Separate out parallel invokes to C file

lib_tflite_micro/src/par_invoke_funcs.c

@@ -0,0 +1,89 @@
+
+#include "/Users/deepakpanickal/code/ai_tools/.venv/lib/python3.10/site-packages/xmos_ai_tools/runtime/include/lib_tflite_micro/src/thread_call.h"
+// #include "xcore/parallel.h"
+
+
+#if defined(__xcore__) || defined(__riscv_xxcore)
+#include <xcore/parallel.h>
+DECLARE_JOB(main_task, (thread_info_t *, synchronizer_t));
+DECLARE_JOB(client_task, (thread_info_t *, int));
+#if defined(__xcore__)
+asm(".linkset __xcore_ugs_shim_main_task.nstackwords, 1024");
+#endif
+#endif
+
+extern void invoke_subgraph_c_trampoline();
+
+void main_task(thread_info_t *t, synchronizer_t sync) {
+   thread_store_sync(t, sync);
+   invoke_subgraph_c_trampoline();
+}
+
+void client_task(thread_info_t *t, int n) {
+  thread_client(t, n);
+}
+
+void par_invoke_1(thread_info_t *ti) {
+#ifdef __xcore__
+  PAR_JOBS(
+     PJOB(main_task, (ti, PAR_SYNC)));
+#else
+  main_task(ti, 0);
+#endif
+}
+
+void par_invoke_2(thread_info_t *ti) {
+#ifdef __xcore__
+  PAR_JOBS(
+     PJOB(main_task, (ti, PAR_SYNC)),
+     PJOB(client_task, (ti, 0)));
+#else
+  client_task(ti, 0);
+  main_task(ti, 0);
+#endif
+}
+
+void par_invoke_3(thread_info_t *ti) {
+#ifdef __xcore__
+  PAR_JOBS(
+     PJOB(main_task, (ti, PAR_SYNC)),
+     PJOB(client_task, (ti, 0)),
+     PJOB(client_task, (ti, 1)));
+#else
+  client_task(ti, 0);
+  client_task(ti, 1);
+  main_task(ti, 0);
+#endif
+}
+
+void par_invoke_4(thread_info_t *ti) {
+#ifdef __xcore__
+  PAR_JOBS(
+     PJOB(main_task, (ti, PAR_SYNC)),
+     PJOB(client_task, (ti, 0)),
+     PJOB(client_task, (ti, 1)),
+     PJOB(client_task, (ti, 2)));
+#else
+  client_task(ti, 0);
+  client_task(ti, 1);
+  client_task(ti, 2);
+  main_task(ti, 0);
+#endif
+}
+
+void par_invoke_5(thread_info_t *ti) {
+#ifdef __xcore__
+  PAR_JOBS(
+     PJOB(main_task, (ti, PAR_SYNC)),
+     PJOB(client_task, (ti, 0)),
+     PJOB(client_task, (ti, 1)),
+     PJOB(client_task, (ti, 2)),
+     PJOB(client_task, (ti, 3)));
+#else
+  client_task(ti, 0);
+  client_task(ti, 1);
+  client_task(ti, 2);
+  client_task(ti, 3);
+  main_task(ti, 0);
+#endif
+}

tflite_micro_compiler/src/Compiler.cc

@@ -1411,12 +1411,33 @@ TfLiteStatus )"
   return )" << prefix_ << R"(init_with_paging(weights_data_ptr, nullptr);
 }
 
+#ifdef __VX4A__
+#define STACKFUNCTION(FN, BYTES) \
+  asm(".globl " # FN ); \
+  asm(".resource_list_empty " # FN ", \"callees\""); \
+  asm(".resource_list_empty " # FN ", \"tail_callees\""); \
+  asm(".resource_list_empty " # FN ", \"parallel_callees\""); \
+  asm(".resource_const " # FN ", \"stack_frame_bytes\", " # BYTES);
+
+STACKFUNCTION(_Z22model_init_with_pagingPvS_, 1000);
+// STACKFUNCTION(_Z12model_invokev, 1000);
+STACKFUNCTION(__call_exitprocs_impl, 1000);
+STACKFUNCTION(invoke_subgraph_c_trampoline, 1000);
+// STACKFUNCTION(_Z10model_initPv);
+#endif
+
+TfLiteStatus mg_status;
+extern "C" void invoke_subgraph_c_trampoline(){
+  mg_status = mg_InvokeSubgraph(0);
+}
+
+extern "C" void par_invoke_)"
+     << numXCThreads_ << R"((void *thread_info);
+
 )";
 wr<<R"(#pragma stackfunction 1000
 TfLiteStatus )"
      << prefix_ << R"(invoke() {
-  thread_init_)"
-     << numXCThreads_ << R"((&xc_config.thread_info);
 
 #ifdef TFLMC_XCORE_PROFILE
   printf("\n\n\nProfiling invoke()...\n");
@@ -1425,12 +1446,11 @@ TfLiteStatus )"
   op_times_summed = 0;
 #endif
 
-  TfLiteStatus status = mg_InvokeSubgraph(0);
-  if (status != kTfLiteOk) {
-    return status;
+  par_invoke_)"
+     << numXCThreads_ << R"((&xc_config.thread_info);
+  if (mg_status != kTfLiteOk) {
+    return mg_status;
   }
-
-  thread_destroy(&xc_config.thread_info);
 )";
   if (has_xc_conv_ops) {
     wr << R"(