HexagonOffload needs to call inject_hvx_lock_unlock() before lower_parallel_tasks()

src/CodeGen_Hexagon.cpp

@@ -3,6 +3,7 @@
 
 #include "AlignLoads.h"
 #include "CSE.h"
+#include "CodeGen_Hexagon.h"
 #include "CodeGen_Internal.h"
 #include "CodeGen_Posix.h"
 #include "Debug.h"
@@ -437,17 +438,6 @@ class InjectHVXLocks : public IRMutator {
  Target target;
 };
 
-Stmt inject_hvx_lock_unlock(Stmt body, const Target &target) {
- InjectHVXLocks i(target);
- body = i.mutate(body);
- if (i.uses_hvx) {
- body = acquire_hvx_context(body, target);
- }
- body = substitute("uses_hvx", i.uses_hvx, body);
- body = simplify(body);
- return body;
-}
-
 void CodeGen_Hexagon::compile_func(const LoweredFunc &f,
  const string &simple_name,
  const string &extern_name) {
@@ -495,6 +485,8 @@ void CodeGen_Hexagon::compile_func(const LoweredFunc &f,
  debug(1) << "Optimizing Hexagon instructions...\n";
  body = optimize_hexagon_instructions(body, target);
 
+ debug(1) << "Before qurt_hvx_lock:\n";
+ debug(1) << body << "\n";
  debug(1) << "Adding calls to qurt_hvx_lock, if necessary...\n";
  body = inject_hvx_lock_unlock(body, target);
 
@@ -2245,13 +2237,29 @@ std::unique_ptr<CodeGen_Posix> new_CodeGen_Hexagon(const Target &target) {
  return std::make_unique<CodeGen_Hexagon>(target);
 }
 
+Stmt inject_hvx_lock_unlock(Stmt body, const Target &target) {
+ InjectHVXLocks i(target);
+ body = i.mutate(body);
+ if (i.uses_hvx) {
+ body = acquire_hvx_context(body, target);
+ }
+ body = substitute("uses_hvx", i.uses_hvx, body);
+ body = simplify(body);
+ return body;
+}
+
 #else // WITH_HEXAGON
 
 std::unique_ptr<CodeGen_Posix> new_CodeGen_Hexagon(const Target &target) {
  user_error << "hexagon not enabled for this build of Halide.\n";
  return nullptr;
 }
 
+Stmt inject_hvx_lock_unlock(Stmt body, const Target &target) {
+ user_error << "hexagon not enabled for this build of Halide.\n";
+ return Stmt();
+}
+
 #endif // WITH_HEXAGON
 
 } // namespace Internal

src/CodeGen_Hexagon.h

@@ -0,0 +1,17 @@
+#ifndef HALIDE_CODEGEN_HEXAGON_H
+#define HALIDE_CODEGEN_HEXAGON_H
+
+#include "Expr.h"
+
+namespace Halide {
+
+struct Target;
+
+namespace Internal {
+
+Stmt inject_hvx_lock_unlock(Stmt body, const Target &target);
+
+} // namespace Internal
+} // namespace Halide
+
+#endif

src/HexagonOffload.cpp

@@ -2,6 +2,7 @@
 #include <memory>
 
 #include "Closure.h"
+#include "CodeGen_Hexagon.h"
 #include "Elf.h"
 #include "HexagonOffload.h"
 #include "IRMutator.h"
@@ -286,7 +287,7 @@ void do_reloc(char *addr, uint32_t mask, uintptr_t val, bool is_signed, bool ver
  // Pull out the subinstructions. They're the low 13
  // bits of each half-word.
  uint32_t hi = (inst >> 16) & ((1 << 13) - 1);
- //uint32_t lo = inst & ((1 << 13) - 1);
+ // uint32_t lo = inst & ((1 << 13) - 1);
 
  // We only understand the ones where hi starts with 010
  internal_assert((hi >> 10) == 2);
@@ -753,6 +754,12 @@ class InjectHexagonRpc : public IRMutator {
  DeviceAPI::None, loop->body);
  }
 
+ debug(2) << "Before inject_hvx_lock_unlock:\n"
+ << body << "\n";
+ body = inject_hvx_lock_unlock(body, device_code.target());
+ debug(2) << "After inject_hvx_lock_unlock:\n"
+ << body << "\n";
+
  // Build a closure for the device code.
  // Note that we must do this *before* calling lower_parallel_tasks();
  // otherwise the Closure may fail to find buffers that are referenced