all the parameters needed for the model.
+ * @param callback A function pointer to the callback that handles the results of the inference.
+ * @return int Returns 0 on success, or a negative error code on failure.
+ */
+int rkllm_init(LLMHandle* handle, RKLLMParam param, LLMResultCallback callback);
+
+/**
+ * @brief Releases the model resources.
+ * 
+ * @param handle The handle to the language model to be destroyed.
+ * @return int Returns 0 on successful release, or a negative error code if an error occurs.
+ */
+int rkllm_destroy(LLMHandle handle);
+
+/**
+ * @brief Runs model inference on the given prompt.
+ * 
+ * @param handle The handle to the initialized language model.
+ * @param prompt The text prompt on which to perform inference.
+ * @param userdata Optional user-defined function that will be passed to the callback.
+ * @return int Returns 0 on success, or a negative error code if an error occurs during inference.
+ */
+int rkllm_run(LLMHandle handle, const char* prompt, void* userdata);
+
+/**
+ * @brief Aborts the current inference process.
+ * 
+ * @param handle The handle to the language model whose inference is to be aborted.
+ * @return int Returns 0 if the process is successfully aborted, or a negative error code
+ *         if no process was running or if the abort fails.
+ */
+int rkllm_abort(LLMHandle handle);
 
 #ifdef __cplusplus
-}
+} //extern "C"
 #endif
 
 #endif