ggml-qnn: enable release build with necessary logs to make reviewers happy

Author: zhouwg

ggml/src/ggml-qnn/ggml-qnn-impl.h

@@ -89,7 +89,7 @@ class  qnn_instance;
 struct ggml_backend_qnn_context;
 void   ggmlqnn_log_internal(ggml_log_level level, const char * file, const char * func, int line, const char * format, ...);
 
-#ifdef NDEBUG
+#if 0//def NDEBUG
 #define GGMLQNN_DEBUG                           0
 #define ENABLE_QNNBACKEND_PERF                  0  // enable/disable op's perf info
 #define GGMLQNN_PRINT_QNN_INTERNAL_LOG          0  // enable/disable QNN's internal log

ggml/src/ggml-qnn/ggml-qnn-ops.cpp

@@ -91,9 +91,7 @@ void ggml_qnn_general_node(ggml_backend_qnn_context * ctx, ggml_tensor * op) {
         p_tensor1 = ggmlqnn_create_compute_tensor(src1);
         p_tensor2 = ggmlqnn_create_compute_tensor(dst);
     }
-#if GGMLQNN_PRINT_OP_ADD_LOG
-    print_tensors_info(__func__, ctx, src0, src1, dst);
-#endif
+    ggmlqnn_print_tensors_info(__func__, ctx, src0, src1, dst);
 
     //ensure QNN tensor has correct tensor type
     QNN_VER_PTR(*p_tensor0)->type = QNN_TENSOR_TYPE_APP_WRITE;

ggml/src/ggml-qnn/ggml-qnn.cpp

@@ -2232,7 +2232,7 @@ static bool ggml_qnn_can_handle_op(const ggml_backend_qnn_context * ctx, const s
     }
 
     if (tensor->op == GGML_OP_MUL_MAT) {
-        dump_op_info(tensor);
+        //dump_op_info(tensor);
         if (src0_rank != src1_rank) // make QNN SDK happy
             return false;
         if (src0_rank < 2) // QNN's limitation, make QNN SDK happy
@@ -2244,15 +2244,16 @@ static bool ggml_qnn_can_handle_op(const ggml_backend_qnn_context * ctx, const s
 
         if (ctx->device == QNN_BACKEND_NPU)
             if (2 == src0_rank)
-                return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16
+                return (src0->type == GGML_TYPE_F32
                     || src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q8_0
                     || src0->type == GGML_TYPE_Q6_K || src0->type == GGML_TYPE_Q8_K
                    ) && (src1->type == GGML_TYPE_F32) && (tensor->type == GGML_TYPE_F32);
            else
                 return (src0->type == GGML_TYPE_F32) && (src1->type == GGML_TYPE_F32) && (tensor->type == GGML_TYPE_F32);
         else
-            return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))
-                   && (src1->type == GGML_TYPE_F32) && (tensor->type == GGML_TYPE_F32);
+            return (src0->type == GGML_TYPE_F32   || src0->type == GGML_TYPE_Q4_0
+                    || src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_Q6_K || src0->type == GGML_TYPE_Q8_K)
+                    && (src1->type == GGML_TYPE_F32) && (tensor->type == GGML_TYPE_F32);
     }
 
     if (tensor->op == GGML_OP_MUL) {

scripts/build-run-android.sh

@@ -74,7 +74,7 @@ function check_and_download_ndk()
 
 function build_arm64
 {
-    cmake -H. -B./out/android -DCMAKE_BUILD_TYPE=Debug -DGGML_USE_QNN=ON -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=latest -DCMAKE_C_FLAGS=-march=armv8.7-a -DGGML_QNN=ON -DGGML_QNN_SDK_PATH=${QNN_SDK_PATH}
+    cmake -H. -B./out/android -DCMAKE_BUILD_TYPE=Release -DGGML_USE_QNN=ON -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=latest -DCMAKE_C_FLAGS=-march=armv8.7-a -DGGML_QNN=ON -DGGML_QNN_SDK_PATH=${QNN_SDK_PATH}
     cd out/android
     make -j16
     show_pwd