release: v1.3.11 (#231)

.gitignore

@@ -56,6 +56,9 @@ tags
 out
 out.*
 build
+android-ndk-*
+ggml-qnn-ut
+ggml-qnn-test
 
 *.mp4
 *.ts

VERSION.txt

@@ -1 +1 @@
-kantv 1.3.10, based on customized FFmpeg 6.1, ggml, whisper.cpp(ASR), llama.cpp(LLM), ncnn
+kantv 1.3.11, based on customized FFmpeg 6.1, ggml, whisper.cpp(ASR), llama.cpp(LLM), ncnn

cdeosplayer/cdeosplayer-lib/src/main/java/cdeos/media/player/CDEUtils.java

@@ -141,7 +141,7 @@ public class CDEUtils {
      private static String mApiGatewayServerUrl = "http://www.cdeos.com:8888/wiseplay/getlicense";
      private static String mLocalEMS = "http://192.168.0.200:81/ems";
 
-     private static String mKANTVAPKVersion = "1.3.10";
+     private static String mKANTVAPKVersion = "1.3.11";
      private static KANTVDRM mKANTVDRM = KANTVDRM.getInstance();
 
      public static final String INVALID_DEVICE_ID = "000000000000000";

cdeosplayer/constants.gradle

@@ -44,8 +44,8 @@ project.ext {
     appTargetSdkVersion = 29
 
     //version and version code of KanTV
-    releaseVersion     = '1.3.10'
-    releaseVersionCode = 1000310
+    releaseVersion     = '1.3.11'
+    releaseVersionCode = 1000311
 
     modulePrefix = ':'
     guavaVersion = '27.1-android'

cdeosplayer/kantv/src/main/assets/config.json

@@ -1,6 +1,6 @@
 {
   "kantvServer": "www.cdeos.com",
   "releaseMode": "0",
-  "apkVersion": "1.3.10",
+  "apkVersion": "1.3.11",
   "apkForTV": "0"
 }

core/ggml/llamacpp/ggml-qnn.cpp

@@ -5,7 +5,7 @@
  *
  *
  * IPR statement:
- * The QNN helper macros and functions are referenced from:
+ * The QNN helper macros and functions are referenced(reverse engineering by many experiments) from:
  * (1) https://github.com/pytorch/executorch/tree/main/backends/qualcomm  (provided by Qualcomm Technologies, Inc.)
  * (2) QNN samples (Qualcomm Technologies, Inc.)
  * (3) /opt/qcom/aistack/qnn/2.20.0.240223/examples/Models/InceptionV3/model/Inception_v3.cpp which is
@@ -16,22 +16,28 @@
  *
  *
  * status:
- * 1. core implementation(data path works fine as expected with whisper.cpp using QNN CPU/GPU backend)
+ * 1. core implementation(data path works fine as expected with whisper.cpp&llama.cpp using QNN CPU/GPU backend)
  *    on Qualcomm's SoC based low-end phone
- * 2. core implementation(data path works fine as expected with whisper.cpp using QNN NPU(aka HTP/DSP)
+ * 2. core implementation(data path works fine as expected with whisper.cpp&llama.cpp using QNN NPU(aka HTP/DSP)
  *    backend on Qualcomm's soC based high-end phone
  * 3. GGML_OP_MUL_MAT & GGML_OP_MUL & GGML_OP_ADD using QNN API has been completed and the dedicated
- *    Android command line UT program passed on Qualcomm's SoC based Andriod phone
+ *    Android command line UT program works fine as expected on Qualcomm's SoC based Android phone
  * 4. PR to upstream GGML community on 04-24-2024: https://github.com/ggerganov/llama.cpp/pull/6869
  *
  * todo:
  * 1. lack of implementation of other GGML OPs using QNN API(only support GGML_OP_MUL_MAT,
- *    GGML_OP_MUL, GGML_OP_ADD currently), would be done in upstream GGML community
+ *    GGML_OP_MUL, GGML_OP_ADD currently). this problem has been done by s standalone PR
+ *    https://github.com/zhouwg/kantv/pull/216 in this project or a standalone PR
+ *    https://github.com/ggerganov/llama.cpp/pull/7641 in upstream.
+ *    it's a general approach for mixed inference between Qualcomm's CPU&GPU / CPU&NPU very
+ *    easily and re-use/refine the existing GGML backend subsystem and no any side-effect to
+ *    any existing backends/codes.
  * 2. only support FP32 / FP16, other(quantized) GGML data type not used currently, data type of
  *    input tensor and output tensor must be same(this is a real big limitation in this backend).
- *    would be done in upstream GGML community
- * 3. QNN's RPC feature(which is required for QNN NPU backend) not used
- * 4. performance fine-tune using QNN backend(mixed inference between CPU&GPU&NPU)
+ *    would be done in upstream GGML community if the PR of ggml-qnn-backend in upstream could be accepted.
+ * 3. QNN's RPC feature(which is required for QNN NPU backend) not used,would be done in upstream
+ *    GGML community if the PR of ggml-qnn-backend in upstream could be accepted.
+ * 4. performance fine-tune(long-term task)
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -3175,6 +3181,10 @@ GGML_CALL static void ggml_backend_qnn_buffer_init_tensor(ggml_backend_buffer_t
     Qnn_ErrorHandle_t error = QNN_SUCCESS;
     ggml_backend_qnn_buffer_context * ctx = (ggml_backend_qnn_buffer_context *)buffer->context;
 
+    static int idx = 0;
+    char tensor_name[GGML_MAX_NAME] = { 0 };
+    snprintf(tensor_name, GGML_MAX_NAME, "tensor_%2d", idx++);
+
     uint32_t dimensions[] = {(uint32_t) tensor->ne[0], (uint32_t) tensor->ne[1], (uint32_t) tensor->ne[2], (uint32_t) tensor->ne[3]};
     //TODO:only support FP32 & FP16
     Qnn_DataType_t  qnn_data_type = QNN_DATATYPE_FLOAT_32;
@@ -3189,7 +3199,7 @@ GGML_CALL static void ggml_backend_qnn_buffer_init_tensor(ggml_backend_buffer_t
             .version= QNN_TENSOR_VERSION_1,
             {.v1= {
                     .id = 0,
-                    .name = tensor->name,
+                    .name = tensor_name,
                     .type = qnn_tensor_type,
                     .dataFormat = QNN_TENSOR_DATA_FORMAT_FLAT_BUFFER,
                     .dataType = qnn_data_type,

core/ggml/llamacpp/tests/ggml-qnn-ut-in-kantv/test-qnn-ops.cpp

@@ -569,7 +569,7 @@ static int qnn_op_ut_automation(int num_threads, int n_backend_type, int n_ggml_
         s += strbuf;
 
 
-        QNN_LOG_DEBUG("%s\n", s.c_str());
+        QNN_LOG_DEBUG("\n%s\n", s.c_str());
     }
 
     n_end_time = ggml_time_us();

core/ggml/llamacpp/tests/ggml-qnn/CMakeLists.txt

@@ -24,20 +24,18 @@ set(SOURCE_FILES
         ../../ggml-quants.c
         ../../ggml-qnn.cpp
         ../../../../../external/ffmpeg/libavutil/cde_log.c
-        test-qnn-ops.cpp
+        ggml-qnn-ut.cpp
 )
 
 
 message("QNN_SDK_PATH         : ${QNN_SDK_PATH}")
 message("QNN_INC_PATH         : ${QNN_INC_PATH}")
 message("QNN_LIB_PATH         : ${QNN_LIB_PATH}")
-message("target name          : ${TARGET_NAME}")
-
 
 add_definitions(-D__ARM_NEON)
 add_definitions(-DGGML_USE_QNN)
 
-if(CMAKE_BUILD_TYPE STREQUAL "release")
+if(CMAKE_BUILD_TYPE STREQUAL "Release")
 add_definitions(-DNDEBUG)
 add_definitions(-O3)
 endif()