Skip to content

Commit

Permalink
Merge pull request halide#2193 from halide/few-buffers
Browse files Browse the repository at this point in the history 
For Hexagon, pack small buffers and scalars into one buffer
  • Loading branch information
Zalman Stern authored Jul 17, 2017
2 parents ec778f9 + 8b7a061 commit 36f34d8
Show file tree
Hide file tree
Showing 11 changed files with 129 additions and 108 deletions.
6 changes: 5 additions & 1 deletion src/runtime/hexagon_remote/Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -106,6 +106,10 @@ bin/%/host_malloc.o: host_malloc.cpp
mkdir -p $(@D)
$(CXX-$*) $(CCFLAGS-$*) -fPIC -c host_malloc.cpp -o $@

bin/%/host_shim.o: host_shim.cpp
mkdir -p $(@D)
$(CXX-$*) $(CCFLAGS-$*) -fPIC -c host_shim.cpp -o $@

bin/%/nearbyint.o: nearbyint.cpp
mkdir -p $(@D)
$(CXX-$*) $(CCFLAGS-$*) -fPIC -c nearbyint.cpp -o $@
Expand All @@ -129,7 +133,7 @@ bin/%/signed_by_debug/libhalide_hexagon_remote_skel.so: bin/%/libhalide_hexagon_
mkdir -p $(@D)
python $(HEXAGON_ELFSIGNER) --no_disclaimer -i $^ -o `dirname $@`

bin/%/libhalide_hexagon_host.so: bin/src/halide_hexagon_remote_stub.c bin/%/host_malloc.o
bin/%/libhalide_hexagon_host.so: bin/src/halide_hexagon_remote_stub.c bin/%/host_malloc.o bin/%/host_shim.o
mkdir -p $(@D)
$(CC-$*) $^ $(CCFLAGS-$*) -Wl,-soname,libhalide_hexagon_host.so -shared -o $@

Expand Down
Binary file modified src/runtime/hexagon_remote/bin/arm-32-android/libhalide_hexagon_host.so
View file
Binary file not shown.
Binary file modified src/runtime/hexagon_remote/bin/arm-64-android/libhalide_hexagon_host.so
View file
Binary file not shown.
4 changes: 3 additions & 1 deletion src/runtime/hexagon_remote/bin/src/halide_hexagon_remote.h
View file
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
#ifndef _HALIDE_HEXAGON_REMOTE_H
#define _HALIDE_HEXAGON_REMOTE_H
#include "AEEStdDef.h"
#ifndef __QAIC_HEADER
#define __QAIC_HEADER(ff) ff
#endif //__QAIC_HEADER
Expand Down Expand Up @@ -33,11 +34,12 @@ struct _halide_hexagon_remote_buffer__seq_octet {
int dataLen;
};
typedef unsigned int halide_hexagon_remote_handle_t;
typedef uint64 halide_hexagon_remote_scalar_t;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_initialize_kernels_v3)(const unsigned char* code, int codeLen, halide_hexagon_remote_handle_t* module_ptr) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_get_symbol_v4)(halide_hexagon_remote_handle_t module_ptr, const char* name, int nameLen, halide_hexagon_remote_handle_t* sym_ptr) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_power_hvx_on)(void) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_power_hvx_off)(void) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_run)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, const halide_hexagon_remote_buffer* input_scalars, int input_scalarsLen) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_run_v2)(halide_hexagon_remote_handle_t module_ptr, halide_hexagon_remote_handle_t symbol, const halide_hexagon_remote_buffer* input_buffers, int input_buffersLen, halide_hexagon_remote_buffer* output_buffers, int output_buffersLen, const halide_hexagon_remote_scalar_t* scalars, int scalarsLen) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_release_kernels_v2)(halide_hexagon_remote_handle_t module_ptr) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_poll_log)(char* log, int logLen, int* read_size) __QAIC_HEADER_ATTRIBUTE;
__QAIC_HEADER_EXPORT int __QAIC_HEADER(halide_hexagon_remote_poll_profiler_state)(int* func, int* threads) __QAIC_HEADER_ATTRIBUTE;
Expand Down
72 changes: 31 additions & 41 deletions src/runtime/hexagon_remote/bin/src/halide_hexagon_remote_skel.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -418,15 +418,15 @@ struct Interface {
#define __QAIC_SLIM_EXPORT
#endif

static const Type types[2];
static const Type types[3];
static const SequenceType sequenceTypes[1] = {{&(types[1]),0x0,0x4,0x4,0x0}};
static const Type types[2] = {{0x1,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x1},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8)}};
static const Parameter parameters[8] = {{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8),0,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x4,3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x4,0,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(sequenceTypes[0]),0}}, 25,SLIM_IFPTR32(0x4,0x8),0,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(sequenceTypes[0]),0}}, 25,SLIM_IFPTR32(0x4,0x8),3,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8),3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)1}}, 2,0x4,3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)1}}, 2,0x4,0,0}};
static const Parameter* const parameterArrays[20] = {(&(parameters[7])),(&(parameters[2])),(&(parameters[2])),(&(parameters[7])),(&(parameters[2])),(&(parameters[2])),(&(parameters[7])),(&(parameters[7])),(&(parameters[2])),(&(parameters[2])),(&(parameters[3])),(&(parameters[4])),(&(parameters[3])),(&(parameters[2])),(&(parameters[0])),(&(parameters[1])),(&(parameters[6])),(&(parameters[6])),(&(parameters[5])),(&(parameters[6]))};
static const Type types[3] = {{0x1,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x1},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8)},{0x8,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x8}};
static const Parameter parameters[9] = {{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8),0,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x4,3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)0}}, 2,0x4,0,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(sequenceTypes[0]),0}}, 25,SLIM_IFPTR32(0x4,0x8),0,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(sequenceTypes[0]),0}}, 25,SLIM_IFPTR32(0x4,0x8),3,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[2]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8),0,0},{SLIM_IFPTR32(0x8,0x10),{{(const uintptr_t)&(types[0]),(const uintptr_t)0x0}}, 9,SLIM_IFPTR32(0x4,0x8),3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)1}}, 2,0x4,3,0},{0x4,{{(const uintptr_t)0,(const uintptr_t)1}}, 2,0x4,0,0}};
static const Parameter* const parameterArrays[20] = {(&(parameters[8])),(&(parameters[2])),(&(parameters[2])),(&(parameters[8])),(&(parameters[2])),(&(parameters[2])),(&(parameters[8])),(&(parameters[8])),(&(parameters[2])),(&(parameters[2])),(&(parameters[3])),(&(parameters[4])),(&(parameters[5])),(&(parameters[2])),(&(parameters[0])),(&(parameters[1])),(&(parameters[7])),(&(parameters[7])),(&(parameters[6])),(&(parameters[7]))};
static const Method methods[9] = {{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x4,0x4,3,2,(&(parameterArrays[14])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x2,0x1,0x0,0x0),0x8,0x4,4,3,(&(parameterArrays[13])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x0,0x0,0x0),0x0,0x0,0,0,0,0x0,0x0},{REMOTE_SCALARS_MAKEX(0,0,255,255,15,15),0x14,0x0,9,5,(&(parameterArrays[8])),0x4,0x1},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[1])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x2,0x0,0x0),0x4,0x4,4,2,(&(parameterArrays[18])),0x4,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x0,0x1,0x0,0x0),0x0,0x8,2,2,(&(parameterArrays[16])),0x1,0x4},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x4,0x0,1,1,(&(parameterArrays[0])),0x4,0x0},{REMOTE_SCALARS_MAKEX(0,0,0x1,0x0,0x0,0x0),0x20,0x0,8,8,(&(parameterArrays[0])),0x4,0x0}};
static const Method* const methodArrays[10] = {&(methods[0]),&(methods[1]),&(methods[2]),&(methods[2]),&(methods[3]),&(methods[4]),&(methods[5]),&(methods[6]),&(methods[7]),&(methods[8])};
static const char strings[349] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0set_performance\0output_buffers\0mipsPerThread\0input_scalars\0input_buffers\0power_hvx_off\0get_symbol_v4\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0sym_ptr\0symbol\0func\0name\0code\0run\0";
static const uint16_t methodStrings[34] = {121,289,152,269,247,103,22,235,239,345,258,323,180,137,166,208,258,335,315,64,330,307,298,303,279,0,340,258,43,59,84,258,194,222};
static const char strings[346] = "initialize_kernels_v3\0busbwUsagePercentage\0set_performance_mode\0poll_profiler_state\0release_kernels_v2\0bwMegabytesPerSec\0set_performance\0output_buffers\0mipsPerThread\0input_buffers\0power_hvx_off\0get_symbol_v4\0power_hvx_on\0set_latency\0set_bus_bw\0module_ptr\0mipsTotal\0read_size\0set_mips\0poll_log\0threads\0scalars\0sym_ptr\0symbol\0run_v2\0func\0name\0code\0";
static const uint16_t methodStrings[34] = {121,275,152,255,233,103,22,221,225,324,244,317,166,137,301,194,244,336,309,64,331,293,284,289,265,0,341,244,43,59,84,244,180,208};
static const uint16_t methodStringsArrays[10] = {25,15,33,32,9,30,22,19,28,0};
__QAIC_SLIM_EXPORT const Interface __QAIC_SLIM(halide_hexagon_remote_slim) = {10,&(methodArrays[0]),0,0,&(methodStringsArrays [0]),methodStrings,strings};
#endif //_HALIDE_HEXAGON_REMOTE_SLIM_H
Expand Down Expand Up @@ -522,63 +522,63 @@ static __inline int _skel_method_3(int (*_pfn)(char*, uint32_t, uint32_t*), uint
_CATCH(_nErr) {}
return _nErr;
}
static __inline int _skel_pack(remote_arg* _praROutPost, remote_arg* _ppraROutPost[1], void* _primROut, char* _in0[1], uint32_t _in0Len[1]) {
static __inline int _skel_pack(remote_arg* _praROutPost, remote_arg* _ppraROutPost[1], void* _primROut, char* _rout0[1], uint32_t _rout0Len[1]) {
int _nErr = 0;
remote_arg* _praROutPostStart = _praROutPost;
remote_arg** _ppraROutPostStart = _ppraROutPost;
_ppraROutPost = &_praROutPost;
_ppraROutPostStart[0] += (_praROutPost - _praROutPostStart) +0;
_ppraROutPostStart[0] += (_praROutPost - _praROutPostStart) +1;
return _nErr;
}
static __inline int _skel_pack_1(remote_arg* _praROutPost, remote_arg* _ppraROutPost[1], void* _primROut, char* _rout0[1], uint32_t _rout0Len[1]) {
static __inline int _skel_pack_1(remote_arg* _praROutPost, remote_arg* _ppraROutPost[1], void* _primROut, char* _in0[1], uint32_t _in0Len[1]) {
int _nErr = 0;
remote_arg* _praROutPostStart = _praROutPost;
remote_arg** _ppraROutPostStart = _ppraROutPost;
_ppraROutPost = &_praROutPost;
_ppraROutPostStart[0] += (_praROutPost - _praROutPostStart) +1;
_ppraROutPostStart[0] += (_praROutPost - _praROutPostStart) +0;
return _nErr;
}
static __inline int _skel_unpack(_allocator* _al, remote_arg* _praIn, remote_arg* _ppraIn[1], remote_arg* _praROut, remote_arg* _ppraROut[1], void* _primIn, void* _primROut, char* _in0[1], uint32_t _in0Len[1]) {
static __inline int _skel_unpack(_allocator* _al, remote_arg* _praIn, remote_arg* _ppraIn[1], remote_arg* _praROut, remote_arg* _ppraROut[1], void* _primIn, void* _primROut, char* _rout0[1], uint32_t _rout0Len[1]) {
int _nErr = 0;
remote_arg* _praInStart = _praIn;
remote_arg** _ppraInStart = _ppraIn;
remote_arg* _praROutStart = _praROut;
remote_arg** _ppraROutStart = _ppraROut;
_ppraIn = &_praIn;
_ppraROut = &_praROut;
_COPY(_in0Len, 0, _primIn, 0, 4);
_ASSERT(_nErr, (int)((_praIn[0].buf.nLen / 1)) >= (int)(_in0Len[0]));
_in0[0] = _praIn[0].buf.pv;
_ppraInStart[0] += (_praIn - _praInStart) + 1;
_ppraROutStart[0] += (_praROut - _praROutStart) +0;
_COPY(_rout0Len, 0, _primIn, 0, 4);
_ASSERT(_nErr, (int)((_praROut[0].buf.nLen / 1)) >= (int)(_rout0Len[0]));
_rout0[0] = _praROut[0].buf.pv;
_ppraInStart[0] += (_praIn - _praInStart) + 0;
_ppraROutStart[0] += (_praROut - _praROutStart) +1;
_CATCH(_nErr) {}
return _nErr;
}
static __inline int _skel_unpack_1(_allocator* _al, remote_arg* _praIn, remote_arg* _ppraIn[1], remote_arg* _praROut, remote_arg* _ppraROut[1], void* _primIn, void* _primROut, char* _rout0[1], uint32_t _rout0Len[1]) {
static __inline int _skel_unpack_1(_allocator* _al, remote_arg* _praIn, remote_arg* _ppraIn[1], remote_arg* _praROut, remote_arg* _ppraROut[1], void* _primIn, void* _primROut, char* _in0[1], uint32_t _in0Len[1]) {
int _nErr = 0;
remote_arg* _praInStart = _praIn;
remote_arg** _ppraInStart = _ppraIn;
remote_arg* _praROutStart = _praROut;
remote_arg** _ppraROutStart = _ppraROut;
_ppraIn = &_praIn;
_ppraROut = &_praROut;
_COPY(_rout0Len, 0, _primIn, 0, 4);
_ASSERT(_nErr, (int)((_praROut[0].buf.nLen / 1)) >= (int)(_rout0Len[0]));
_rout0[0] = _praROut[0].buf.pv;
_ppraInStart[0] += (_praIn - _praInStart) + 0;
_ppraROutStart[0] += (_praROut - _praROutStart) +1;
_COPY(_in0Len, 0, _primIn, 0, 4);
_ASSERT(_nErr, (int)((_praIn[0].buf.nLen / 1)) >= (int)(_in0Len[0]));
_in0[0] = _praIn[0].buf.pv;
_ppraInStart[0] += (_praIn - _praInStart) + 1;
_ppraROutStart[0] += (_praROut - _praROutStart) +0;
_CATCH(_nErr) {}
return _nErr;
}
static __inline int _skel_method_4(int (*_pfn)(uint32_t, uint32_t, void*, uint32_t, void*, uint32_t, void*, uint32_t), uint32_t _sc, remote_arg* _pra) {
static __inline int _skel_method_4(int (*_pfn)(uint32_t, uint32_t, void*, uint32_t, void*, uint32_t, char*, uint32_t), uint32_t _sc, remote_arg* _pra) {
remote_arg* _praEnd;
uint32_t _in0[1];
uint32_t _in1[1];
void* _in2[1];
uint32_t _in2Len[1];
void* _rout3[1];
uint32_t _rout3Len[1];
void* _in4[1];
char* _in4[1];
uint32_t _in4Len[1];
uint32_t* _primIn;
int _numIn[1];
Expand All @@ -595,8 +595,6 @@ static __inline int _skel_method_4(int (*_pfn)(uint32_t, uint32_t, void*, uint32
int _nErr = 0;
char* _seq_primIn3;
char* _seq_nat3;
char* _seq_primIn4;
char* _seq_nat4;
_praEnd = ((_pra + REMOTE_SCALARS_INBUFS(_sc)) + REMOTE_SCALARS_OUTBUFS(_sc) + REMOTE_SCALARS_INHANDLES(_sc) + REMOTE_SCALARS_OUTHANDLES(_sc));
_ASSERT(_nErr, (_pra + ((4 + 0) + (0 + 0))) <= _praEnd);
_numIn[0] = (REMOTE_SCALARS_INBUFS(_sc) - 1);
Expand All @@ -613,34 +611,26 @@ static __inline int _skel_method_4(int (*_pfn)(uint32_t, uint32_t, void*, uint32
_ALLOCATE(_nErr, _al, (_in2Len[0] * SLIM_IFPTR32(8, 16)), SLIM_IFPTR32(4, 8), _in2[0]);
for(_ii = 0, _seq_primIn2 = (char*)_praIn[0].buf.pv, _seq_nat2 = (char*)_in2[0];_ii < (int)_in2Len[0];++_ii, _seq_primIn2 = (_seq_primIn2 + 4), _seq_nat2 = (_seq_nat2 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_unpack(_al, (_praIn + 1), _ppraIn, (_praROut + 0), _ppraROut, _seq_primIn2, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat2)[0]), (char**)&(((uint64_t*)_seq_nat2)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat2)[1]), (uint32_t*)&(((uint32_t*)_seq_nat2)[2]))));
_TRY(_nErr, _skel_unpack_1(_al, (_praIn + 1), _ppraIn, (_praROut + 0), _ppraROut, _seq_primIn2, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat2)[0]), (char**)&(((uint64_t*)_seq_nat2)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat2)[1]), (uint32_t*)&(((uint32_t*)_seq_nat2)[2]))));
}
_COPY(_rout3Len, 0, _primIn, 12, 4);
_ASSERT(_nErr, (int)((_praIn[1].buf.nLen / 4)) >= (int)(_rout3Len[0]));
_ALLOCATE(_nErr, _al, (_rout3Len[0] * SLIM_IFPTR32(8, 16)), SLIM_IFPTR32(4, 8), _rout3[0]);
for(_ii = 0, _seq_primIn3 = (char*)_praIn[1].buf.pv, _seq_nat3 = (char*)_rout3[0];_ii < (int)_rout3Len[0];++_ii, _seq_primIn3 = (_seq_primIn3 + 4), _seq_nat3 = (_seq_nat3 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_unpack_1(_al, (_praIn + 2), _ppraIn, (_praROut + 0), _ppraROut, _seq_primIn3, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat3)[0]), (char**)&(((uint64_t*)_seq_nat3)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat3)[1]), (uint32_t*)&(((uint32_t*)_seq_nat3)[2]))));
_TRY(_nErr, _skel_unpack(_al, (_praIn + 2), _ppraIn, (_praROut + 0), _ppraROut, _seq_primIn3, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat3)[0]), (char**)&(((uint64_t*)_seq_nat3)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat3)[1]), (uint32_t*)&(((uint32_t*)_seq_nat3)[2]))));
}
_COPY(_in4Len, 0, _primIn, 16, 4);
_ASSERT(_nErr, (int)((_praIn[2].buf.nLen / 4)) >= (int)(_in4Len[0]));
_ALLOCATE(_nErr, _al, (_in4Len[0] * SLIM_IFPTR32(8, 16)), SLIM_IFPTR32(4, 8), _in4[0]);
for(_ii = 0, _seq_primIn4 = (char*)_praIn[2].buf.pv, _seq_nat4 = (char*)_in4[0];_ii < (int)_in4Len[0];++_ii, _seq_primIn4 = (_seq_primIn4 + 4), _seq_nat4 = (_seq_nat4 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_unpack(_al, (_praIn + 3), _ppraIn, (_praROut + 0), _ppraROut, _seq_primIn4, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat4)[0]), (char**)&(((uint64_t*)_seq_nat4)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat4)[1]), (uint32_t*)&(((uint32_t*)_seq_nat4)[2]))));
}
_ASSERT(_nErr, (int)((_praIn[2].buf.nLen / 8)) >= (int)(_in4Len[0]));
_in4[0] = _praIn[2].buf.pv;
_TRY(_nErr, _pfn(*_in0, *_in1, *_in2, *_in2Len, *_rout3, *_rout3Len, *_in4, *_in4Len));
for(_ii = 0, _seq_nat2 = (char*)_in2[0];_ii < (int)_in2Len[0];++_ii, _seq_nat2 = (_seq_nat2 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_pack((_praROutPost + 0), _ppraROutPost, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat2)[0]), (char**)&(((uint64_t*)_seq_nat2)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat2)[1]), (uint32_t*)&(((uint32_t*)_seq_nat2)[2]))));
_TRY(_nErr, _skel_pack_1((_praROutPost + 0), _ppraROutPost, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat2)[0]), (char**)&(((uint64_t*)_seq_nat2)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat2)[1]), (uint32_t*)&(((uint32_t*)_seq_nat2)[2]))));
}
for(_ii = 0, _seq_nat3 = (char*)_rout3[0];_ii < (int)_rout3Len[0];++_ii, _seq_nat3 = (_seq_nat3 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_pack_1((_praROutPost + 0), _ppraROutPost, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat3)[0]), (char**)&(((uint64_t*)_seq_nat3)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat3)[1]), (uint32_t*)&(((uint32_t*)_seq_nat3)[2]))));
}
for(_ii = 0, _seq_nat4 = (char*)_in4[0];_ii < (int)_in4Len[0];++_ii, _seq_nat4 = (_seq_nat4 + SLIM_IFPTR32(8, 16)))
{
_TRY(_nErr, _skel_pack((_praROutPost + 0), _ppraROutPost, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat4)[0]), (char**)&(((uint64_t*)_seq_nat4)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat4)[1]), (uint32_t*)&(((uint32_t*)_seq_nat4)[2]))));
_TRY(_nErr, _skel_pack((_praROutPost + 0), _ppraROutPost, 0, SLIM_IFPTR32((char**)&(((uint32_t*)_seq_nat3)[0]), (char**)&(((uint64_t*)_seq_nat3)[0])), SLIM_IFPTR32((uint32_t*)&(((uint32_t*)_seq_nat3)[1]), (uint32_t*)&(((uint32_t*)_seq_nat3)[2]))));
}
_CATCH(_nErr) {}
_allocator_deinit(_al);
Expand Down Expand Up @@ -721,7 +711,7 @@ __QAIC_SKEL_EXPORT int __QAIC_SKEL(halide_hexagon_remote_skel_invoke)(uint32_t _
case 3:
return _skel_method_5((void*)__QAIC_IMPL(halide_hexagon_remote_power_hvx_off), _sc, _pra);
case 4:
return _skel_method_4((void*)__QAIC_IMPL(halide_hexagon_remote_run), _sc, _pra);
return _skel_method_4((void*)__QAIC_IMPL(halide_hexagon_remote_run_v2), _sc, _pra);
case 5:
return _skel_method_1((void*)__QAIC_IMPL(halide_hexagon_remote_release_kernels_v2), _sc, _pra);
case 6:
Expand Down
Loading

0 comments on commit 36f34d8

Please sign in to comment.