iswaalex commited on
Commit
885a4cf
·
verified ·
1 Parent(s): eb50e91

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_10.bin +3 -0
  2. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_101.bin +3 -0
  3. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_102.bin +3 -0
  4. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_104.bin +3 -0
  5. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_105.bin +3 -0
  6. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_106.bin +3 -0
  7. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_107.bin +3 -0
  8. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_138.bin +3 -0
  9. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_144.bin +3 -0
  10. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_145.bin +3 -0
  11. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/Release/0_0_reloadable22.txt +0 -0
  12. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/scripts/0_0_reloadable22.bcf +16 -0
  13. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/scripts/0_0_reloadable22.prx +13 -0
  14. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/src/0_0_reloadable22.cc +81 -0
  15. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/timestamped_log/0_0_reloadable22.log +0 -0
  16. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/0_0_reloadable23.log +0 -0
  17. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.# +2 -0
  18. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.## +2 -0
  19. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.calltree +92 -0
  20. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.cmic2 +0 -0
  21. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.cmico +1 -0
  22. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.lst +0 -0
  23. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.map +296 -0
  24. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.o.lst +0 -0
  25. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.sdr +125 -0
  26. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.srv +0 -0
  27. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.txt +0 -0
  28. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/scripts/0_0_reloadable23.bcf +16 -0
  29. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/scripts/0_0_reloadable23.prx +13 -0
  30. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/src/0_0_reloadable23.cc +124 -0
  31. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/timestamped_log/0_0_reloadable23.log +0 -0
  32. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/xlopt.log +482 -0
  33. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/0_0_reloadable24.log +0 -0
  34. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.# +2 -0
  35. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.## +2 -0
  36. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.calltree +88 -0
  37. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.cmic2 +0 -0
  38. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.cmico +1 -0
  39. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.lst +0 -0
  40. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.map +294 -0
  41. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.o.lst +0 -0
  42. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.sdr +125 -0
  43. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.srv +0 -0
  44. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.txt +0 -0
  45. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/scripts/0_0_reloadable24.bcf +16 -0
  46. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/scripts/0_0_reloadable24.prx +13 -0
  47. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/src/0_0_reloadable24.cc +113 -0
  48. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/timestamped_log/0_0_reloadable24.log +0 -0
  49. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/xlopt.log +437 -0
  50. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable25/0_0_reloadable25.log +0 -0
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_10.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d879ffada09c8269291fd9cf3bdb2b7e50406eb6df005ecb945752bcbb9c7897
3
+ size 468
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_101.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5100b5f145a2a15caa641f98633e597029e1544687a2e9d2da66a66c572bc71
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_102.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18edbb71396011e9afb08421953f12d77c9a9d2515d6233f5199f44249553442
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_104.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c8fec3a09bf7510cd8ce1ab86c293d7b466805391ace22f6b975a073c96e6e03
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_105.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:541b329c484a094eeac656452e18f0be0816bb2438bdfd19f8ad1e0a9b3b9cc3
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_106.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6fb9253ded0289c5b796164e40201bd8f5d660b2373d91b5b58eb314375af7e7
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_107.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ae43a8dcc0cd99130f42249d4f2a9f1fc91c3c99699db65bb98ad360a326ef6
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_138.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1478e76c247bdc716175f91eaec431c57078b8990e57acb91029527c65ed934
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_144.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1a98d8789fe0e97e55dd1e7973c6dcfdce60fdf56b3f49d0b75e4bf4cf87aa1b
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_145.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:165102e6e1386d3966ba3ffeb0747f572c719a1eb2bc25086f46217b1e5fa9af
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/Release/0_0_reloadable22.txt ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/scripts/0_0_reloadable22.bcf ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _reserved DMb 0x0 0x40000
2
+
3
+ _reserved PM 0x0 0x930 //reserved for main elf
4
+
5
+ _entry_point _Z13kernelWrapperPPvjjjj
6
+ _symbol _Z13kernelWrapperPPvjjjj 0x930
7
+
8
+ _reserved DMb 0x7b280 0x800 //reserved for lcp ping-pong buffers
9
+ _reserved DMb 0x7ba80 0x40 //reserved for sync buffer
10
+ _stack DM_stack 0x7bac0 0x940 //stack for core
11
+ _reserved DMb 0x7c400 0x40 //reserved for main elf heap
12
+ //space for synopsys compiler at 0x7c440 0x880//heap
13
+ _reserved DMb 0x40000 0x3b280
14
+
15
+ _reserved DMb 0x7ccc0 0x3340
16
+ _reserved DMb 0x80000 0x80000 // And everything else the core can't see
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/scripts/0_0_reloadable22.prx ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <project name="Project" processor="me">
2
+ <file type="lbc" name="0_0_reloadable22.ll" path="../../ir"/>
3
+ <issinit/>
4
+ <option id="cpp.define" value="__AIENGINE__ __AIE_ARCH__=21" inherit = "1"/>
5
+ <option id="llvm.xargs" value="-fno-jump-tables -fno-discard-value-names" inherit = "1"/>
6
+ <option id="llvm.lang" value="Follow file extension"/>
7
+ <option id="bridge.cfg" value="0_0_reloadable22.bcf"/>
8
+ <option id="cpp.include" value="&lt;XILINX_VITIS_AIETOOLS&gt;/include /app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend /usr/local/lib/python3.10/site-packages/include/aie_api /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf /app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/&lt;AIE_SRC_DIR&gt; " inherit="1"/>
9
+ <option id="project.dir" value="&lt;CONFIG&gt;/0_0_reloadable22"/>
10
+ <option id="project.name" value="0_0_reloadable22"/>
11
+ <option id="project.type" value="exe"/>
12
+ <option id="simrun.issca" value="&lt;XILINX_VITIS_AIETOOLS&gt;/bin/unwrapped/lnx64.o/aie2p/ca_udm_dbg" inherit="0"/>
13
+ </project>
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/src/0_0_reloadable22.cc ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Automatically generated processor driver using AIEngine tool-chain
2
+
3
+ #include <stdio.h>
4
+ #include <memory>
5
+ #include <adf.h>
6
+
7
+
8
+ // Declare Kernel functions and initializers
9
+ #include "resize_adf_wrapper.cpp"
10
+ void conv2d_maxpool(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
11
+ #include "slice_adf_wrapper.cpp"
12
+ #include "concat_adf_wrapper.cpp"
13
+
14
+ // Declare Kernel objects and external arrays
15
+
16
+
17
+ void _b787_wrapper(void* args[])
18
+ {
19
+ mllib_graphs::resize_adf_wrapper<bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>, 1, 0, 0>(
20
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
21
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
22
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]));
23
+ }
24
+
25
+ void _b896_wrapper(void* args[])
26
+ {
27
+ conv2d_maxpool(
28
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
29
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
30
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
31
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
32
+ }
33
+
34
+ void _b806_wrapper(void* args[])
35
+ {
36
+ mllib_graphs::slice_adf_wrapper<1, bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>>(
37
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
38
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
39
+ *reinterpret_cast<uint32 const(*)[7]>(args[2]));
40
+ }
41
+
42
+ void _b820_wrapper(void* args[])
43
+ {
44
+ mllib_graphs::concat_adf_wrapper<bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>>(
45
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
46
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
47
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[2]),
48
+ *reinterpret_cast<uint32 const(*)[8]>(args[3]));
49
+ }
50
+
51
+ using UniformKernelFunc = void (*)(void **);
52
+
53
+ static UniformKernelFunc g_uniformKernelFuncs[4] = {
54
+ _b787_wrapper,
55
+ _b896_wrapper,
56
+ _b806_wrapper,
57
+ _b820_wrapper
58
+ };
59
+
60
+ __attribute__((always_inline)) void kernelWrapper(void* args[], uint32 kernelId, uint32 numSyncIn, uint32 numAsyncIn, uint32 numSyncOut)
61
+ {
62
+ uint32 idx = 0;
63
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 0);
64
+ idx += (numSyncIn > 0) ? 1 : 0;
65
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 1);
66
+ idx += (numSyncIn > 1) ? 1 : 0;
67
+ idx += numAsyncIn;
68
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::out>(numSyncOut > 0);
69
+ idx += (numSyncOut > 0) ? 1 : 0;
70
+
71
+ (*(g_uniformKernelFuncs[kernelId]))(args);
72
+
73
+ idx = 0;
74
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 0);
75
+ idx += (numSyncIn > 0) ? 1 : 0;
76
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 1);
77
+ idx += (numSyncIn > 1) ? 1 : 0;
78
+ idx += numAsyncIn;
79
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::out>(numSyncOut > 0);
80
+ idx += (numSyncOut > 0) ? 1 : 0;
81
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable22/timestamped_log/0_0_reloadable22.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/0_0_reloadable23.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.# ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 23ecfe0fdedd1a524eaff21654c4c29bced37405
2
+ 1a735f496f1284ab86135ede88939a8aeeb375a3
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.## ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 1972f852fee958b21b761e09c2a029813ccf5832
2
+ 92e103875c6a39bc8cdbe4a21c02c5a94ba3cc96
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.calltree ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:02:00 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable23 ../Release/0_0_reloadable23.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable23.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6583 -pme
5
+
6
+
7
+ // Release: ipp V-2024.06-TGT-241219
8
+
9
+ _Z13kernelWrapperPPvjjjj
10
+ _Z14_b1638_wrapperPPv (referenced text)
11
+ _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
12
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
13
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
14
+ _Z13_b806_wrapperPPv (referenced text)
15
+ _ZN12mllib_graphs17slice_adf_wrapperILi1E8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_EEvRNS2_9io_bufferIT0_NS2_9direction2inET1_EERNSD_ISE_NSF_3outET2_EERA7_Kj
16
+ _Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params
17
+ _Z14_b1655_wrapperPPv (referenced text)
18
+ _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
19
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
20
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
21
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
22
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
23
+ _Z13_b891_wrapperPPv (referenced text)
24
+ _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
25
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
26
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
27
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
28
+ _Z13_b820_wrapperPPv (referenced text)
29
+ _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj
30
+ _Z13_b896_wrapperPPv (referenced text)
31
+ _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
32
+ _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
33
+ _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
34
+ _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
35
+ _Z14_b1672_wrapperPPv (referenced text)
36
+ _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
37
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
38
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
39
+ _Z13_b886_wrapperPPv (referenced text)
40
+ _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
41
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
42
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
43
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
44
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (*)
45
+
46
+
47
+ Call tree stack and functions sizes:
48
+
49
+ stack stack stack call func func function name
50
+ desc level level desc
51
+ ----- ----- ----- ----- ----- ----- --------------------------------------------------------------
52
+ 64 512 0 0 546 13444 _Z13kernelWrapperPPvjjjj
53
+ 0 64 1 1 32 960 _Z14_b1638_wrapperPPv
54
+ 64 64 1 2 488 928 _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
55
+ 0 0 2 3 62 62 _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
56
+ 0 0 2 3 378 378 _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
57
+ 0 192 1 1 32 1488 _Z13_b806_wrapperPPv
58
+ 192 192 1 2 674 1456 _ZN12mllib_graphs17slice_adf_wrapperILi1E8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_EEvRNS2_9io_bufferIT0_NS2_9direction2inET1_EERNSD_ISE_NSF_3outET2_EERA7_Kj <tail call>
59
+ 0 0 2 3 782 782 _Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params
60
+ 0 192 1 1 36 1378 _Z14_b1655_wrapperPPv
61
+ 64 192 1 2 602 1342 _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
62
+ 64 64 2 3 74 190 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
63
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
64
+ 0 128 2 3 16 550 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
65
+ 128 128 2 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE <tail call>
66
+ 0 128 1 1 36 1092 _Z13_b891_wrapperPPv
67
+ 64 128 1 2 602 1056 _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
68
+ 64 64 2 3 138 162 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
69
+ 0 0 3 4 24 24 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
70
+ 0 0 2 3 292 292 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
71
+ 0 0 1 1 36 854 _Z13_b820_wrapperPPv
72
+ 0 0 1 2 818 818 _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj <tail call>
73
+ 0 192 1 1 36 4714 _Z13_b896_wrapperPPv
74
+ 64 192 1 2 568 4678 _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
75
+ 64 64 2 3 1430 1430 _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
76
+ 128 128 2 3 2410 2680 _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
77
+ 0 0 3 4 270 270 _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
78
+ 0 448 1 1 32 1544 _Z14_b1672_wrapperPPv
79
+ 64 448 1 2 488 1512 _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
80
+ 0 0 2 3 62 62 _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
81
+ 384 384 2 3 962 962 _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
82
+ 0 192 1 1 36 1402 _Z13_b886_wrapperPPv
83
+ 64 192 1 2 602 1366 _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
84
+ 64 64 2 3 98 214 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
85
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
86
+ 0 128 2 3 16 550 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
87
+ 128 128 2 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE <tail call>
88
+
89
+
90
+ Maximum call level : 4
91
+ Maximum stack level: 3
92
+ Maximum stack size : 512
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.cmic2 ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.cmico ADDED
@@ -0,0 +1 @@
 
 
1
+ +Mdec
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.lst ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.map ADDED
@@ -0,0 +1,296 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:02:00 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable23 ../Release/0_0_reloadable23.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable23.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6583 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ Memory map for memory 'DM_stack':
9
+
10
+ Size = 1048576
11
+ Width = 8 bits
12
+ Offset = 0
13
+ Used = 2368
14
+
15
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
16
+
17
+ Memory map for memory 'DMb':
18
+
19
+ Size = 1048576
20
+ Width = 8 bits
21
+ Offset = 0
22
+ Used = 3217
23
+
24
+ 0x00000000..0x0007babf ( 506560 items) : Reserved
25
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
26
+ 0x0007c400..0x0007c43f ( 64 items) : Reserved
27
+ 0x0007c440..0x0007c443 ( 4 items) : ../Release/0_0_reloadable23.o::_ZL9curr_iter (Data, Local, .bss.DMb.4)
28
+ 0x0007c444..0x0007c447 ( 4 items) : ../Release/0_0_reloadable23.o::_ZL14num_depth_iter (Data, Local, .bss.DMb.4)
29
+ 0x0007c448..0x0007c44b ( 4 items) : ../Release/0_0_reloadable23.o::_ZL10depth_iter (Data, Local, .bss.DMb.4)
30
+ 0x0007c44c..0x0007c44f ( 4 items) : ../Release/0_0_reloadable23.o::_ZL11total_iters (Data, Local, .bss.DMb.4)
31
+ 0x0007c450..0x0007c453 ( 4 items) : ../Release/0_0_reloadable23.o::_ZL8core_row (Data, Local, .bss.DMb.4)
32
+ 0x0007c454..0x0007c457 ( 4 items) : ../Release/0_0_reloadable23.o::_ZL11ifm1_offset (Data, Local, .bss.DMb.4)
33
+ 0x0007c458..0x0007c45b ( 4 items) : ../Release/0_0_reloadable23.o::_ZL11ifm2_offset (Data, Local, .bss.DMb.4)
34
+ 0x0007c45c..0x0007c45f ( 4 items) : ../Release/0_0_reloadable23.o::_ZN9ns_concatL11kernel_iterE (Data, Local, .bss.DMb.4)
35
+ 0x0007c460..0x0007c463 ( 4 items) : ../Release/0_0_reloadable23.o::_ZN9ns_concatL8core_colE (Data, Local, .bss.DMb.4)
36
+ 0x0007c464..0x0007c467 ( 4 items) : ../Release/0_0_reloadable23.o::_ZN9ns_concatL8core_rowE (Data, Local, .bss.DMb.4)
37
+ 0x0007c468..0x0007c46b ( 4 items) : ../Release/0_0_reloadable23.o::_ZL8num_iter (Data, Local, .data.DMb.4)
38
+ 0x0007c46c..0x0007c46f ( 4 items) : me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_ZN12me_primitive11control_satE (Data, Global, .bss.DMb.4)
39
+ 0x0007c470..0x0007c470 ( 1 items) : me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_ZN12me_primitive11control_rndE (Data, Global, .bss.DMb.1)
40
+ 0x0007c480..0x0007c49f ( 32 items) : ../Release/0_0_reloadable23.o::_ZL20g_uniformKernelFuncs (Data, Local, .rodata.DMb.64)
41
+
42
+ Called functions : _Z14_b1638_wrapperPPv
43
+ _Z13_b806_wrapperPPv
44
+ _Z14_b1655_wrapperPPv
45
+ _Z13_b891_wrapperPPv
46
+ _Z13_b820_wrapperPPv
47
+ _Z13_b896_wrapperPPv
48
+ _Z14_b1672_wrapperPPv
49
+ _Z13_b886_wrapperPPv
50
+
51
+ 0x0007c4c0..0x0007c4ff ( 64 items) : ../Release/0_0_reloadable23.o::add1d_params (Data, Global, .bss.DMb.64)
52
+ 0x0007c500..0x0007c53f ( 64 items) : ../Release/0_0_reloadable23.o::mul1d_params (Data, Global, .bss.DMb.64)
53
+ 0x0007c540..0x0007c57f ( 64 items) : ../Release/0_0_reloadable23.o::sub1d_params (Data, Global, .bss.DMb.64)
54
+ 0x0007c580..0x0007c5bf ( 64 items) : ../Release/0_0_reloadable23.o::sigmoid1d_params (Data, Global, .bss.DMb.64)
55
+ 0x0007c5c0..0x0007c5ff ( 64 items) : ../Release/0_0_reloadable23.o::tanh1d_params (Data, Global, .bss.DMb.64)
56
+ 0x0007c600..0x0007c7bf ( 448 items) : ../Release/0_0_reloadable23.o::conv2d_params (Data, Global, .bss.DMb.64)
57
+ 0x0007ccc0..0x000fffff ( 537408 items) : Reserved
58
+
59
+ Memory map for memory 'PM':
60
+
61
+ Size = 1048576
62
+ Width = 8 bits
63
+ Offset = 0
64
+ Used = 13444
65
+
66
+ 0x00000000..0x0000092f ( 2352 items) : Reserved
67
+ 0x00000930..0x00000b51 ( 546 items) : ../Release/0_0_reloadable23.o::_Z13kernelWrapperPPvjjjj (Function, Global, .text) (stack frame size = 64)
68
+
69
+ Referenced symbols: _ZL20g_uniformKernelFuncs
70
+
71
+ 0x00000b60..0x000010f5 ( 1430 items) : ../Release/0_0_reloadable23.o::_Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh (Function, Weak, .text) (stack frame size = 64)
72
+ 0x00001100..0x0000120d ( 270 items) : ../Release/0_0_reloadable23.o::_Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams (Function, Weak, .text) (stack frame size = 0)
73
+
74
+ Referenced symbols: _ZN12me_primitive11control_rndE
75
+
76
+ 0x00001210..0x00001b79 ( 2410 items) : ../Release/0_0_reloadable23.o::_Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params (Function, Weak, .text) (stack frame size = 128)
77
+
78
+ Called functions : _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
79
+
80
+ Referenced symbols: _ZN12me_primitive11control_rndE
81
+
82
+ 0x00001b80..0x00001db7 ( 568 items) : ../Release/0_0_reloadable23.o::_Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
83
+
84
+ Called functions : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
85
+ _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
86
+
87
+ Referenced symbols: _ZL9curr_iter
88
+ conv2d_params
89
+ _ZN12me_primitive11control_rndE
90
+ _ZN12me_primitive11control_satE
91
+ _ZL14num_depth_iter
92
+ _ZL8num_iter
93
+ _ZL10depth_iter
94
+ _ZL11total_iters
95
+
96
+ 0x00001dc0..0x00001fd5 ( 534 items) : ../Release/0_0_reloadable23.o::_ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (Function, Local, .text) (stack frame size = 128)
97
+
98
+ Referenced symbols: _ZN12me_primitive11control_rndE
99
+
100
+ 0x00001fe0..0x0000201d ( 62 items) : ../Release/0_0_reloadable23.o::_ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
101
+ 0x00002020..0x00002199 ( 378 items) : ../Release/0_0_reloadable23.o::_ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
102
+
103
+ Referenced symbols: _ZN12me_primitive11control_rndE
104
+
105
+ 0x000021a0..0x00002387 ( 488 items) : ../Release/0_0_reloadable23.o::_Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
106
+
107
+ Called functions : _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
108
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
109
+
110
+ Referenced symbols: _ZL9curr_iter
111
+ _ZL8core_row
112
+ sigmoid1d_params
113
+ _ZN12me_primitive11control_rndE
114
+ _ZN12me_primitive11control_satE
115
+ _ZL11ifm1_offset
116
+ _ZL8num_iter
117
+
118
+ 0x00002390..0x000023cd ( 62 items) : ../Release/0_0_reloadable23.o::_ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
119
+ 0x000023d0..0x00002791 ( 962 items) : ../Release/0_0_reloadable23.o::_ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E (Function, Weak, .text) (stack frame size = 384)
120
+
121
+ Referenced symbols: _ZN12me_primitive11control_rndE
122
+
123
+ 0x000027a0..0x00002987 ( 488 items) : ../Release/0_0_reloadable23.o::_Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
124
+
125
+ Called functions : _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
126
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
127
+
128
+ Referenced symbols: _ZL9curr_iter
129
+ _ZL8core_row
130
+ tanh1d_params
131
+ _ZN12me_primitive11control_rndE
132
+ _ZN12me_primitive11control_satE
133
+ _ZL11ifm1_offset
134
+ _ZL8num_iter
135
+
136
+ 0x00002990..0x00002a03 ( 116 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 0)
137
+ 0x00002a10..0x00002a71 ( 98 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 64)
138
+
139
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
140
+
141
+ 0x00002a80..0x00002a8f ( 16 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 0)
142
+
143
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
144
+
145
+ 0x00002a90..0x00002ce9 ( 602 items) : ../Release/0_0_reloadable23.o::_Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
146
+
147
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
148
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
149
+
150
+ Referenced symbols: _ZL9curr_iter
151
+ _ZL8core_row
152
+ add1d_params
153
+ _ZN12me_primitive11control_rndE
154
+ _ZN12me_primitive11control_satE
155
+ _ZL11ifm1_offset
156
+ _ZL11ifm2_offset
157
+ _ZL8num_iter
158
+
159
+ 0x00002cf0..0x00002d07 ( 24 items) : ../Release/0_0_reloadable23.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
160
+ 0x00002d10..0x00002d99 ( 138 items) : ../Release/0_0_reloadable23.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
161
+
162
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
163
+
164
+ 0x00002da0..0x00002ec3 ( 292 items) : ../Release/0_0_reloadable23.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
165
+
166
+ Referenced symbols: _ZN12me_primitive11control_rndE
167
+
168
+ 0x00002ed0..0x00003129 ( 602 items) : ../Release/0_0_reloadable23.o::_Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
169
+
170
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
171
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
172
+
173
+ Referenced symbols: _ZL9curr_iter
174
+ _ZL8core_row
175
+ mul1d_params
176
+ _ZN12me_primitive11control_rndE
177
+ _ZN12me_primitive11control_satE
178
+ _ZL11ifm1_offset
179
+ _ZL11ifm2_offset
180
+ _ZL8num_iter
181
+
182
+ 0x00003130..0x000031a3 ( 116 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
183
+ 0x000031b0..0x000031f9 ( 74 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
184
+
185
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
186
+
187
+ 0x00003200..0x0000320f ( 16 items) : ../Release/0_0_reloadable23.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
188
+
189
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
190
+
191
+ 0x00003210..0x00003469 ( 602 items) : ../Release/0_0_reloadable23.o::_Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
192
+
193
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
194
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
195
+
196
+ Referenced symbols: _ZL9curr_iter
197
+ _ZL8core_row
198
+ sub1d_params
199
+ _ZN12me_primitive11control_rndE
200
+ _ZN12me_primitive11control_satE
201
+ _ZL11ifm1_offset
202
+ _ZL11ifm2_offset
203
+ _ZL8num_iter
204
+
205
+ 0x00003470..0x0000348f ( 32 items) : ../Release/0_0_reloadable23.o::_Z14_b1638_wrapperPPv (Function, Global, .text) (stack frame size = 0)
206
+
207
+ Called functions : _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
208
+
209
+ 0x00003490..0x0000379d ( 782 items) : ../Release/0_0_reloadable23.o::_Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params (Function, Weak, .text) (stack frame size = 0)
210
+ 0x000037a0..0x00003a41 ( 674 items) : ../Release/0_0_reloadable23.o::_ZN12mllib_graphs17slice_adf_wrapperILi1E8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_EEvRNS2_9io_bufferIT0_NS2_9direction2inET1_EERNSD_ISE_NSF_3outET2_EERA7_Kj (Function, Weak, .text) (stack frame size = 192)
211
+
212
+ Called functions : _Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params
213
+
214
+ 0x00003a50..0x00003a6f ( 32 items) : ../Release/0_0_reloadable23.o::_Z13_b806_wrapperPPv (Function, Global, .text) (stack frame size = 0)
215
+
216
+ Called functions : _ZN12mllib_graphs17slice_adf_wrapperILi1E8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_EEvRNS2_9io_bufferIT0_NS2_9direction2inET1_EERNSD_ISE_NSF_3outET2_EERA7_Kj
217
+
218
+ 0x00003a70..0x00003a93 ( 36 items) : ../Release/0_0_reloadable23.o::_Z14_b1655_wrapperPPv (Function, Global, .text) (stack frame size = 0)
219
+
220
+ Called functions : _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
221
+
222
+ 0x00003aa0..0x00003ac3 ( 36 items) : ../Release/0_0_reloadable23.o::_Z13_b891_wrapperPPv (Function, Global, .text) (stack frame size = 0)
223
+
224
+ Called functions : _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
225
+
226
+ 0x00003ad0..0x00003e01 ( 818 items) : ../Release/0_0_reloadable23.o::_ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj (Function, Weak, .text) (stack frame size = 0)
227
+
228
+ Referenced symbols: _ZN9ns_concatL11kernel_iterE
229
+ _ZN12me_primitive11control_rndE
230
+ _ZN9ns_concatL8core_rowE
231
+ _ZN12me_primitive11control_satE
232
+ _ZN9ns_concatL8core_colE
233
+
234
+ 0x00003e10..0x00003e33 ( 36 items) : ../Release/0_0_reloadable23.o::_Z13_b820_wrapperPPv (Function, Global, .text) (stack frame size = 0)
235
+
236
+ Called functions : _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj
237
+
238
+ 0x00003e40..0x00003e63 ( 36 items) : ../Release/0_0_reloadable23.o::_Z13_b896_wrapperPPv (Function, Global, .text) (stack frame size = 0)
239
+
240
+ Called functions : _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
241
+
242
+ 0x00003e70..0x00003e8f ( 32 items) : ../Release/0_0_reloadable23.o::_Z14_b1672_wrapperPPv (Function, Global, .text) (stack frame size = 0)
243
+
244
+ Called functions : _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
245
+
246
+ 0x00003e90..0x00003eb3 ( 36 items) : ../Release/0_0_reloadable23.o::_Z13_b886_wrapperPPv (Function, Global, .text) (stack frame size = 0)
247
+
248
+ Called functions : _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
249
+
250
+
251
+ External symbols:
252
+
253
+ __dso_handle = 0x0
254
+ _ctors_end = 0x0
255
+ _ctors_start = 0x0
256
+ _dtors_end = 0x0
257
+ _dtors_start = 0x0
258
+ _pc_end = 0x3eb4
259
+ _pc_start = 0x930
260
+ _sp_end_DM_stack = 0x7c400
261
+ _sp_start_DM_stack = 0x7bac0
262
+
263
+ Section summary for memory 'DM_stack':
264
+
265
+ .stack File
266
+ ---------- ----------
267
+ 2368
268
+ ---------- ----------
269
+ 2368 Total
270
+
271
+ Section summary for memory 'DMb':
272
+
273
+ .bss .data .rodata File
274
+ ---------- ---------- ---------- ----------
275
+ 808 4 32 ../Release/0_0_reloadable23.o
276
+ 5 0 0 me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
277
+ ---------- ---------- ---------- ----------
278
+ 813 4 32 Total
279
+
280
+ Section summary for memory 'PM':
281
+
282
+ .text File
283
+ ---------- ----------
284
+ 13444 ../Release/0_0_reloadable23.o
285
+ ---------- ----------
286
+ 13444 Total
287
+
288
+ File summary:
289
+
290
+ ../Release/0_0_reloadable23.o
291
+ DMb 844
292
+ PM 13444
293
+
294
+ me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
295
+ DMb 5
296
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.o.lst ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.sdr ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:02:00 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable23 ../Release/0_0_reloadable23.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable23.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6583 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ // Symbols in memory 'DM_bankA':
9
+ // Symbols in memory 'DM_bankAB':
10
+ // Symbols in memory 'DM_bankAC':
11
+ // Symbols in memory 'DM_bankAD':
12
+ // Symbols in memory 'DM_bankB':
13
+ // Symbols in memory 'DM_bankBC':
14
+ // Symbols in memory 'DM_bankBD':
15
+ // Symbols in memory 'DM_bankC':
16
+ // Symbols in memory 'DM_bankCD':
17
+ // Symbols in memory 'DM_bankD':
18
+ // Symbols in memory 'DM_stack':
19
+ // Symbols in memory 'DM_test':
20
+ // Symbols in memory 'DMb':
21
+ _symbol _ZN12me_primitive11control_satE 0x0007c46c
22
+ _symbol _ZN12me_primitive11control_rndE 0x0007c470
23
+ _symbol add1d_params 0x0007c4c0
24
+ _symbol mul1d_params 0x0007c500
25
+ _symbol sub1d_params 0x0007c540
26
+ _symbol sigmoid1d_params 0x0007c580
27
+ _symbol tanh1d_params 0x0007c5c0
28
+ _symbol conv2d_params 0x0007c600
29
+ // Symbols in memory 'DMh':
30
+ // Symbols in memory 'DMh_bankA':
31
+ // Symbols in memory 'DMh_bankAB':
32
+ // Symbols in memory 'DMh_bankAC':
33
+ // Symbols in memory 'DMh_bankAD':
34
+ // Symbols in memory 'DMh_bankB':
35
+ // Symbols in memory 'DMh_bankBC':
36
+ // Symbols in memory 'DMh_bankBD':
37
+ // Symbols in memory 'DMh_bankC':
38
+ // Symbols in memory 'DMh_bankCD':
39
+ // Symbols in memory 'DMh_bankD':
40
+ // Symbols in memory 'DMh_stack':
41
+ // Symbols in memory 'DMs':
42
+ // Symbols in memory 'DMs_bankA':
43
+ // Symbols in memory 'DMs_bankAB':
44
+ // Symbols in memory 'DMs_bankAC':
45
+ // Symbols in memory 'DMs_bankAD':
46
+ // Symbols in memory 'DMs_bankB':
47
+ // Symbols in memory 'DMs_bankBC':
48
+ // Symbols in memory 'DMs_bankBD':
49
+ // Symbols in memory 'DMs_bankC':
50
+ // Symbols in memory 'DMs_bankCD':
51
+ // Symbols in memory 'DMs_bankD':
52
+ // Symbols in memory 'DMs_stack':
53
+ // Symbols in memory 'DMv':
54
+ // Symbols in memory 'DMv_bankA':
55
+ // Symbols in memory 'DMv_bankAB':
56
+ // Symbols in memory 'DMv_bankAC':
57
+ // Symbols in memory 'DMv_bankAD':
58
+ // Symbols in memory 'DMv_bankB':
59
+ // Symbols in memory 'DMv_bankBC':
60
+ // Symbols in memory 'DMv_bankBD':
61
+ // Symbols in memory 'DMv_bankC':
62
+ // Symbols in memory 'DMv_bankCD':
63
+ // Symbols in memory 'DMv_bankD':
64
+ // Symbols in memory 'DMv_stack':
65
+ // Symbols in memory 'DMw':
66
+ // Symbols in memory 'DMw_bankA':
67
+ // Symbols in memory 'DMw_bankAB':
68
+ // Symbols in memory 'DMw_bankAC':
69
+ // Symbols in memory 'DMw_bankAD':
70
+ // Symbols in memory 'DMw_bankB':
71
+ // Symbols in memory 'DMw_bankBC':
72
+ // Symbols in memory 'DMw_bankBD':
73
+ // Symbols in memory 'DMw_bankC':
74
+ // Symbols in memory 'DMw_bankCD':
75
+ // Symbols in memory 'DMw_bankD':
76
+ // Symbols in memory 'DMw_stack':
77
+ // Symbols in memory 'DMx':
78
+ // Symbols in memory 'DMx_bankA':
79
+ // Symbols in memory 'DMx_bankAB':
80
+ // Symbols in memory 'DMx_bankAC':
81
+ // Symbols in memory 'DMx_bankAD':
82
+ // Symbols in memory 'DMx_bankB':
83
+ // Symbols in memory 'DMx_bankBC':
84
+ // Symbols in memory 'DMx_bankBD':
85
+ // Symbols in memory 'DMx_bankC':
86
+ // Symbols in memory 'DMx_bankCD':
87
+ // Symbols in memory 'DMx_bankD':
88
+ // Symbols in memory 'DMx_stack':
89
+ // Symbols in memory 'PM':
90
+ _symbol _Z13kernelWrapperPPvjjjj 0x00000930
91
+ _symbol _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh 0x00000b60
92
+ _symbol _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams 0x00001100
93
+ _symbol _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params 0x00001210
94
+ _symbol _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00001b80
95
+ _symbol _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv 0x00001fe0
96
+ _symbol _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E 0x00002020
97
+ _symbol _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x000021a0
98
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv 0x00002390
99
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E 0x000023d0
100
+ _symbol _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x000027a0
101
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv 0x00002990
102
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv 0x00002a10
103
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E 0x00002a80
104
+ _symbol _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00002a90
105
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E 0x00002cf0
106
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv 0x00002d10
107
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E 0x00002da0
108
+ _symbol _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00002ed0
109
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv 0x00003130
110
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv 0x000031b0
111
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E 0x00003200
112
+ _symbol _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00003210
113
+ _symbol _Z14_b1638_wrapperPPv 0x00003470
114
+ _symbol _Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params 0x00003490
115
+ _symbol _ZN12mllib_graphs17slice_adf_wrapperILi1E8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_EEvRNS2_9io_bufferIT0_NS2_9direction2inET1_EERNSD_ISE_NSF_3outET2_EERA7_Kj 0x000037a0
116
+ _symbol _Z13_b806_wrapperPPv 0x00003a50
117
+ _symbol _Z14_b1655_wrapperPPv 0x00003a70
118
+ _symbol _Z13_b891_wrapperPPv 0x00003aa0
119
+ _symbol _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj 0x00003ad0
120
+ _symbol _Z13_b820_wrapperPPv 0x00003e10
121
+ _symbol _Z13_b896_wrapperPPv 0x00003e40
122
+ _symbol _Z14_b1672_wrapperPPv 0x00003e70
123
+ _symbol _Z13_b886_wrapperPPv 0x00003e90
124
+ // Symbols in memory 'PMw':
125
+ // Symbols in memory 'TM4':
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.srv ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/Release/0_0_reloadable23.txt ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/scripts/0_0_reloadable23.bcf ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _reserved DMb 0x0 0x40000
2
+
3
+ _reserved PM 0x0 0x930 //reserved for main elf
4
+
5
+ _entry_point _Z13kernelWrapperPPvjjjj
6
+ _symbol _Z13kernelWrapperPPvjjjj 0x930
7
+
8
+ _reserved DMb 0x7b280 0x800 //reserved for lcp ping-pong buffers
9
+ _reserved DMb 0x7ba80 0x40 //reserved for sync buffer
10
+ _stack DM_stack 0x7bac0 0x940 //stack for core
11
+ _reserved DMb 0x7c400 0x40 //reserved for main elf heap
12
+ //space for synopsys compiler at 0x7c440 0x880//heap
13
+ _reserved DMb 0x40000 0x3b280
14
+
15
+ _reserved DMb 0x7ccc0 0x3340
16
+ _reserved DMb 0x80000 0x80000 // And everything else the core can't see
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/scripts/0_0_reloadable23.prx ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <project name="Project" processor="me">
2
+ <file type="lbc" name="0_0_reloadable23.ll" path="../../ir"/>
3
+ <issinit/>
4
+ <option id="cpp.define" value="__AIENGINE__ __AIE_ARCH__=21" inherit = "1"/>
5
+ <option id="llvm.xargs" value="-fno-jump-tables -fno-discard-value-names" inherit = "1"/>
6
+ <option id="llvm.lang" value="Follow file extension"/>
7
+ <option id="bridge.cfg" value="0_0_reloadable23.bcf"/>
8
+ <option id="cpp.include" value="&lt;XILINX_VITIS_AIETOOLS&gt;/include /app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend /usr/local/lib/python3.10/site-packages/include/aie_api /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf /app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/&lt;AIE_SRC_DIR&gt; " inherit="1"/>
9
+ <option id="project.dir" value="&lt;CONFIG&gt;/0_0_reloadable23"/>
10
+ <option id="project.name" value="0_0_reloadable23"/>
11
+ <option id="project.type" value="exe"/>
12
+ <option id="simrun.issca" value="&lt;XILINX_VITIS_AIETOOLS&gt;/bin/unwrapped/lnx64.o/aie2p/ca_udm_dbg" inherit="0"/>
13
+ </project>
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/src/0_0_reloadable23.cc ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Automatically generated processor driver using AIEngine tool-chain
2
+
3
+ #include <stdio.h>
4
+ #include <memory>
5
+ #include <adf.h>
6
+
7
+
8
+ // Declare Kernel functions and initializers
9
+ void superkernel_sigmoid1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
10
+ #include "slice_adf_wrapper.cpp"
11
+ void superkernel_sub1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
12
+ void superkernel_mul1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
13
+ #include "concat_adf_wrapper.cpp"
14
+ void conv2d_maxpool(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
15
+ void superkernel_tanh1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
16
+ void superkernel_add1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
17
+
18
+ // Declare Kernel objects and external arrays
19
+
20
+
21
+ void _b1638_wrapper(void* args[])
22
+ {
23
+ superkernel_sigmoid1d(
24
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
25
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
26
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
27
+ }
28
+
29
+ void _b806_wrapper(void* args[])
30
+ {
31
+ mllib_graphs::slice_adf_wrapper<1, bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>>(
32
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
33
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
34
+ *reinterpret_cast<uint32 const(*)[7]>(args[2]));
35
+ }
36
+
37
+ void _b1655_wrapper(void* args[])
38
+ {
39
+ superkernel_sub1d(
40
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
41
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
42
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
43
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
44
+ }
45
+
46
+ void _b891_wrapper(void* args[])
47
+ {
48
+ superkernel_mul1d(
49
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
50
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
51
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
52
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
53
+ }
54
+
55
+ void _b820_wrapper(void* args[])
56
+ {
57
+ mllib_graphs::concat_adf_wrapper<bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>>(
58
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
59
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
60
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[2]),
61
+ *reinterpret_cast<uint32 const(*)[8]>(args[3]));
62
+ }
63
+
64
+ void _b896_wrapper(void* args[])
65
+ {
66
+ conv2d_maxpool(
67
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
68
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
69
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
70
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
71
+ }
72
+
73
+ void _b1672_wrapper(void* args[])
74
+ {
75
+ superkernel_tanh1d(
76
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
77
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
78
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
79
+ }
80
+
81
+ void _b886_wrapper(void* args[])
82
+ {
83
+ superkernel_add1d(
84
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
85
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
86
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
87
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
88
+ }
89
+
90
+ using UniformKernelFunc = void (*)(void **);
91
+
92
+ static UniformKernelFunc g_uniformKernelFuncs[8] = {
93
+ _b1638_wrapper,
94
+ _b806_wrapper,
95
+ _b1655_wrapper,
96
+ _b891_wrapper,
97
+ _b820_wrapper,
98
+ _b896_wrapper,
99
+ _b1672_wrapper,
100
+ _b886_wrapper
101
+ };
102
+
103
+ __attribute__((always_inline)) void kernelWrapper(void* args[], uint32 kernelId, uint32 numSyncIn, uint32 numAsyncIn, uint32 numSyncOut)
104
+ {
105
+ uint32 idx = 0;
106
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 0);
107
+ idx += (numSyncIn > 0) ? 1 : 0;
108
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 1);
109
+ idx += (numSyncIn > 1) ? 1 : 0;
110
+ idx += numAsyncIn;
111
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::out>(numSyncOut > 0);
112
+ idx += (numSyncOut > 0) ? 1 : 0;
113
+
114
+ (*(g_uniformKernelFuncs[kernelId]))(args);
115
+
116
+ idx = 0;
117
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 0);
118
+ idx += (numSyncIn > 0) ? 1 : 0;
119
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 1);
120
+ idx += (numSyncIn > 1) ? 1 : 0;
121
+ idx += numAsyncIn;
122
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::out>(numSyncOut > 0);
123
+ idx += (numSyncOut > 0) ? 1 : 0;
124
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/timestamped_log/0_0_reloadable23.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable23/xlopt.log ADDED
@@ -0,0 +1,482 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
2
+
3
+
4
+ --------------- FILTER ANALYSIS INFO LOG ---------------
5
+
6
+ Reading Header IR from ir/_header.ll
7
+
8
+ --------------------------------------------------------
9
+
10
+ Add module pass 363{anonymous}::GuidancePass
11
+ Add module pass 363{anonymous}::ChessOptionsPass
12
+ Add module pass 363{anonymous}::DisableInliningInMainPass
13
+ Add module pass 363cdno::xlopt::AIEMergeSubWordStoresOpt
14
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::PromotePass>
15
+ Add module pass 363{anonymous}::IpConstPropPass
16
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::SimplifyCFGPass>
17
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::LoopSimplifyPass>
18
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::FunctionToLoopPassAdaptor>
19
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopInfoPass>
20
+ Add module pass 363cdno::xlopt::AIEAnnotatePragmaPass
21
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopPeelPass>
22
+ Add module pass 363{anonymous}::AIEAliasAnalysisPass
23
+
24
+
25
+ --------------- MEMORY MANAGEMENT GUIDANCE LOG ---------------
26
+
27
+ ALIGNMENT_HINT: Alignment of global array g_uniformKernelFuncs is 4 bytes; automatically aligning it to 64 bytes.
28
+ SIZE_HINT: Global array sigmoid_lut<0u, 256u>::data_ab is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
29
+ SIZE_HINT: Global array sigmoid_lut<0u, 256u>::data_cd is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
30
+ SIZE_HINT: Global array gelu_lut_32<0u, 512u>::data_ab is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
31
+ SIZE_HINT: Global array gelu_lut_32<0u, 512u>::data_cd is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
32
+ SIZE_HINT: Global array tanh_lut<0u, 512u>::data_ab is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
33
+ SIZE_HINT: Global array tanh_lut<0u, 512u>::data_cd is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
34
+ SIZE_HINT: Global array log_f32_lut<256u>::fraction_table_ab_f32 is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
35
+ SIZE_HINT: Global array log_f32_lut<256u>::fraction_table_cd_f32 is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
36
+ SIZE_HINT: Global array exp2_lut<512u>::exp2_table_ab is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
37
+ SIZE_HINT: Global array exp2_lut<512u>::exp2_table_cd is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
38
+
39
+ --------------------------------------------------------------
40
+
41
+
42
+
43
+ --------------- MERGING SUBWORD STORES OPT LOG ---------------
44
+
45
+
46
+
47
+ --------------- LOOP STATISTICS : _ZN18conv2d_bf16_paramsC2Ev ---------------
48
+
49
+ Total loops = 1
50
+ Loops with prepare for pipelining pragma = 0
51
+ Loops with unroll pragma = 0
52
+ Loops with flatten pragma = 0
53
+ Loops with min range pragma = 0
54
+ Loops with max range pragma = 0
55
+ Loops with known trip count = 0
56
+
57
+ -----------------------------------------------------------------------------
58
+
59
+
60
+
61
+ --------------- LOOP STATISTICS : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh ---------------
62
+
63
+ Total loops = 1
64
+ Loops with prepare for pipelining pragma = 0
65
+ Loops with unroll pragma = 0
66
+ Loops with flatten pragma = 0
67
+ Loops with min range pragma = 0
68
+ Loops with max range pragma = 0
69
+ Loops with known trip count = 1
70
+
71
+ -------------------------------------------------------------------------------------------------------------------
72
+
73
+
74
+
75
+ --------------- LOOP STATISTICS : _Z14conv2d_genericILh1EL5act_t0ELb0ELb1ELb0E8bfloat16EvPS1_S2_S2_S2_R18conv2d_bf16_params10out_mode_t ---------------
76
+
77
+ Total loops = 4
78
+ Loops with prepare for pipelining pragma = 4
79
+ Loops with unroll pragma = 0
80
+ Loops with flatten pragma = 0
81
+ Loops with min range pragma = 4
82
+ Loops with max range pragma = 0
83
+ Loops with known trip count = 0
84
+
85
+ -------------------------------------------------------------------------------------------------------------------------------------------------------
86
+
87
+
88
+
89
+ --------------- LOOP STATISTICS : _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams ---------------
90
+
91
+ Total loops = 1
92
+ Loops with prepare for pipelining pragma = 1
93
+ Loops with unroll pragma = 0
94
+ Loops with flatten pragma = 0
95
+ Loops with min range pragma = 1
96
+ Loops with max range pragma = 0
97
+ Loops with known trip count = 0
98
+
99
+ -------------------------------------------------------------------------------------------------------------------
100
+
101
+
102
+
103
+ --------------- LOOP STATISTICS : _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E ---------------
104
+
105
+ Total loops = 1
106
+ Loops with prepare for pipelining pragma = 1
107
+ Loops with unroll pragma = 0
108
+ Loops with flatten pragma = 0
109
+ Loops with min range pragma = 1
110
+ Loops with max range pragma = 0
111
+ Loops with known trip count = 0
112
+
113
+ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
114
+
115
+
116
+
117
+ --------------- LOOP STATISTICS : _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj32EEELj2EEC2Ev ---------------
118
+
119
+ Total loops = 1
120
+ Loops with prepare for pipelining pragma = 0
121
+ Loops with unroll pragma = 0
122
+ Loops with flatten pragma = 0
123
+ Loops with min range pragma = 0
124
+ Loops with max range pragma = 0
125
+ Loops with known trip count = 0
126
+
127
+ --------------------------------------------------------------------------------------------------------
128
+
129
+
130
+
131
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
132
+
133
+ Total loops = 1
134
+ Loops with prepare for pipelining pragma = 1
135
+ Loops with unroll pragma = 0
136
+ Loops with flatten pragma = 0
137
+ Loops with min range pragma = 1
138
+ Loops with max range pragma = 0
139
+ Loops with known trip count = 0
140
+
141
+ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
142
+
143
+
144
+
145
+ --------------- LOOP STATISTICS : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE ---------------
146
+
147
+ Total loops = 1
148
+ Loops with prepare for pipelining pragma = 1
149
+ Loops with unroll pragma = 0
150
+ Loops with flatten pragma = 0
151
+ Loops with min range pragma = 1
152
+ Loops with max range pragma = 0
153
+ Loops with known trip count = 0
154
+
155
+ ---------------------------------------------------------------------------------------------------------------------------------------------------------------
156
+
157
+
158
+
159
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
160
+
161
+ Total loops = 1
162
+ Loops with prepare for pipelining pragma = 1
163
+ Loops with unroll pragma = 0
164
+ Loops with flatten pragma = 0
165
+ Loops with min range pragma = 1
166
+ Loops with max range pragma = 0
167
+ Loops with known trip count = 0
168
+
169
+ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
170
+
171
+
172
+
173
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
174
+
175
+ Total loops = 1
176
+ Loops with prepare for pipelining pragma = 1
177
+ Loops with unroll pragma = 0
178
+ Loops with flatten pragma = 0
179
+ Loops with min range pragma = 1
180
+ Loops with max range pragma = 0
181
+ Loops with known trip count = 0
182
+
183
+ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
184
+
185
+
186
+
187
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj32EEC2Eb ---------------
188
+
189
+ Total loops = 2
190
+ Loops with prepare for pipelining pragma = 0
191
+ Loops with unroll pragma = 2
192
+ Loops with flatten pragma = 0
193
+ Loops with min range pragma = 0
194
+ Loops with max range pragma = 0
195
+ Loops with known trip count = 2
196
+
197
+ -------------------------------------------------------------------------
198
+
199
+
200
+
201
+ --------------- LOOP STATISTICS : _Z9avgpool2dILh1E8bfloat16Qsr5mllib5utilsE11is_one_of_vIT0_ahS0_EEvPS1_S2_R25avgpool2d_internal_paramsIS1_E ---------------
202
+
203
+ Total loops = 1
204
+ Loops with prepare for pipelining pragma = 1
205
+ Loops with unroll pragma = 0
206
+ Loops with flatten pragma = 0
207
+ Loops with min range pragma = 1
208
+ Loops with max range pragma = 0
209
+ Loops with known trip count = 0
210
+
211
+ -------------------------------------------------------------------------------------------------------------------------------------------------------------
212
+
213
+
214
+
215
+ --------------- LOOP STATISTICS : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E ---------------
216
+
217
+ Total loops = 1
218
+ Loops with prepare for pipelining pragma = 1
219
+ Loops with unroll pragma = 0
220
+ Loops with flatten pragma = 0
221
+ Loops with min range pragma = 1
222
+ Loops with max range pragma = 0
223
+ Loops with known trip count = 0
224
+
225
+ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
226
+
227
+
228
+
229
+ --------------- LOOP STATISTICS : _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh ---------------
230
+
231
+ Total loops = 1
232
+ Loops with prepare for pipelining pragma = 0
233
+ Loops with unroll pragma = 0
234
+ Loops with flatten pragma = 0
235
+ Loops with min range pragma = 0
236
+ Loops with max range pragma = 0
237
+ Loops with known trip count = 1
238
+
239
+ --------------------------------------------------------------------------------------------------------------
240
+
241
+
242
+
243
+ --------------- LOOP STATISTICS : _Z9conv2d_dwILh1E8bfloat16S0_S0_N3adf16io_buffer_configINS1_7extentsIJEEENS1_7locking4syncENS1_10addressing6linearENS1_6marginILj0EEEEESB_NS2_IS4_NS5_5asyncES8_SA_EEQsr3stdE9is_same_vIT0_S0_EEvRNS1_9io_bufferISE_NS1_9direction2inET3_EERNSF_IT1_SH_T4_EERNSF_IT2_NSG_3outET5_EER21conv2d_dw_bf16_params ---------------
244
+
245
+ Total loops = 2
246
+ Loops with prepare for pipelining pragma = 2
247
+ Loops with unroll pragma = 0
248
+ Loops with flatten pragma = 0
249
+ Loops with min range pragma = 2
250
+ Loops with max range pragma = 0
251
+ Loops with known trip count = 0
252
+
253
+ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
254
+
255
+
256
+
257
+ --------------- LOOP STATISTICS : _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE3runEPS0_S6_R18reduce_c8_params_tIS4_E ---------------
258
+
259
+ Total loops = 2
260
+ Loops with prepare for pipelining pragma = 1
261
+ Loops with unroll pragma = 0
262
+ Loops with flatten pragma = 0
263
+ Loops with min range pragma = 1
264
+ Loops with max range pragma = 0
265
+ Loops with known trip count = 0
266
+
267
+ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
268
+
269
+
270
+
271
+ --------------- LOOP STATISTICS : _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj16EEELj4EEC2Ev ---------------
272
+
273
+ Total loops = 1
274
+ Loops with prepare for pipelining pragma = 0
275
+ Loops with unroll pragma = 0
276
+ Loops with flatten pragma = 0
277
+ Loops with min range pragma = 0
278
+ Loops with max range pragma = 0
279
+ Loops with known trip count = 0
280
+
281
+ --------------------------------------------------------------------------------------------------------
282
+
283
+
284
+
285
+ --------------- LOOP STATISTICS : _ZN19reduce_mean_c8_implI8bfloat16E16body_postprocessEPvR18reduce_c8_params_tI23reduce_mean_c8_params_tIS0_EE ---------------
286
+
287
+ Total loops = 2
288
+ Loops with prepare for pipelining pragma = 2
289
+ Loops with unroll pragma = 0
290
+ Loops with flatten pragma = 0
291
+ Loops with min range pragma = 2
292
+ Loops with max range pragma = 0
293
+ Loops with known trip count = 0
294
+
295
+ ---------------------------------------------------------------------------------------------------------------------------------------------------------------
296
+
297
+
298
+
299
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj8EEC2Eb ---------------
300
+
301
+ Total loops = 2
302
+ Loops with prepare for pipelining pragma = 0
303
+ Loops with unroll pragma = 2
304
+ Loops with flatten pragma = 0
305
+ Loops with min range pragma = 0
306
+ Loops with max range pragma = 0
307
+ Loops with known trip count = 2
308
+
309
+ ------------------------------------------------------------------------
310
+
311
+
312
+
313
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj16EEC2Eb ---------------
314
+
315
+ Total loops = 2
316
+ Loops with prepare for pipelining pragma = 0
317
+ Loops with unroll pragma = 2
318
+ Loops with flatten pragma = 0
319
+ Loops with min range pragma = 0
320
+ Loops with max range pragma = 0
321
+ Loops with known trip count = 2
322
+
323
+ -------------------------------------------------------------------------
324
+
325
+
326
+
327
+ --------------- LOOP STATISTICS : _ZNK3aie4maskILj8EEeqERKS1_ ---------------
328
+
329
+ Total loops = 1
330
+ Loops with prepare for pipelining pragma = 0
331
+ Loops with unroll pragma = 1
332
+ Loops with flatten pragma = 0
333
+ Loops with min range pragma = 0
334
+ Loops with max range pragma = 0
335
+ Loops with known trip count = 1
336
+
337
+ -----------------------------------------------------------------------------
338
+
339
+
340
+
341
+ --------------- LOOP STATISTICS : _ZNK3aie4maskILj16EEeqERKS1_ ---------------
342
+
343
+ Total loops = 1
344
+ Loops with prepare for pipelining pragma = 0
345
+ Loops with unroll pragma = 1
346
+ Loops with flatten pragma = 0
347
+ Loops with min range pragma = 0
348
+ Loops with max range pragma = 0
349
+ Loops with known trip count = 1
350
+
351
+ ------------------------------------------------------------------------------
352
+
353
+
354
+
355
+ --------------- LOOP STATISTICS : _Z26pad_3d_bfloat16_simplifiedP8bfloat16S0_R15pad_3d_params_t ---------------
356
+
357
+ Total loops = 3
358
+ Loops with prepare for pipelining pragma = 0
359
+ Loops with unroll pragma = 0
360
+ Loops with flatten pragma = 0
361
+ Loops with min range pragma = 0
362
+ Loops with max range pragma = 0
363
+ Loops with known trip count = 0
364
+
365
+ ---------------------------------------------------------------------------------------------------------------
366
+
367
+
368
+
369
+ --------------- LOOP STATISTICS : _Z11slice_hcwc8I8bfloat16EvPT_S2_R18slice_hcwc8_params ---------------
370
+
371
+ Total loops = 2
372
+ Loops with prepare for pipelining pragma = 1
373
+ Loops with unroll pragma = 0
374
+ Loops with flatten pragma = 0
375
+ Loops with min range pragma = 0
376
+ Loops with max range pragma = 0
377
+ Loops with known trip count = 0
378
+
379
+ --------------------------------------------------------------------------------------------------------
380
+
381
+
382
+
383
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj64EEC2Eb ---------------
384
+
385
+ Total loops = 2
386
+ Loops with prepare for pipelining pragma = 0
387
+ Loops with unroll pragma = 2
388
+ Loops with flatten pragma = 0
389
+ Loops with min range pragma = 0
390
+ Loops with max range pragma = 0
391
+ Loops with known trip count = 2
392
+
393
+ -------------------------------------------------------------------------
394
+
395
+
396
+
397
+ --------------- LOOP STATISTICS : _ZNK3aie4maskILj64EEeqERKS1_ ---------------
398
+
399
+ Total loops = 1
400
+ Loops with prepare for pipelining pragma = 0
401
+ Loops with unroll pragma = 1
402
+ Loops with flatten pragma = 0
403
+ Loops with min range pragma = 0
404
+ Loops with max range pragma = 0
405
+ Loops with known trip count = 1
406
+
407
+ ------------------------------------------------------------------------------
408
+
409
+
410
+
411
+ --------------- LOOP STATISTICS : _Z7maskgenIyLj16EET_jj ---------------
412
+
413
+ Total loops = 1
414
+ Loops with prepare for pipelining pragma = 0
415
+ Loops with unroll pragma = 0
416
+ Loops with flatten pragma = 0
417
+ Loops with min range pragma = 0
418
+ Loops with max range pragma = 0
419
+ Loops with known trip count = 1
420
+
421
+ ------------------------------------------------------------------------
422
+
423
+
424
+
425
+ --------------- LOOP STATISTICS : _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj ---------------
426
+
427
+ Total loops = 4
428
+ Loops with prepare for pipelining pragma = 0
429
+ Loops with unroll pragma = 0
430
+ Loops with flatten pragma = 0
431
+ Loops with min range pragma = 0
432
+ Loops with max range pragma = 0
433
+ Loops with known trip count = 0
434
+
435
+ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
436
+
437
+
438
+
439
+ --------------- PRAGMA INSERTION LOG (unroll threshold = 3000, max unroll factor = 1) ---------------
440
+
441
+ Adding pragma to function _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh :
442
+ Pragma added to loop at line 437 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common/../../include/conv/conv2d_bf16_params.h : chess_loop_range(8,8)
443
+ Adding pragma to function _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj32EEELj2EEC2Ev :
444
+ Pragma added to loop at line 135 in file /usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite/../include/array : chess_prepare_for_pipelining
445
+ Adding pragma to function _ZN3aie4maskILj32EEC2Eb :
446
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
447
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
448
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
449
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
450
+ Adding pragma to function _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh :
451
+ Pragma added to loop at line 214 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common/../../include/conv/conv2d_dw_bf16_params.h : chess_loop_range(4,4)
452
+ Adding pragma to function _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj16EEELj4EEC2Ev :
453
+ Pragma added to loop at line 135 in file /usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite/../include/array : chess_prepare_for_pipelining
454
+ Adding pragma to function _ZN3aie4maskILj8EEC2Eb :
455
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
456
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
457
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
458
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
459
+ Adding pragma to function _ZN3aie4maskILj16EEC2Eb :
460
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
461
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
462
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
463
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
464
+ Adding pragma to function _ZNK3aie4maskILj16EEeqERKS1_ :
465
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
466
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
467
+ Adding pragma to function _ZNK3aie4maskILj8EEeqERKS1_ :
468
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
469
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
470
+ Adding pragma to function _Z7maskgenIyLj16EET_jj :
471
+ Pragma added to loop at line 25 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc/slice_hcwc8_params.h : chess_loop_range(4,4)
472
+ Adding pragma to function _ZN3aie4maskILj64EEC2Eb :
473
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
474
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
475
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(2,2)
476
+ Adding pragma to function _ZNK3aie4maskILj64EEeqERKS1_ :
477
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(2,2)
478
+ Adding pragma to function _ZN12mllib_graphs18concat_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_SC_EEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_SG_T1_EERNSD_ISE_NSF_3outET2_EERA8_Kj :
479
+ Pragma added to loop at line 112 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf/concat_adf_wrapper.cpp : chess_prepare_for_pipelining
480
+
481
+ -----------------------------------------------------------------------------------------------------
482
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/0_0_reloadable24.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.# ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 1d9bafb03ba3e7e6997e128da917dae80aa02b0b
2
+ 1a735f496f1284ab86135ede88939a8aeeb375a3
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.## ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 63337a66bfb0c84b1afa7a155502d28566676a87
2
+ 92e103875c6a39bc8cdbe4a21c02c5a94ba3cc96
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.calltree ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:03:19 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable24 ../Release/0_0_reloadable24.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable24.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6947 -pme
5
+
6
+
7
+ // Release: ipp V-2024.06-TGT-241219
8
+
9
+ _Z13kernelWrapperPPvjjjj
10
+ _Z13_b853_wrapperPPv (referenced text)
11
+ _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj
12
+ _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj
13
+ _Z13_b896_wrapperPPv (referenced text)
14
+ _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
15
+ _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
16
+ _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
17
+ _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
18
+ _Z14_b1638_wrapperPPv (referenced text)
19
+ _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
20
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
21
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
22
+ _Z14_b1655_wrapperPPv (referenced text)
23
+ _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
24
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
25
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
26
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
27
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
28
+ _Z13_b891_wrapperPPv (referenced text)
29
+ _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
30
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
31
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
32
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
33
+ _Z14_b1672_wrapperPPv (referenced text)
34
+ _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
35
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
36
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
37
+ _Z13_b886_wrapperPPv (referenced text)
38
+ _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
39
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
40
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
41
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
42
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (*)
43
+
44
+
45
+ Call tree stack and functions sizes:
46
+
47
+ stack stack stack call func func function name
48
+ desc level level desc
49
+ ----- ----- ----- ----- ----- ----- --------------------------------------------------------------
50
+ 64 512 0 0 546 13188 _Z13kernelWrapperPPvjjjj
51
+ 0 0 1 1 32 2086 _Z13_b853_wrapperPPv
52
+ 0 0 1 2 470 2054 _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj <tail call>
53
+ 0 0 1 3 1584 1584 _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj <tail call>
54
+ 0 192 1 1 36 4714 _Z13_b896_wrapperPPv
55
+ 64 192 1 2 568 4678 _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
56
+ 64 64 2 3 1430 1430 _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
57
+ 128 128 2 3 2410 2680 _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
58
+ 0 0 3 4 270 270 _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
59
+ 0 64 1 1 32 960 _Z14_b1638_wrapperPPv
60
+ 64 64 1 2 488 928 _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
61
+ 0 0 2 3 62 62 _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
62
+ 0 0 2 3 378 378 _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
63
+ 0 192 1 1 36 1378 _Z14_b1655_wrapperPPv
64
+ 64 192 1 2 602 1342 _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
65
+ 64 64 2 3 74 190 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
66
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
67
+ 0 128 2 3 16 550 _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
68
+ 128 128 2 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE <tail call>
69
+ 0 128 1 1 36 1092 _Z13_b891_wrapperPPv
70
+ 64 128 1 2 602 1056 _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
71
+ 64 64 2 3 138 162 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
72
+ 0 0 3 4 24 24 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
73
+ 0 0 2 3 292 292 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
74
+ 0 448 1 1 32 1544 _Z14_b1672_wrapperPPv
75
+ 64 448 1 2 488 1512 _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
76
+ 0 0 2 3 62 62 _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
77
+ 384 384 2 3 962 962 _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
78
+ 0 192 1 1 36 1402 _Z13_b886_wrapperPPv
79
+ 64 192 1 2 602 1366 _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE <tail call>
80
+ 64 64 2 3 98 214 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
81
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
82
+ 0 128 2 3 16 550 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
83
+ 128 128 2 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE <tail call>
84
+
85
+
86
+ Maximum call level : 4
87
+ Maximum stack level: 3
88
+ Maximum stack size : 512
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.cmic2 ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.cmico ADDED
@@ -0,0 +1 @@
 
 
1
+ +Mdec
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.lst ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.map ADDED
@@ -0,0 +1,294 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:03:20 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable24 ../Release/0_0_reloadable24.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable24.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6947 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ Memory map for memory 'DM_stack':
9
+
10
+ Size = 1048576
11
+ Width = 8 bits
12
+ Offset = 0
13
+ Used = 2368
14
+
15
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
16
+
17
+ Memory map for memory 'DMb':
18
+
19
+ Size = 1048576
20
+ Width = 8 bits
21
+ Offset = 0
22
+ Used = 3397
23
+
24
+ 0x00000000..0x0007babf ( 506560 items) : Reserved
25
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
26
+ 0x0007c400..0x0007c43f ( 64 items) : Reserved
27
+ 0x0007c440..0x0007c443 ( 4 items) : ../Release/0_0_reloadable24.o::_ZL9curr_iter (Data, Local, .bss.DMb.4)
28
+ 0x0007c444..0x0007c447 ( 4 items) : ../Release/0_0_reloadable24.o::_ZL14num_depth_iter (Data, Local, .bss.DMb.4)
29
+ 0x0007c448..0x0007c44b ( 4 items) : ../Release/0_0_reloadable24.o::_ZL10depth_iter (Data, Local, .bss.DMb.4)
30
+ 0x0007c44c..0x0007c44f ( 4 items) : ../Release/0_0_reloadable24.o::_ZL11total_iters (Data, Local, .bss.DMb.4)
31
+ 0x0007c450..0x0007c453 ( 4 items) : ../Release/0_0_reloadable24.o::_ZL8core_row (Data, Local, .bss.DMb.4)
32
+ 0x0007c454..0x0007c457 ( 4 items) : ../Release/0_0_reloadable24.o::_ZL11ifm1_offset (Data, Local, .bss.DMb.4)
33
+ 0x0007c458..0x0007c45b ( 4 items) : ../Release/0_0_reloadable24.o::_ZL11ifm2_offset (Data, Local, .bss.DMb.4)
34
+ 0x0007c45c..0x0007c45f ( 4 items) : ../Release/0_0_reloadable24.o::_ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE8cur_iter (Data, Weak, .bss.DMb.4)
35
+ 0x0007c460..0x0007c463 ( 4 items) : ../Release/0_0_reloadable24.o::_ZL8num_iter (Data, Local, .data.DMb.4)
36
+ 0x0007c464..0x0007c467 ( 4 items) : me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_ZN12me_primitive11control_satE (Data, Global, .bss.DMb.4)
37
+ 0x0007c468..0x0007c468 ( 1 items) : me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_ZN12me_primitive11control_rndE (Data, Global, .bss.DMb.1)
38
+ 0x0007c480..0x0007c49b ( 28 items) : ../Release/0_0_reloadable24.o::_ZL20g_uniformKernelFuncs (Data, Local, .rodata.DMb.64)
39
+
40
+ Called functions : _Z13_b853_wrapperPPv
41
+ _Z13_b896_wrapperPPv
42
+ _Z14_b1638_wrapperPPv
43
+ _Z14_b1655_wrapperPPv
44
+ _Z13_b891_wrapperPPv
45
+ _Z14_b1672_wrapperPPv
46
+ _Z13_b886_wrapperPPv
47
+
48
+ 0x0007c4c0..0x0007c4df ( 32 items) : ../Release/0_0_reloadable24.o::__const____Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj___coeffs (Data, Local, .rodata.DMb.64)
49
+ 0x0007c500..0x0007c51f ( 32 items) : ../Release/0_0_reloadable24.o::__const____Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj___coeffs_asymmetric (Data, Local, .rodata.DMb.64)
50
+ 0x0007c540..0x0007c57f ( 64 items) : ../Release/0_0_reloadable24.o::add1d_params (Data, Global, .bss.DMb.64)
51
+ 0x0007c580..0x0007c5bf ( 64 items) : ../Release/0_0_reloadable24.o::mul1d_params (Data, Global, .bss.DMb.64)
52
+ 0x0007c5c0..0x0007c5ff ( 64 items) : ../Release/0_0_reloadable24.o::sub1d_params (Data, Global, .bss.DMb.64)
53
+ 0x0007c600..0x0007c63f ( 64 items) : ../Release/0_0_reloadable24.o::sigmoid1d_params (Data, Global, .bss.DMb.64)
54
+ 0x0007c640..0x0007c67f ( 64 items) : ../Release/0_0_reloadable24.o::tanh1d_params (Data, Global, .bss.DMb.64)
55
+ 0x0007c680..0x0007c83f ( 448 items) : ../Release/0_0_reloadable24.o::conv2d_params (Data, Global, .bss.DMb.64)
56
+ 0x0007c840..0x0007c8bf ( 128 items) : ../Release/0_0_reloadable24.o::_ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE6params (Data, Weak, .bss.DMb.64)
57
+ 0x0007ccc0..0x000fffff ( 537408 items) : Reserved
58
+
59
+ Memory map for memory 'PM':
60
+
61
+ Size = 1048576
62
+ Width = 8 bits
63
+ Offset = 0
64
+ Used = 13188
65
+
66
+ 0x00000000..0x0000092f ( 2352 items) : Reserved
67
+ 0x00000930..0x00000b51 ( 546 items) : ../Release/0_0_reloadable24.o::_Z13kernelWrapperPPvjjjj (Function, Global, .text) (stack frame size = 64)
68
+
69
+ Referenced symbols: _ZL20g_uniformKernelFuncs
70
+
71
+ 0x00000b60..0x000010f5 ( 1430 items) : ../Release/0_0_reloadable24.o::_Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh (Function, Weak, .text) (stack frame size = 64)
72
+ 0x00001100..0x0000120d ( 270 items) : ../Release/0_0_reloadable24.o::_Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams (Function, Weak, .text) (stack frame size = 0)
73
+
74
+ Referenced symbols: _ZN12me_primitive11control_rndE
75
+
76
+ 0x00001210..0x00001b79 ( 2410 items) : ../Release/0_0_reloadable24.o::_Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params (Function, Weak, .text) (stack frame size = 128)
77
+
78
+ Called functions : _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
79
+
80
+ Referenced symbols: _ZN12me_primitive11control_rndE
81
+
82
+ 0x00001b80..0x00001db7 ( 568 items) : ../Release/0_0_reloadable24.o::_Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
83
+
84
+ Called functions : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
85
+ _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params
86
+
87
+ Referenced symbols: _ZL9curr_iter
88
+ conv2d_params
89
+ _ZN12me_primitive11control_rndE
90
+ _ZN12me_primitive11control_satE
91
+ _ZL14num_depth_iter
92
+ _ZL8num_iter
93
+ _ZL10depth_iter
94
+ _ZL11total_iters
95
+
96
+ 0x00001dc0..0x00001fd5 ( 534 items) : ../Release/0_0_reloadable24.o::_ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (Function, Local, .text) (stack frame size = 128)
97
+
98
+ Referenced symbols: _ZN12me_primitive11control_rndE
99
+
100
+ 0x00001fe0..0x0000201d ( 62 items) : ../Release/0_0_reloadable24.o::_ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
101
+ 0x00002020..0x00002199 ( 378 items) : ../Release/0_0_reloadable24.o::_ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
102
+
103
+ Referenced symbols: _ZN12me_primitive11control_rndE
104
+
105
+ 0x000021a0..0x00002387 ( 488 items) : ../Release/0_0_reloadable24.o::_Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
106
+
107
+ Called functions : _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
108
+ _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
109
+
110
+ Referenced symbols: _ZL9curr_iter
111
+ _ZL8core_row
112
+ sigmoid1d_params
113
+ _ZN12me_primitive11control_rndE
114
+ _ZN12me_primitive11control_satE
115
+ _ZL11ifm1_offset
116
+ _ZL8num_iter
117
+
118
+ 0x00002390..0x000023cd ( 62 items) : ../Release/0_0_reloadable24.o::_ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
119
+ 0x000023d0..0x00002791 ( 962 items) : ../Release/0_0_reloadable24.o::_ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E (Function, Weak, .text) (stack frame size = 384)
120
+
121
+ Referenced symbols: _ZN12me_primitive11control_rndE
122
+
123
+ 0x000027a0..0x00002987 ( 488 items) : ../Release/0_0_reloadable24.o::_Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE (Function, Global, .text) (stack frame size = 64)
124
+
125
+ Called functions : _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
126
+ _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
127
+
128
+ Referenced symbols: _ZL9curr_iter
129
+ _ZL8core_row
130
+ tanh1d_params
131
+ _ZN12me_primitive11control_rndE
132
+ _ZN12me_primitive11control_satE
133
+ _ZL11ifm1_offset
134
+ _ZL8num_iter
135
+
136
+ 0x00002990..0x00002a03 ( 116 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 0)
137
+ 0x00002a10..0x00002a71 ( 98 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 64)
138
+
139
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
140
+
141
+ 0x00002a80..0x00002a8f ( 16 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 0)
142
+
143
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
144
+
145
+ 0x00002a90..0x00002ce9 ( 602 items) : ../Release/0_0_reloadable24.o::_Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
146
+
147
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
148
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
149
+
150
+ Referenced symbols: _ZL9curr_iter
151
+ _ZL8core_row
152
+ add1d_params
153
+ _ZN12me_primitive11control_rndE
154
+ _ZN12me_primitive11control_satE
155
+ _ZL11ifm1_offset
156
+ _ZL11ifm2_offset
157
+ _ZL8num_iter
158
+
159
+ 0x00002cf0..0x00002d07 ( 24 items) : ../Release/0_0_reloadable24.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
160
+ 0x00002d10..0x00002d99 ( 138 items) : ../Release/0_0_reloadable24.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
161
+
162
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
163
+
164
+ 0x00002da0..0x00002ec3 ( 292 items) : ../Release/0_0_reloadable24.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
165
+
166
+ Referenced symbols: _ZN12me_primitive11control_rndE
167
+
168
+ 0x00002ed0..0x00003129 ( 602 items) : ../Release/0_0_reloadable24.o::_Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
169
+
170
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
171
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
172
+
173
+ Referenced symbols: _ZL9curr_iter
174
+ _ZL8core_row
175
+ mul1d_params
176
+ _ZN12me_primitive11control_rndE
177
+ _ZN12me_primitive11control_satE
178
+ _ZL11ifm1_offset
179
+ _ZL11ifm2_offset
180
+ _ZL8num_iter
181
+
182
+ 0x00003130..0x000031a3 ( 116 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
183
+ 0x000031b0..0x000031f9 ( 74 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
184
+
185
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
186
+
187
+ 0x00003200..0x0000320f ( 16 items) : ../Release/0_0_reloadable24.o::_ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
188
+
189
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
190
+
191
+ 0x00003210..0x00003469 ( 602 items) : ../Release/0_0_reloadable24.o::_Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE (Function, Global, .text) (stack frame size = 64)
192
+
193
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
194
+ _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E
195
+
196
+ Referenced symbols: _ZL9curr_iter
197
+ _ZL8core_row
198
+ sub1d_params
199
+ _ZN12me_primitive11control_rndE
200
+ _ZN12me_primitive11control_satE
201
+ _ZL11ifm1_offset
202
+ _ZL11ifm2_offset
203
+ _ZL8num_iter
204
+
205
+ 0x00003470..0x00003a9f ( 1584 items) : ../Release/0_0_reloadable24.o::_Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj (Function, Weak, .text) (stack frame size = 0)
206
+
207
+ Referenced symbols: __const____Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj___coeffs
208
+ __const____Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj___coeffs_asymmetric
209
+ _ZN12me_primitive11control_rndE
210
+
211
+ 0x00003aa0..0x00003c75 ( 470 items) : ../Release/0_0_reloadable24.o::_ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj (Function, Weak, .text) (stack frame size = 0)
212
+
213
+ Called functions : _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj
214
+
215
+ Referenced symbols: _ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE8cur_iter
216
+ _ZN12me_primitive11control_satE
217
+ _ZN12me_primitive11control_rndE
218
+ _ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE6params
219
+
220
+ 0x00003c80..0x00003c9f ( 32 items) : ../Release/0_0_reloadable24.o::_Z13_b853_wrapperPPv (Function, Global, .text) (stack frame size = 0)
221
+
222
+ Called functions : _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj
223
+
224
+ 0x00003ca0..0x00003cc3 ( 36 items) : ../Release/0_0_reloadable24.o::_Z13_b896_wrapperPPv (Function, Global, .text) (stack frame size = 0)
225
+
226
+ Called functions : _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
227
+
228
+ 0x00003cd0..0x00003cef ( 32 items) : ../Release/0_0_reloadable24.o::_Z14_b1638_wrapperPPv (Function, Global, .text) (stack frame size = 0)
229
+
230
+ Called functions : _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
231
+
232
+ 0x00003cf0..0x00003d13 ( 36 items) : ../Release/0_0_reloadable24.o::_Z14_b1655_wrapperPPv (Function, Global, .text) (stack frame size = 0)
233
+
234
+ Called functions : _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
235
+
236
+ 0x00003d20..0x00003d43 ( 36 items) : ../Release/0_0_reloadable24.o::_Z13_b891_wrapperPPv (Function, Global, .text) (stack frame size = 0)
237
+
238
+ Called functions : _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
239
+
240
+ 0x00003d50..0x00003d6f ( 32 items) : ../Release/0_0_reloadable24.o::_Z14_b1672_wrapperPPv (Function, Global, .text) (stack frame size = 0)
241
+
242
+ Called functions : _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
243
+
244
+ 0x00003d70..0x00003d93 ( 36 items) : ../Release/0_0_reloadable24.o::_Z13_b886_wrapperPPv (Function, Global, .text) (stack frame size = 0)
245
+
246
+ Called functions : _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
247
+
248
+
249
+ External symbols:
250
+
251
+ __dso_handle = 0x0
252
+ _ctors_end = 0x0
253
+ _ctors_start = 0x0
254
+ _dtors_end = 0x0
255
+ _dtors_start = 0x0
256
+ _pc_end = 0x3d94
257
+ _pc_start = 0x930
258
+ _sp_end_DM_stack = 0x7c400
259
+ _sp_start_DM_stack = 0x7bac0
260
+
261
+ Section summary for memory 'DM_stack':
262
+
263
+ .stack File
264
+ ---------- ----------
265
+ 2368
266
+ ---------- ----------
267
+ 2368 Total
268
+
269
+ Section summary for memory 'DMb':
270
+
271
+ .bss .data .rodata File
272
+ ---------- ---------- ---------- ----------
273
+ 928 4 92 ../Release/0_0_reloadable24.o
274
+ 5 0 0 me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
275
+ ---------- ---------- ---------- ----------
276
+ 933 4 92 Total
277
+
278
+ Section summary for memory 'PM':
279
+
280
+ .text File
281
+ ---------- ----------
282
+ 13188 ../Release/0_0_reloadable24.o
283
+ ---------- ----------
284
+ 13188 Total
285
+
286
+ File summary:
287
+
288
+ ../Release/0_0_reloadable24.o
289
+ DMb 1024
290
+ PM 13188
291
+
292
+ me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
293
+ DMb 5
294
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.o.lst ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.sdr ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 04:03:20 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable24 ../Release/0_0_reloadable24.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0_reloadable24.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork6947 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ // Symbols in memory 'DM_bankA':
9
+ // Symbols in memory 'DM_bankAB':
10
+ // Symbols in memory 'DM_bankAC':
11
+ // Symbols in memory 'DM_bankAD':
12
+ // Symbols in memory 'DM_bankB':
13
+ // Symbols in memory 'DM_bankBC':
14
+ // Symbols in memory 'DM_bankBD':
15
+ // Symbols in memory 'DM_bankC':
16
+ // Symbols in memory 'DM_bankCD':
17
+ // Symbols in memory 'DM_bankD':
18
+ // Symbols in memory 'DM_stack':
19
+ // Symbols in memory 'DM_test':
20
+ // Symbols in memory 'DMb':
21
+ _symbol _ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE8cur_iter 0x0007c45c
22
+ _symbol _ZN12me_primitive11control_satE 0x0007c464
23
+ _symbol _ZN12me_primitive11control_rndE 0x0007c468
24
+ _symbol add1d_params 0x0007c540
25
+ _symbol mul1d_params 0x0007c580
26
+ _symbol sub1d_params 0x0007c5c0
27
+ _symbol sigmoid1d_params 0x0007c600
28
+ _symbol tanh1d_params 0x0007c640
29
+ _symbol conv2d_params 0x0007c680
30
+ _symbol _ZZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_KjE6params 0x0007c840
31
+ // Symbols in memory 'DMh':
32
+ // Symbols in memory 'DMh_bankA':
33
+ // Symbols in memory 'DMh_bankAB':
34
+ // Symbols in memory 'DMh_bankAC':
35
+ // Symbols in memory 'DMh_bankAD':
36
+ // Symbols in memory 'DMh_bankB':
37
+ // Symbols in memory 'DMh_bankBC':
38
+ // Symbols in memory 'DMh_bankBD':
39
+ // Symbols in memory 'DMh_bankC':
40
+ // Symbols in memory 'DMh_bankCD':
41
+ // Symbols in memory 'DMh_bankD':
42
+ // Symbols in memory 'DMh_stack':
43
+ // Symbols in memory 'DMs':
44
+ // Symbols in memory 'DMs_bankA':
45
+ // Symbols in memory 'DMs_bankAB':
46
+ // Symbols in memory 'DMs_bankAC':
47
+ // Symbols in memory 'DMs_bankAD':
48
+ // Symbols in memory 'DMs_bankB':
49
+ // Symbols in memory 'DMs_bankBC':
50
+ // Symbols in memory 'DMs_bankBD':
51
+ // Symbols in memory 'DMs_bankC':
52
+ // Symbols in memory 'DMs_bankCD':
53
+ // Symbols in memory 'DMs_bankD':
54
+ // Symbols in memory 'DMs_stack':
55
+ // Symbols in memory 'DMv':
56
+ // Symbols in memory 'DMv_bankA':
57
+ // Symbols in memory 'DMv_bankAB':
58
+ // Symbols in memory 'DMv_bankAC':
59
+ // Symbols in memory 'DMv_bankAD':
60
+ // Symbols in memory 'DMv_bankB':
61
+ // Symbols in memory 'DMv_bankBC':
62
+ // Symbols in memory 'DMv_bankBD':
63
+ // Symbols in memory 'DMv_bankC':
64
+ // Symbols in memory 'DMv_bankCD':
65
+ // Symbols in memory 'DMv_bankD':
66
+ // Symbols in memory 'DMv_stack':
67
+ // Symbols in memory 'DMw':
68
+ // Symbols in memory 'DMw_bankA':
69
+ // Symbols in memory 'DMw_bankAB':
70
+ // Symbols in memory 'DMw_bankAC':
71
+ // Symbols in memory 'DMw_bankAD':
72
+ // Symbols in memory 'DMw_bankB':
73
+ // Symbols in memory 'DMw_bankBC':
74
+ // Symbols in memory 'DMw_bankBD':
75
+ // Symbols in memory 'DMw_bankC':
76
+ // Symbols in memory 'DMw_bankCD':
77
+ // Symbols in memory 'DMw_bankD':
78
+ // Symbols in memory 'DMw_stack':
79
+ // Symbols in memory 'DMx':
80
+ // Symbols in memory 'DMx_bankA':
81
+ // Symbols in memory 'DMx_bankAB':
82
+ // Symbols in memory 'DMx_bankAC':
83
+ // Symbols in memory 'DMx_bankAD':
84
+ // Symbols in memory 'DMx_bankB':
85
+ // Symbols in memory 'DMx_bankBC':
86
+ // Symbols in memory 'DMx_bankBD':
87
+ // Symbols in memory 'DMx_bankC':
88
+ // Symbols in memory 'DMx_bankCD':
89
+ // Symbols in memory 'DMx_bankD':
90
+ // Symbols in memory 'DMx_stack':
91
+ // Symbols in memory 'PM':
92
+ _symbol _Z13kernelWrapperPPvjjjj 0x00000930
93
+ _symbol _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh 0x00000b60
94
+ _symbol _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams 0x00001100
95
+ _symbol _Z11conv2d_bf16ILh1EL5act_t0E8bfloat16S1_S1_N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_NS3_IS5_NS6_5asyncES9_SB_EELb0ELb0ELb1ELb0EEvRNS2_9io_bufferIT1_NS2_9direction2inET4_EERNSF_IT2_SI_T5_EERNSF_IT3_NSH_3outET6_EER18conv2d_bf16_params 0x00001210
96
+ _symbol _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00001b80
97
+ _symbol _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv 0x00001fe0
98
+ _symbol _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E 0x00002020
99
+ _symbol _Z21superkernel_sigmoid1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x000021a0
100
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv 0x00002390
101
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E 0x000023d0
102
+ _symbol _Z18superkernel_tanh1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x000027a0
103
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv 0x00002990
104
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv 0x00002a10
105
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E 0x00002a80
106
+ _symbol _Z17superkernel_add1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00002a90
107
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E 0x00002cf0
108
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv 0x00002d10
109
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E 0x00002da0
110
+ _symbol _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00002ed0
111
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv 0x00003130
112
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv 0x000031b0
113
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168sub_implIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS4_E 0x00003200
114
+ _symbol _Z17superkernel_sub1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE 0x00003210
115
+ _symbol _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj 0x00003470
116
+ _symbol _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj 0x00003aa0
117
+ _symbol _Z13_b853_wrapperPPv 0x00003c80
118
+ _symbol _Z13_b896_wrapperPPv 0x00003ca0
119
+ _symbol _Z14_b1638_wrapperPPv 0x00003cd0
120
+ _symbol _Z14_b1655_wrapperPPv 0x00003cf0
121
+ _symbol _Z13_b891_wrapperPPv 0x00003d20
122
+ _symbol _Z14_b1672_wrapperPPv 0x00003d50
123
+ _symbol _Z13_b886_wrapperPPv 0x00003d70
124
+ // Symbols in memory 'PMw':
125
+ // Symbols in memory 'TM4':
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.srv ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/Release/0_0_reloadable24.txt ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/scripts/0_0_reloadable24.bcf ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _reserved DMb 0x0 0x40000
2
+
3
+ _reserved PM 0x0 0x930 //reserved for main elf
4
+
5
+ _entry_point _Z13kernelWrapperPPvjjjj
6
+ _symbol _Z13kernelWrapperPPvjjjj 0x930
7
+
8
+ _reserved DMb 0x7b280 0x800 //reserved for lcp ping-pong buffers
9
+ _reserved DMb 0x7ba80 0x40 //reserved for sync buffer
10
+ _stack DM_stack 0x7bac0 0x940 //stack for core
11
+ _reserved DMb 0x7c400 0x40 //reserved for main elf heap
12
+ //space for synopsys compiler at 0x7c440 0x880//heap
13
+ _reserved DMb 0x40000 0x3b280
14
+
15
+ _reserved DMb 0x7ccc0 0x3340
16
+ _reserved DMb 0x80000 0x80000 // And everything else the core can't see
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/scripts/0_0_reloadable24.prx ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <project name="Project" processor="me">
2
+ <file type="lbc" name="0_0_reloadable24.ll" path="../../ir"/>
3
+ <issinit/>
4
+ <option id="cpp.define" value="__AIENGINE__ __AIE_ARCH__=21" inherit = "1"/>
5
+ <option id="llvm.xargs" value="-fno-jump-tables -fno-discard-value-names" inherit = "1"/>
6
+ <option id="llvm.lang" value="Follow file extension"/>
7
+ <option id="bridge.cfg" value="0_0_reloadable24.bcf"/>
8
+ <option id="cpp.include" value="&lt;XILINX_VITIS_AIETOOLS&gt;/include /app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend /usr/local/lib/python3.10/site-packages/include/aie_api /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common /usr/local/lib/python3.10/dist-packages/vitis_mllib /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf /app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/&lt;AIE_SRC_DIR&gt; " inherit="1"/>
9
+ <option id="project.dir" value="&lt;CONFIG&gt;/0_0_reloadable24"/>
10
+ <option id="project.name" value="0_0_reloadable24"/>
11
+ <option id="project.type" value="exe"/>
12
+ <option id="simrun.issca" value="&lt;XILINX_VITIS_AIETOOLS&gt;/bin/unwrapped/lnx64.o/aie2p/ca_udm_dbg" inherit="0"/>
13
+ </project>
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/src/0_0_reloadable24.cc ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Automatically generated processor driver using AIEngine tool-chain
2
+
3
+ #include <stdio.h>
4
+ #include <memory>
5
+ #include <adf.h>
6
+
7
+
8
+ // Declare Kernel functions and initializers
9
+ #include "resize_adf_wrapper.cpp"
10
+ void conv2d_maxpool(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
11
+ void superkernel_sigmoid1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
12
+ void superkernel_sub1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
13
+ void superkernel_mul1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
14
+ void superkernel_tanh1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
15
+ void superkernel_add1d(adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>>> &__restrict,const unsigned int (&)[16],adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
16
+
17
+ // Declare Kernel objects and external arrays
18
+
19
+
20
+ void _b853_wrapper(void* args[])
21
+ {
22
+ mllib_graphs::resize_adf_wrapper<bfloat16, adf::io_buffer_config<adf::extents<>>, adf::io_buffer_config<adf::extents<>>, 1, 0, 2>(
23
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
24
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
25
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]));
26
+ }
27
+
28
+ void _b896_wrapper(void* args[])
29
+ {
30
+ conv2d_maxpool(
31
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
32
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
33
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
34
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
35
+ }
36
+
37
+ void _b1638_wrapper(void* args[])
38
+ {
39
+ superkernel_sigmoid1d(
40
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
41
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
42
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
43
+ }
44
+
45
+ void _b1655_wrapper(void* args[])
46
+ {
47
+ superkernel_sub1d(
48
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
49
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
50
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
51
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
52
+ }
53
+
54
+ void _b891_wrapper(void* args[])
55
+ {
56
+ superkernel_mul1d(
57
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
58
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
59
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
60
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
61
+ }
62
+
63
+ void _b1672_wrapper(void* args[])
64
+ {
65
+ superkernel_tanh1d(
66
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
67
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
68
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
69
+ }
70
+
71
+ void _b886_wrapper(void* args[])
72
+ {
73
+ superkernel_add1d(
74
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
75
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
76
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
77
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
78
+ }
79
+
80
+ using UniformKernelFunc = void (*)(void **);
81
+
82
+ static UniformKernelFunc g_uniformKernelFuncs[7] = {
83
+ _b853_wrapper,
84
+ _b896_wrapper,
85
+ _b1638_wrapper,
86
+ _b1655_wrapper,
87
+ _b891_wrapper,
88
+ _b1672_wrapper,
89
+ _b886_wrapper
90
+ };
91
+
92
+ __attribute__((always_inline)) void kernelWrapper(void* args[], uint32 kernelId, uint32 numSyncIn, uint32 numAsyncIn, uint32 numSyncOut)
93
+ {
94
+ uint32 idx = 0;
95
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 0);
96
+ idx += (numSyncIn > 0) ? 1 : 0;
97
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 1);
98
+ idx += (numSyncIn > 1) ? 1 : 0;
99
+ idx += numAsyncIn;
100
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::out>(numSyncOut > 0);
101
+ idx += (numSyncOut > 0) ? 1 : 0;
102
+
103
+ (*(g_uniformKernelFuncs[kernelId]))(args);
104
+
105
+ idx = 0;
106
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 0);
107
+ idx += (numSyncIn > 0) ? 1 : 0;
108
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 1);
109
+ idx += (numSyncIn > 1) ? 1 : 0;
110
+ idx += numAsyncIn;
111
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::out>(numSyncOut > 0);
112
+ idx += (numSyncOut > 0) ? 1 : 0;
113
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/timestamped_log/0_0_reloadable24.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable24/xlopt.log ADDED
@@ -0,0 +1,437 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
2
+
3
+
4
+ --------------- FILTER ANALYSIS INFO LOG ---------------
5
+
6
+ Reading Header IR from ir/_header.ll
7
+
8
+ --------------------------------------------------------
9
+
10
+ Add module pass 363{anonymous}::GuidancePass
11
+ Add module pass 363{anonymous}::ChessOptionsPass
12
+ Add module pass 363{anonymous}::DisableInliningInMainPass
13
+ Add module pass 363cdno::xlopt::AIEMergeSubWordStoresOpt
14
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::PromotePass>
15
+ Add module pass 363{anonymous}::IpConstPropPass
16
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::SimplifyCFGPass>
17
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::LoopSimplifyPass>
18
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::FunctionToLoopPassAdaptor>
19
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopInfoPass>
20
+ Add module pass 363cdno::xlopt::AIEAnnotatePragmaPass
21
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopPeelPass>
22
+ Add module pass 363{anonymous}::AIEAliasAnalysisPass
23
+
24
+
25
+ --------------- MEMORY MANAGEMENT GUIDANCE LOG ---------------
26
+
27
+ ALIGNMENT_HINT: Alignment of global array g_uniformKernelFuncs is 4 bytes; automatically aligning it to 64 bytes.
28
+ ALIGNMENT_HINT: Alignment of global array __const._Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj.coeffs is 2 bytes; automatically aligning it to 64 bytes.
29
+ ALIGNMENT_HINT: Alignment of global array __const._Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj.coeffs_asymmetric is 2 bytes; automatically aligning it to 64 bytes.
30
+ SIZE_HINT: Global array sigmoid_lut<0u, 256u>::data_ab is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
31
+ SIZE_HINT: Global array sigmoid_lut<0u, 256u>::data_cd is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
32
+ SIZE_HINT: Global array gelu_lut_32<0u, 512u>::data_ab is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
33
+ SIZE_HINT: Global array gelu_lut_32<0u, 512u>::data_cd is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
34
+ SIZE_HINT: Global array tanh_lut<0u, 512u>::data_ab is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
35
+ SIZE_HINT: Global array tanh_lut<0u, 512u>::data_cd is 2048 bytes. Consider making it mapper-managed LUT or memory buffer.
36
+ SIZE_HINT: Global array log_f32_lut<256u>::fraction_table_ab_f32 is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
37
+ SIZE_HINT: Global array log_f32_lut<256u>::fraction_table_cd_f32 is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
38
+ SIZE_HINT: Global array exp2_lut<512u>::exp2_table_ab is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
39
+ SIZE_HINT: Global array exp2_lut<512u>::exp2_table_cd is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
40
+
41
+ --------------------------------------------------------------
42
+
43
+
44
+
45
+ --------------- MERGING SUBWORD STORES OPT LOG ---------------
46
+
47
+
48
+
49
+ --------------- LOOP STATISTICS : _ZN18conv2d_bf16_paramsC2Ev ---------------
50
+
51
+ Total loops = 1
52
+ Loops with prepare for pipelining pragma = 0
53
+ Loops with unroll pragma = 0
54
+ Loops with flatten pragma = 0
55
+ Loops with min range pragma = 0
56
+ Loops with max range pragma = 0
57
+ Loops with known trip count = 0
58
+
59
+ -----------------------------------------------------------------------------
60
+
61
+
62
+
63
+ --------------- LOOP STATISTICS : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh ---------------
64
+
65
+ Total loops = 1
66
+ Loops with prepare for pipelining pragma = 0
67
+ Loops with unroll pragma = 0
68
+ Loops with flatten pragma = 0
69
+ Loops with min range pragma = 0
70
+ Loops with max range pragma = 0
71
+ Loops with known trip count = 1
72
+
73
+ -------------------------------------------------------------------------------------------------------------------
74
+
75
+
76
+
77
+ --------------- LOOP STATISTICS : _Z14conv2d_genericILh1EL5act_t0ELb0ELb1ELb0E8bfloat16EvPS1_S2_S2_S2_R18conv2d_bf16_params10out_mode_t ---------------
78
+
79
+ Total loops = 4
80
+ Loops with prepare for pipelining pragma = 4
81
+ Loops with unroll pragma = 0
82
+ Loops with flatten pragma = 0
83
+ Loops with min range pragma = 4
84
+ Loops with max range pragma = 0
85
+ Loops with known trip count = 0
86
+
87
+ -------------------------------------------------------------------------------------------------------------------------------------------------------
88
+
89
+
90
+
91
+ --------------- LOOP STATISTICS : _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams ---------------
92
+
93
+ Total loops = 1
94
+ Loops with prepare for pipelining pragma = 1
95
+ Loops with unroll pragma = 0
96
+ Loops with flatten pragma = 0
97
+ Loops with min range pragma = 1
98
+ Loops with max range pragma = 0
99
+ Loops with known trip count = 0
100
+
101
+ -------------------------------------------------------------------------------------------------------------------
102
+
103
+
104
+
105
+ --------------- LOOP STATISTICS : _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E ---------------
106
+
107
+ Total loops = 1
108
+ Loops with prepare for pipelining pragma = 1
109
+ Loops with unroll pragma = 0
110
+ Loops with flatten pragma = 0
111
+ Loops with min range pragma = 1
112
+ Loops with max range pragma = 0
113
+ Loops with known trip count = 0
114
+
115
+ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
116
+
117
+
118
+
119
+ --------------- LOOP STATISTICS : _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj32EEELj2EEC2Ev ---------------
120
+
121
+ Total loops = 1
122
+ Loops with prepare for pipelining pragma = 0
123
+ Loops with unroll pragma = 0
124
+ Loops with flatten pragma = 0
125
+ Loops with min range pragma = 0
126
+ Loops with max range pragma = 0
127
+ Loops with known trip count = 0
128
+
129
+ --------------------------------------------------------------------------------------------------------
130
+
131
+
132
+
133
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
134
+
135
+ Total loops = 1
136
+ Loops with prepare for pipelining pragma = 1
137
+ Loops with unroll pragma = 0
138
+ Loops with flatten pragma = 0
139
+ Loops with min range pragma = 1
140
+ Loops with max range pragma = 0
141
+ Loops with known trip count = 0
142
+
143
+ -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
144
+
145
+
146
+
147
+ --------------- LOOP STATISTICS : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE ---------------
148
+
149
+ Total loops = 1
150
+ Loops with prepare for pipelining pragma = 1
151
+ Loops with unroll pragma = 0
152
+ Loops with flatten pragma = 0
153
+ Loops with min range pragma = 1
154
+ Loops with max range pragma = 0
155
+ Loops with known trip count = 0
156
+
157
+ ---------------------------------------------------------------------------------------------------------------------------------------------------------------
158
+
159
+
160
+
161
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1619elementwise_sigmoidIS0_E26sigmoid_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
162
+
163
+ Total loops = 1
164
+ Loops with prepare for pipelining pragma = 1
165
+ Loops with unroll pragma = 0
166
+ Loops with flatten pragma = 0
167
+ Loops with min range pragma = 1
168
+ Loops with max range pragma = 0
169
+ Loops with known trip count = 0
170
+
171
+ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
172
+
173
+
174
+
175
+ --------------- LOOP STATISTICS : _ZN17elementwise_unaryI8bfloat1616elementwise_tanhIS0_E23tanh_templated_params_tIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E ---------------
176
+
177
+ Total loops = 1
178
+ Loops with prepare for pipelining pragma = 1
179
+ Loops with unroll pragma = 0
180
+ Loops with flatten pragma = 0
181
+ Loops with min range pragma = 1
182
+ Loops with max range pragma = 0
183
+ Loops with known trip count = 0
184
+
185
+ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
186
+
187
+
188
+
189
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj32EEC2Eb ---------------
190
+
191
+ Total loops = 2
192
+ Loops with prepare for pipelining pragma = 0
193
+ Loops with unroll pragma = 2
194
+ Loops with flatten pragma = 0
195
+ Loops with min range pragma = 0
196
+ Loops with max range pragma = 0
197
+ Loops with known trip count = 2
198
+
199
+ -------------------------------------------------------------------------
200
+
201
+
202
+
203
+ --------------- LOOP STATISTICS : _Z9avgpool2dILh1E8bfloat16Qsr5mllib5utilsE11is_one_of_vIT0_ahS0_EEvPS1_S2_R25avgpool2d_internal_paramsIS1_E ---------------
204
+
205
+ Total loops = 1
206
+ Loops with prepare for pipelining pragma = 1
207
+ Loops with unroll pragma = 0
208
+ Loops with flatten pragma = 0
209
+ Loops with min range pragma = 1
210
+ Loops with max range pragma = 0
211
+ Loops with known trip count = 0
212
+
213
+ -------------------------------------------------------------------------------------------------------------------------------------------------------------
214
+
215
+
216
+
217
+ --------------- LOOP STATISTICS : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E ---------------
218
+
219
+ Total loops = 1
220
+ Loops with prepare for pipelining pragma = 1
221
+ Loops with unroll pragma = 0
222
+ Loops with flatten pragma = 0
223
+ Loops with min range pragma = 1
224
+ Loops with max range pragma = 0
225
+ Loops with known trip count = 0
226
+
227
+ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
228
+
229
+
230
+
231
+ --------------- LOOP STATISTICS : _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh ---------------
232
+
233
+ Total loops = 1
234
+ Loops with prepare for pipelining pragma = 0
235
+ Loops with unroll pragma = 0
236
+ Loops with flatten pragma = 0
237
+ Loops with min range pragma = 0
238
+ Loops with max range pragma = 0
239
+ Loops with known trip count = 1
240
+
241
+ --------------------------------------------------------------------------------------------------------------
242
+
243
+
244
+
245
+ --------------- LOOP STATISTICS : _Z9conv2d_dwILh1E8bfloat16S0_S0_N3adf16io_buffer_configINS1_7extentsIJEEENS1_7locking4syncENS1_10addressing6linearENS1_6marginILj0EEEEESB_NS2_IS4_NS5_5asyncES8_SA_EEQsr3stdE9is_same_vIT0_S0_EEvRNS1_9io_bufferISE_NS1_9direction2inET3_EERNSF_IT1_SH_T4_EERNSF_IT2_NSG_3outET5_EER21conv2d_dw_bf16_params ---------------
246
+
247
+ Total loops = 2
248
+ Loops with prepare for pipelining pragma = 2
249
+ Loops with unroll pragma = 0
250
+ Loops with flatten pragma = 0
251
+ Loops with min range pragma = 2
252
+ Loops with max range pragma = 0
253
+ Loops with known trip count = 0
254
+
255
+ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
256
+
257
+
258
+
259
+ --------------- LOOP STATISTICS : _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE3runEPS0_S6_R18reduce_c8_params_tIS4_E ---------------
260
+
261
+ Total loops = 2
262
+ Loops with prepare for pipelining pragma = 1
263
+ Loops with unroll pragma = 0
264
+ Loops with flatten pragma = 0
265
+ Loops with min range pragma = 1
266
+ Loops with max range pragma = 0
267
+ Loops with known trip count = 0
268
+
269
+ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
270
+
271
+
272
+
273
+ --------------- LOOP STATISTICS : _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj16EEELj4EEC2Ev ---------------
274
+
275
+ Total loops = 1
276
+ Loops with prepare for pipelining pragma = 0
277
+ Loops with unroll pragma = 0
278
+ Loops with flatten pragma = 0
279
+ Loops with min range pragma = 0
280
+ Loops with max range pragma = 0
281
+ Loops with known trip count = 0
282
+
283
+ --------------------------------------------------------------------------------------------------------
284
+
285
+
286
+
287
+ --------------- LOOP STATISTICS : _ZN19reduce_mean_c8_implI8bfloat16E16body_postprocessEPvR18reduce_c8_params_tI23reduce_mean_c8_params_tIS0_EE ---------------
288
+
289
+ Total loops = 2
290
+ Loops with prepare for pipelining pragma = 2
291
+ Loops with unroll pragma = 0
292
+ Loops with flatten pragma = 0
293
+ Loops with min range pragma = 2
294
+ Loops with max range pragma = 0
295
+ Loops with known trip count = 0
296
+
297
+ ---------------------------------------------------------------------------------------------------------------------------------------------------------------
298
+
299
+
300
+
301
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj8EEC2Eb ---------------
302
+
303
+ Total loops = 2
304
+ Loops with prepare for pipelining pragma = 0
305
+ Loops with unroll pragma = 2
306
+ Loops with flatten pragma = 0
307
+ Loops with min range pragma = 0
308
+ Loops with max range pragma = 0
309
+ Loops with known trip count = 2
310
+
311
+ ------------------------------------------------------------------------
312
+
313
+
314
+
315
+ --------------- LOOP STATISTICS : _ZN3aie4maskILj16EEC2Eb ---------------
316
+
317
+ Total loops = 2
318
+ Loops with prepare for pipelining pragma = 0
319
+ Loops with unroll pragma = 2
320
+ Loops with flatten pragma = 0
321
+ Loops with min range pragma = 0
322
+ Loops with max range pragma = 0
323
+ Loops with known trip count = 2
324
+
325
+ -------------------------------------------------------------------------
326
+
327
+
328
+
329
+ --------------- LOOP STATISTICS : _ZNK3aie4maskILj8EEeqERKS1_ ---------------
330
+
331
+ Total loops = 1
332
+ Loops with prepare for pipelining pragma = 0
333
+ Loops with unroll pragma = 1
334
+ Loops with flatten pragma = 0
335
+ Loops with min range pragma = 0
336
+ Loops with max range pragma = 0
337
+ Loops with known trip count = 1
338
+
339
+ -----------------------------------------------------------------------------
340
+
341
+
342
+
343
+ --------------- LOOP STATISTICS : _ZNK3aie4maskILj16EEeqERKS1_ ---------------
344
+
345
+ Total loops = 1
346
+ Loops with prepare for pipelining pragma = 0
347
+ Loops with unroll pragma = 1
348
+ Loops with flatten pragma = 0
349
+ Loops with min range pragma = 0
350
+ Loops with max range pragma = 0
351
+ Loops with known trip count = 1
352
+
353
+ ------------------------------------------------------------------------------
354
+
355
+
356
+
357
+ --------------- LOOP STATISTICS : _Z26pad_3d_bfloat16_simplifiedP8bfloat16S0_R15pad_3d_params_t ---------------
358
+
359
+ Total loops = 3
360
+ Loops with prepare for pipelining pragma = 0
361
+ Loops with unroll pragma = 0
362
+ Loops with flatten pragma = 0
363
+ Loops with min range pragma = 0
364
+ Loops with max range pragma = 0
365
+ Loops with known trip count = 0
366
+
367
+ ---------------------------------------------------------------------------------------------------------------
368
+
369
+
370
+
371
+ --------------- LOOP STATISTICS : _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj ---------------
372
+
373
+ Total loops = 1
374
+ Loops with prepare for pipelining pragma = 0
375
+ Loops with unroll pragma = 0
376
+ Loops with flatten pragma = 0
377
+ Loops with min range pragma = 0
378
+ Loops with max range pragma = 0
379
+ Loops with known trip count = 1
380
+
381
+ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
382
+
383
+
384
+
385
+ --------------- LOOP STATISTICS : _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj ---------------
386
+
387
+ Total loops = 6
388
+ Loops with prepare for pipelining pragma = 2
389
+ Loops with unroll pragma = 0
390
+ Loops with flatten pragma = 0
391
+ Loops with min range pragma = 0
392
+ Loops with max range pragma = 0
393
+ Loops with known trip count = 0
394
+
395
+ ------------------------------------------------------------------------------------------------------------------------------------------
396
+
397
+
398
+
399
+ --------------- PRAGMA INSERTION LOG (unroll threshold = 3000, max unroll factor = 1) ---------------
400
+
401
+ Adding pragma to function _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh :
402
+ Pragma added to loop at line 437 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common/../../include/conv/conv2d_bf16_params.h : chess_loop_range(8,8)
403
+ Adding pragma to function _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj32EEELj2EEC2Ev :
404
+ Pragma added to loop at line 135 in file /usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite/../include/array : chess_prepare_for_pipelining
405
+ Adding pragma to function _ZN3aie4maskILj32EEC2Eb :
406
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
407
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
408
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
409
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
410
+ Adding pragma to function _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh :
411
+ Pragma added to loop at line 214 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common/../../include/conv/conv2d_dw_bf16_params.h : chess_loop_range(4,4)
412
+ Adding pragma to function _ZNSt3__25arrayIN3aie6vectorI8bfloat16Lj16EEELj4EEC2Ev :
413
+ Pragma added to loop at line 135 in file /usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite/../include/array : chess_prepare_for_pipelining
414
+ Adding pragma to function _ZN3aie4maskILj8EEC2Eb :
415
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
416
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
417
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
418
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
419
+ Adding pragma to function _ZN3aie4maskILj16EEC2Eb :
420
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
421
+ Pragma added to loop at line 122 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
422
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
423
+ Pragma added to loop at line 128 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
424
+ Adding pragma to function _ZNK3aie4maskILj16EEeqERKS1_ :
425
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
426
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
427
+ Adding pragma to function _ZNK3aie4maskILj8EEeqERKS1_ :
428
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_loop_range(1,1)
429
+ Pragma added to loop at line 249 in file /usr/local/lib/python3.10/dist-packages/include/aie_api/detail/aie2/../../mask.hpp : chess_flatten_loop
430
+ Adding pragma to function _Z23resize_bilinear_scale_2I8bfloat16Qsr5mllib5utilsE11is_one_of_vIT_S0_EEvPS1_PfS2_jjjj :
431
+ Pragma added to loop at line 283 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc/resize_bilinear.h : chess_prepare_for_pipelining
432
+ Pragma added to loop at line 262 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc/resize_bilinear.h : chess_prepare_for_pipelining
433
+ Adding pragma to function _ZN12mllib_graphs18resize_adf_wrapperI8bfloat16N3adf16io_buffer_configINS2_7extentsIJEEENS2_7locking4syncENS2_10addressing6linearENS2_6marginILj0EEEEESC_Li1ELi0ELi2EEEvRNS2_9io_bufferIT_NS2_9direction2inET0_EERNSD_ISE_NSF_3outET1_EERA16_Kj :
434
+ Pragma added to loop at line 81 in file /usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf/resize_adf_wrapper.cpp : chess_loop_range(6,6)
435
+
436
+ -----------------------------------------------------------------------------------------------------
437
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0_reloadable25/0_0_reloadable25.log ADDED
The diff for this file is too large to render. See raw diff