ggml: add hw_accel in data structure

Author: zhouwg

ggml.c

@@ -2195,6 +2195,7 @@ struct ggml_context {
     bool   mem_buffer_owned;
     bool   no_alloc;
     bool   no_alloc_save; // this is used to save the no_alloc state when using scratch buffers
+    bool   use_hwaccel;
 
     int    n_objects;
 
@@ -2754,6 +2755,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         /*.mem_buffer_owned   =*/ params.mem_buffer ? false : true,
         /*.no_alloc           =*/ params.no_alloc,
         /*.no_alloc_save      =*/ params.no_alloc,
+        /*.use_hwaccel        =*/ params.use_hwaccel,
         /*.n_objects          =*/ 0,
         /*.objects_begin      =*/ NULL,
         /*.objects_end        =*/ NULL,

ggml.h

@@ -657,6 +657,7 @@ extern "C" {
         size_t mem_size;   // bytes
         void * mem_buffer; // if NULL, memory will be allocated internally
         bool   no_alloc;   // don't allocate memory for the tensor data
+        bool   use_hwaccel;
     };
 
 

whisper.cpp

@@ -4066,6 +4066,14 @@ static int whisper_has_openvino(void) {
 #endif
 }
 
+static int whisper_has_qnn(void) {
+#ifdef GGML_USE_QNN
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 const char * whisper_print_system_info(void) {
     static std::string s;
 
@@ -4086,7 +4094,9 @@ const char * whisper_print_system_info(void) {
     s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
     s += "CUDA = "      + std::to_string(ggml_cpu_has_cuda())      + " | ";
     s += "COREML = "    + std::to_string(whisper_has_coreml())     + " | ";
-    s += "OPENVINO = "  + std::to_string(whisper_has_openvino())          ;
+    s += "OPENVINO = "  + std::to_string(whisper_has_openvino())   + " | ";
+    s += "QNN = "       + std::to_string(whisper_has_qnn())               ;
+
 
     return s.c_str();
 }
@@ -6518,6 +6528,9 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
                 /*.no_alloc   =*/ false,
             };
 
+#ifdef GGML_USE_QNN
+            gparams.use_hwaccel   = true;
+#endif
             struct ggml_context * ctx0 = ggml_init(gparams);
 
             struct ggml_tensor * a = ggml_new_tensor_2d(ctx0, wtype,         N, N);