ggml-hexagon: setting enable_rpc_ion_mempool to 1 and make test-backend-ops happy

Author: zhouwg

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -359,7 +359,7 @@ static struct hexagon_appcfg_t g_hexagon_appcfg = {
 #elif defined(_WIN32)
         .qnn_runtimelib_path    = "C:\\",
 #endif
-        .ggml_hexagon_version   = {"1.03"},
+        .ggml_hexagon_version   = {"1.04"},
         .ggml_dsp_version       = {"0.61"},
 };
 
@@ -5107,9 +5107,8 @@ static int ggmlhexagon_init_rpcmempool(ggml_backend_hexagon_context * ctx) {
         GGML_ASSERT(ctx->rpc_mempool_capacity > (8 * SIZE_IN_MB));
         ctx->rpc_mempool_len = ctx->rpc_mempool_capacity - (8 * SIZE_IN_MB);
 
-        //FIXME: it seems there is unknown issue with another ION memory pool
-        ctx->rpc_mempool = rpcmem_alloc(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS,
-                                        ctx->rpc_mempool_len);
+        //FIXME: it seems there is unknown issue with 2+ GiB memory pool
+        ctx->rpc_mempool = rpcmem_alloc(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_TRY_MAP_STATIC, ctx->rpc_mempool_len);
         if (nullptr == ctx->rpc_mempool) {
             GGMLHEXAGON_LOG_WARN("alloc rpc memorypool %d failed", ctx->rpc_mempool_len);
             return 2;

ggml/src/ggml-hexagon/kernels/ggml-dsp.c

@@ -888,8 +888,7 @@ int ggmlop_dsp_open(const char*uri, remote_handle64* handle) {
 
     ggml_init();
 
-    unsigned int api_version = qurt_api_version();
-    GGMLHEXAGON_LOG_DEBUG("api_version = 0x%x", api_version);
+    GGMLHEXAGON_LOG_DEBUG("api_version = 0x%x", qurt_api_version());
     GGMLHEXAGON_LOG_DEBUG("hvx units = 0x%d", qurt_hvx_get_units());
     qurt_arch_version_t  vers;
     qurt_sysenv_get_arch_version(&vers);

scripts/ggml-hexagon.cfg

@@ -23,7 +23,7 @@
 #
 [general]
 #version of ggml-hexagon.cpp on ARM-AP side
-version = "1.03"
+version = "1.04"
 #version of ggml-dsp.c on cDSP side
 ggmldsp_version = "0.61"
 
@@ -69,8 +69,8 @@ enable_perf = 1
 
 
 # enablie/disable profiler feature to visually compare NPU performance between HWACCEL_CDSP and HWACCEL_QNN
-# this is not default setting but useful before PR in upstream can be approved
-enable_profiler = 1
+# this is default setting
+enable_profiler = 0
 #threshold duration of NPU performance profiler, per seconds
 profiler_duration = 5
 #threshold counst of NPU performance profiler
@@ -94,7 +94,7 @@ precision_mode = "fp16"
 #hwaccel approach through cDSP(offload ggml op to Hexagon cDSP directly)
 [cdsp]
 #enable/disable rpc ion memory pool
-enable_rpc_ion_mempool = 0
+enable_rpc_ion_mempool = 1
 
 #enable/disable offload all quantized type mulmat to cDSP
 enable_all_q_mulmat = 0