ggml-dsp: probe QuRT RTOS information in function ggmlop_dsp_open

Author: zhouwg

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

@@ -1872,8 +1872,10 @@ static void ggmlhexagon_print_running_timestamp(ggml_backend_hexagon_context * c
     if (HWACCEL_CDSP == g_hexagon_appcfg.hwaccel_approach) {
         GGMLHEXAGON_LOG_INFO("offload quantize GGML_OP_MUL_MAT: %s", g_hexagon_appcfg.enable_q_mulmat ? "YES" : "NO");
         GGMLHEXAGON_LOG_INFO("using rpc ion memory pool:        %s", g_hexagon_appcfg.enable_rpc_ion_mempool ? "YES" : "NO");
+        GGMLHEXAGON_LOG_INFO("thread_counts with HWACCEL_CDSP: %d", g_hexagon_appcfg.thread_counts);
         ggmlhexagon_probe_dspinfo(ctx);
     } else {
+        GGMLHEXAGON_LOG_INFO("thread_counts with HWACCEL_QNN: %d", g_hexagon_appcfg.hvx_threads);
         GGMLHEXAGON_LOG_INFO("offload quantize GGML_OP_MUL_MAT: %s", g_hexagon_appcfg.enable_q_mulmat ? "YES" : "NO");
     }
     GGMLHEXAGON_LOG_INFO("running timestamp:%s", timestamp);

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

@@ -32,6 +32,8 @@
 #include "HAP_vtcm_mgr.h"
 #include "HAP_compute_res.h"
 
+#include "qurt.h"
+
 #include "AEEStdErr.h"
 #include "hexagon_types.h"
 #include "hexagon_protos.h"
@@ -886,6 +888,19 @@ 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("hvx units = 0x%d", qurt_hvx_get_units());
+    qurt_arch_version_t  vers;
+    qurt_sysenv_get_arch_version(&vers);
+    GGMLHEXAGON_LOG_DEBUG("arch_version=0x%x", vers.arch_version);
+    qurt_sysenv_app_heap_t aheap;
+    qurt_sysenv_get_app_heap(&aheap);
+    GGMLHEXAGON_LOG_DEBUG("aheap.heap_base=0x%x, aheap.heap_limit=0x%x", aheap.heap_base, aheap.heap_limit);
+    qurt_sysenv_max_hthreads_t mhwt;
+    qurt_sysenv_get_max_hw_threads(&mhwt);
+    GGMLHEXAGON_LOG_DEBUG("max hardware threads=%d", mhwt.max_hthreads);
+
     return 0;
 }
 

scripts/ggml-hexagon.cfg

@@ -85,7 +85,7 @@ profiler_counts = 200
 # enable/disable QNN SDK's internal log, this will very helpful for troubleshooting in HWACCEL_QNN approach
 print_qnn_internal_log = 0
 
-hvx_threads = 4
+hvx_threads = 8
 vtcm_size_in_mb = 8
 enable_dlbc = 1
 precision_mode = "fp16"
@@ -102,9 +102,7 @@ enable_all_q_mulmat = 0
 #ensure enable_q_mulmat = 1 when set enable_all_q_mulmat = 1
 
 #enable/disable multi-threading on cDSP side
-# 0  disable multi-threading on cDSP side
-# 1  disable multi-threading on cDSP side
-# 2  setting thread_counts to 2 on cDSP side
-# 3  setting thread_counts to 3 on cDSP side
-# 4  setting thread_counts to 4 on cDSP side
+# 0    disable multi-threading on cDSP side
+# 1    disable multi-threading on cDSP side
+# 2-8  thread_counts on cDSP side
 thread_counts = 1