iswaalex commited on
Commit
1b04451
·
verified ·
1 Parent(s): f254366

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.onnx +3 -0
  2. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_128.bin +3 -0
  3. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_129.bin +3 -0
  4. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_130.bin +3 -0
  5. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_131.bin +3 -0
  6. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_136.bin +3 -0
  7. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_250.bin +3 -0
  8. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_37.bin +3 -0
  9. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_38.bin +3 -0
  10. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_4.bin +3 -0
  11. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_40.bin +3 -0
  12. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_41.bin +3 -0
  13. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_42.bin +3 -0
  14. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_43.bin +3 -0
  15. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_44.bin +3 -0
  16. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_45.bin +3 -0
  17. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_46.bin +3 -0
  18. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_56.bin +3 -0
  19. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_57.bin +3 -0
  20. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_58.bin +3 -0
  21. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_59.bin +3 -0
  22. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_6.bin +3 -0
  23. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_60.bin +3 -0
  24. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_61.bin +3 -0
  25. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_62.bin +3 -0
  26. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_63.bin +3 -0
  27. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_64.bin +3 -0
  28. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_65.bin +3 -0
  29. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_66.bin +3 -0
  30. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_67.bin +3 -0
  31. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_68.bin +3 -0
  32. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_69.bin +3 -0
  33. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_7.bin +3 -0
  34. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_70.bin +3 -0
  35. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_71.bin +3 -0
  36. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_72.bin +3 -0
  37. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_73.bin +3 -0
  38. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable4/src/0_2_reloadable4.cc +41 -0
  39. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.calltree +108 -0
  40. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.cmic2 +0 -0
  41. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.cmico +1 -0
  42. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.lst +0 -0
  43. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.map +324 -0
  44. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.sdr +129 -0
  45. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.srv +0 -0
  46. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.txt +0 -0
  47. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/scripts/0_2_reloadable5.bcf +16 -0
  48. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/scripts/0_2_reloadable5.prx +13 -0
  49. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/src/0_2_reloadable5.cc +110 -0
  50. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable6/Release/0_2_reloadable6.calltree +54 -0
segmentation_1_4_0_fp32_combined.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a73e7b69981004bd3df16a7d8494b40a757c422c999f310feee416e6e58b3d81
3
+ size 14975310
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_128.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:801429b6a86c59d0a139acc086a4c0b19dea4fcf010750b8af911795779b3ef7
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_129.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1545a2de7f8e377ec48bc84db8572b4a4733a86ba5f59ba34b08237b130101c4
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_130.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8eb16cc09bc51217dbd78f0fa4ff8c297513bb3ab6176fd7c20784771e643ea4
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_131.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8986a453226b3ab1a63a3f018f4e1368dc71e078b43de4a7b0812f2355fbf6e
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_136.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8986a453226b3ab1a63a3f018f4e1368dc71e078b43de4a7b0812f2355fbf6e
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_250.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9deb3e2e988c1fb7596f08c77c56782705cd603ff0f4db22b6f730ce92a3d9da
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_37.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a98dcc2951bc9192d34bfb6307858452851a99bccf04a0e6f0751da03ee9aed
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_38.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7e3f5788d2f47a2ff92b6afba037dfc1cb27150a5d686b3a890ab7396dc16171
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_4.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f901ed3e0570f6938244bfcab71f035049bb7eb40aca251115b78c1c6230eb7b
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_40.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:10fe86e880716f51f9b88c3bfd1f1a3a2a6bf5d80aa35a23fe02afc67a0bf8ce
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_41.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8be2ecd7a3db1c78360a3a8fdbe93c3f05a70ce09b8a87c58877275bceef18b9
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_42.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_43.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_44.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_45.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_46.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f160f47a8817f836489ba77f94fd815c17526fb3448a8ce9f2c7e6a14630ebe2
3
+ size 516
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_56.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b94f047a402bfc8e5e365a3a8d8d97d3513b9df28c496a07106a7e4c18f8ba3e
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_57.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be5bafa37310e0c9e5f314af01061c87eaa59c6b7c6b8940be41b557e434fa49
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_58.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb7e05ff34424a0c9fe31a13c78e1ed7edff1daf5f7a36ef57e9dce32b8bf80a
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_59.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac9a09a34c4d7ce81628e63bd5fa99c91004ce4c80d282cbd1e5a6931a1d97cc
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_6.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ef1a050a26e3498db0b30095a869f974c7c8b96c8c5824b4f6f1d81c25a597e
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_60.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e9022fcedc09886b890a1051e3db46a07cfe8e11f911a04af20c132431970bc
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_61.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d103c8f0b6bab9ffe3211caa0f368d276d23b6b251daba563d2537fac9a0dfb9
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_62.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b94f047a402bfc8e5e365a3a8d8d97d3513b9df28c496a07106a7e4c18f8ba3e
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_63.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f448e8669da6d20a453fb22041d00e74696063efae178dd0d9a366bcd5e11182
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_64.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac9a09a34c4d7ce81628e63bd5fa99c91004ce4c80d282cbd1e5a6931a1d97cc
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_65.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e9022fcedc09886b890a1051e3db46a07cfe8e11f911a04af20c132431970bc
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_66.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d103c8f0b6bab9ffe3211caa0f368d276d23b6b251daba563d2537fac9a0dfb9
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_67.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b94f047a402bfc8e5e365a3a8d8d97d3513b9df28c496a07106a7e4c18f8ba3e
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_68.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:002cf5d23c8929cc8c27573ba3129673938dd0c2fcd7a017f3bff7800e5134ef
3
+ size 516
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_69.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93e19fd229d0a4bcb70568ea1e79e2309157429f8aa5e82cf98d41f3462ba35f
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_7.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3f728557041508f674178e958e5df414570f923862e516e9c49f57290a915720
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_70.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d68d94744795723d9de33f9801a38c386d18ee297c760856c06a0937f9a4bc0
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_71.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c53dfc2035a1838cfbc626818b0648aa57434b654f90e9509b786589ae076c1
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_72.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:414db66a66f7ecb2350828a73e9496c75a63bd6657996d697cb372292a535d1b
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_73.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a5b30355951604c93a16f54d1605fa81e54ed33661924f5348d80270534c3c4
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable4/src/0_2_reloadable4.cc ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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_reduce_mean_c8(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
+
11
+ // Declare Kernel objects and external arrays
12
+
13
+
14
+ void _b961_wrapper(void* args[])
15
+ {
16
+ superkernel_reduce_mean_c8(
17
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
18
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
19
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
20
+ }
21
+
22
+ using UniformKernelFunc = void (*)(void **);
23
+
24
+ static UniformKernelFunc g_uniformKernelFuncs[1] = {
25
+ _b961_wrapper
26
+ };
27
+
28
+ __attribute__((always_inline)) void kernelWrapper(void* args[], uint32 kernelId, uint32 numSyncIn, uint32 numAsyncIn, uint32 numSyncOut)
29
+ {
30
+ uint32 idx = 0;
31
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 0);
32
+ idx += (numSyncIn > 0) ? 1 : 0;
33
+ idx += numAsyncIn;
34
+
35
+ (*(g_uniformKernelFuncs[kernelId]))(args);
36
+
37
+ idx = 0;
38
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 0);
39
+ idx += (numSyncIn > 0) ? 1 : 0;
40
+ idx += numAsyncIn;
41
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.calltree ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:49:20 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable5 ../Release/0_0_reloadable5.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_reloadable5.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/chesswork4008 -pme
5
+
6
+
7
+ // Release: ipp V-2024.06-TGT-241219
8
+
9
+ _Z13kernelWrapperPPvjjjj
10
+ _Z13_b896_wrapperPPv (referenced text)
11
+ _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
12
+ _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
13
+ _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
14
+ _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
15
+ _Z13_b901_wrapperPPv (referenced text)
16
+ _Z40superkernel_add1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
17
+ _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_EPKv
18
+ _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_EPKv
19
+ _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_E
20
+ _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_E
21
+ _ZN41elementwise_binary_attribute_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_R27elementwise_binary_params_tIS5_E
22
+ _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
23
+ _Z13_b906_wrapperPPv (referenced text)
24
+ _Z18superkernel_clip1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
25
+ _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
26
+ _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
27
+ _Z13_b881_wrapperPPv (referenced text)
28
+ _Z40superkernel_mul1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
29
+ _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
30
+ _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
31
+ _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_R27elementwise_binary_params_tIS4_E
32
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
33
+ _Z13_b891_wrapperPPv (referenced text)
34
+ _Z17superkernel_mul1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERNS0_IS1_NS2_3outESK_EE
35
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
36
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
37
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
38
+ _Z13_b924_wrapperPPv (referenced text)
39
+ _Z26superkernel_conv_eltbinaryRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERA17_KjRNS0_IS1_NS2_3outESH_EE
40
+ _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh (*)
41
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv (*)
42
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
43
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
44
+ _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 (*)
45
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E (*)
46
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
47
+ _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (*)
48
+ _Z13_b919_wrapperPPv (referenced text)
49
+ _Z22superkernel_conv2d_dwcRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
50
+ _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh
51
+ _ZN12me_primitive10udiv_dstepEjjRjS0_
52
+ _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
53
+
54
+
55
+ Call tree stack and functions sizes:
56
+
57
+ stack stack stack call func func function name
58
+ desc level level desc
59
+ ----- ----- ----- ----- ----- ----- --------------------------------------------------------------
60
+ 64 320 0 0 390 13150 _Z13kernelWrapperPPvjjjj
61
+ 0 192 1 1 36 4714 _Z13_b896_wrapperPPv
62
+ 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>
63
+ 64 64 2 3 1430 1430 _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
64
+ 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
65
+ 0 0 3 4 270 270 _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
66
+ 0 192 1 1 32 1252 _Z13_b901_wrapperPPv
67
+ 64 192 1 2 488 1220 _Z40superkernel_add1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
68
+ 64 128 2 3 62 304 _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_EPKv
69
+ 64 64 3 4 162 186 _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_EPKv
70
+ 0 0 4 5 24 24 _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_E
71
+ 0 0 2 4 56 56 _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_E <tail call>
72
+ 128 128 2 3 114 428 _ZN41elementwise_binary_attribute_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_R27elementwise_binary_params_tIS5_E
73
+ 0 0 3 4 314 314 _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
74
+ 0 64 1 1 32 862 _Z13_b906_wrapperPPv
75
+ 64 64 1 2 488 830 _Z18superkernel_clip1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
76
+ 0 0 2 3 100 100 _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
77
+ 0 0 2 3 242 242 _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
78
+ 0 256 1 1 32 1394 _Z13_b881_wrapperPPv
79
+ 64 256 1 2 488 1362 _Z40superkernel_mul1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
80
+ 64 64 2 3 74 190 _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
81
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
82
+ 64 192 2 3 150 684 _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_R27elementwise_binary_params_tIS4_E
83
+ 128 128 3 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
84
+ 0 128 1 1 36 1092 _Z13_b891_wrapperPPv
85
+ 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>
86
+ 64 64 2 3 138 162 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
87
+ 0 0 3 4 24 24 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
88
+ 0 0 2 3 292 292 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
89
+ 0 192 1 1 40 6494 _Z13_b924_wrapperPPv
90
+ 64 192 1 2 1126 6454 _Z26superkernel_conv_eltbinaryRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERA17_KjRNS0_IS1_NS2_3outESH_EE <tail call>
91
+ 64 64 2 3 1430 1430 _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
92
+ 64 64 2 3 138 162 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv (*)
93
+ 64 64 2 3 98 214 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
94
+ 0 0 3 4 116 116 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
95
+ 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 (*)
96
+ 0 0 2 3 292 292 _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
97
+ 0 128 2 3 16 550 _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
98
+ 128 128 2 4 534 534 _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE <tail call>
99
+ 0 192 1 1 36 2050 _Z13_b919_wrapperPPv
100
+ 128 192 1 2 478 2014 _Z22superkernel_conv2d_dwcRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE <tail call>
101
+ 64 64 2 3 672 814 _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh
102
+ 0 0 3 4 142 142 _ZN12me_primitive10udiv_dstepEjjRjS0_
103
+ 0 0 2 3 722 722 _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
104
+
105
+
106
+ Maximum call level : 5
107
+ Maximum stack level: 4
108
+ Maximum stack size : 320
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.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_2_reloadable5/Release/0_2_reloadable5.cmico ADDED
@@ -0,0 +1 @@
 
 
1
+ +Mdec
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.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_2_reloadable5/Release/0_2_reloadable5.map ADDED
@@ -0,0 +1,324 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:49:21 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable5 ../Release/0_0_reloadable5.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_reloadable5.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/chesswork4008 -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 = 3461
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_reloadable5.o::_ZL9curr_iter (Data, Local, .bss.DMb.4)
28
+ 0x0007c444..0x0007c447 ( 4 items) : ../Release/0_0_reloadable5.o::_ZL14num_depth_iter (Data, Local, .bss.DMb.4)
29
+ 0x0007c448..0x0007c44b ( 4 items) : ../Release/0_0_reloadable5.o::_ZL10depth_iter (Data, Local, .bss.DMb.4)
30
+ 0x0007c44c..0x0007c44f ( 4 items) : ../Release/0_0_reloadable5.o::_ZL11total_iters (Data, Local, .bss.DMb.4)
31
+ 0x0007c450..0x0007c453 ( 4 items) : ../Release/0_0_reloadable5.o::_ZL8core_row (Data, Local, .bss.DMb.4)
32
+ 0x0007c454..0x0007c457 ( 4 items) : ../Release/0_0_reloadable5.o::_ZL11ifm1_offset (Data, Local, .bss.DMb.4)
33
+ 0x0007c458..0x0007c45b ( 4 items) : ../Release/0_0_reloadable5.o::_ZL10ifmsv_size (Data, Local, .bss.DMb.4)
34
+ 0x0007c45c..0x0007c45f ( 4 items) : ../Release/0_0_reloadable5.o::_ZL11ifm2_offset (Data, Local, .bss.DMb.4)
35
+ 0x0007c460..0x0007c463 ( 4 items) : ../Release/0_0_reloadable5.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_reloadable5.o::_ZL20g_uniformKernelFuncs (Data, Local, .rodata.DMb.64)
39
+
40
+ Called functions : _Z13_b896_wrapperPPv
41
+ _Z13_b901_wrapperPPv
42
+ _Z13_b906_wrapperPPv
43
+ _Z13_b881_wrapperPPv
44
+ _Z13_b891_wrapperPPv
45
+ _Z13_b924_wrapperPPv
46
+ _Z13_b919_wrapperPPv
47
+
48
+ 0x0007c4c0..0x0007c4ff ( 64 items) : ../Release/0_0_reloadable5.o::add1d_attribute_broadcasting_params (Data, Global, .bss.DMb.64)
49
+ 0x0007c500..0x0007c53f ( 64 items) : ../Release/0_0_reloadable5.o::mul1d_attribute_broadcasting_params (Data, Global, .bss.DMb.64)
50
+ 0x0007c540..0x0007c57f ( 64 items) : ../Release/0_0_reloadable5.o::add1d_params (Data, Global, .bss.DMb.64)
51
+ 0x0007c580..0x0007c5bf ( 64 items) : ../Release/0_0_reloadable5.o::mul1d_params (Data, Global, .bss.DMb.64)
52
+ 0x0007c5c0..0x0007c5ff ( 64 items) : ../Release/0_0_reloadable5.o::clip1d_params (Data, Global, .bss.DMb.64)
53
+ 0x0007c600..0x0007c7bf ( 448 items) : ../Release/0_0_reloadable5.o::conv2d_params (Data, Global, .bss.DMb.64)
54
+ 0x0007c7c0..0x0007c8bf ( 256 items) : ../Release/0_0_reloadable5.o::conv2d_dw_params (Data, Global, .bss.DMb.64)
55
+ 0x0007ccc0..0x000fffff ( 537408 items) : Reserved
56
+
57
+ Memory map for memory 'PM':
58
+
59
+ Size = 1048576
60
+ Width = 8 bits
61
+ Offset = 0
62
+ Used = 13150
63
+
64
+ 0x00000000..0x0000092f ( 2352 items) : Reserved
65
+ 0x00000930..0x00000ab5 ( 390 items) : ../Release/0_0_reloadable5.o::_Z13kernelWrapperPPvjjjj (Function, Global, .text) (stack frame size = 64)
66
+
67
+ Referenced symbols: _ZL20g_uniformKernelFuncs
68
+
69
+ 0x00000ac0..0x00001055 ( 1430 items) : ../Release/0_0_reloadable5.o::_Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh (Function, Weak, .text) (stack frame size = 64)
70
+ 0x00001060..0x0000116d ( 270 items) : ../Release/0_0_reloadable5.o::_Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams (Function, Weak, .text) (stack frame size = 0)
71
+
72
+ Referenced symbols: _ZN12me_primitive11control_rndE
73
+
74
+ 0x00001170..0x00001ad9 ( 2410 items) : ../Release/0_0_reloadable5.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)
75
+
76
+ Called functions : _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams
77
+
78
+ Referenced symbols: _ZN12me_primitive11control_rndE
79
+
80
+ 0x00001ae0..0x00001d17 ( 568 items) : ../Release/0_0_reloadable5.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)
81
+
82
+ Called functions : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
83
+ _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
84
+
85
+ Referenced symbols: _ZL9curr_iter
86
+ conv2d_params
87
+ _ZN12me_primitive11control_rndE
88
+ _ZN12me_primitive11control_satE
89
+ _ZL14num_depth_iter
90
+ _ZL8num_iter
91
+ _ZL10depth_iter
92
+ _ZL11total_iters
93
+
94
+ 0x00001d20..0x00001d37 ( 24 items) : ../Release/0_0_reloadable5.o::_ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 0)
95
+ 0x00001d40..0x00001de1 ( 162 items) : ../Release/0_0_reloadable5.o::_ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 64)
96
+
97
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_E
98
+
99
+ 0x00001df0..0x00001e27 ( 56 items) : ../Release/0_0_reloadable5.o::_ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 0)
100
+ 0x00001e30..0x00001e6d ( 62 items) : ../Release/0_0_reloadable5.o::_ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 64)
101
+
102
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_EPKv
103
+ _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_E
104
+
105
+ 0x00001e70..0x00001fa9 ( 314 items) : ../Release/0_0_reloadable5.o::_ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 0)
106
+
107
+ Referenced symbols: _ZN12me_primitive11control_rndE
108
+
109
+ 0x00001fb0..0x00002021 ( 114 items) : ../Release/0_0_reloadable5.o::_ZN41elementwise_binary_attribute_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_R27elementwise_binary_params_tIS5_E (Function, Weak, .text) (stack frame size = 128)
110
+
111
+ Called functions : _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
112
+
113
+ 0x00002030..0x00002217 ( 488 items) : ../Release/0_0_reloadable5.o::_Z40superkernel_add1d_attribute_broadcastingRN3adf9io_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)
114
+
115
+ Called functions : _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_EPKv
116
+ _ZN41elementwise_binary_attribute_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_R27elementwise_binary_params_tIS5_E
117
+
118
+ Referenced symbols: _ZL9curr_iter
119
+ _ZL8core_row
120
+ add1d_attribute_broadcasting_params
121
+ _ZN12me_primitive11control_rndE
122
+ _ZN12me_primitive11control_satE
123
+ _ZL11ifm1_offset
124
+ _ZL8num_iter
125
+
126
+ 0x00002220..0x00002283 ( 100 items) : ../Release/0_0_reloadable5.o::_ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
127
+ 0x00002290..0x00002381 ( 242 items) : ../Release/0_0_reloadable5.o::_ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
128
+ 0x00002390..0x00002577 ( 488 items) : ../Release/0_0_reloadable5.o::_Z18superkernel_clip1dRN3adf9io_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)
129
+
130
+ Called functions : _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv
131
+ _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E
132
+
133
+ Referenced symbols: _ZL9curr_iter
134
+ _ZL8core_row
135
+ clip1d_params
136
+ _ZN12me_primitive11control_rndE
137
+ _ZN12me_primitive11control_satE
138
+ _ZL11ifm1_offset
139
+ _ZL8num_iter
140
+
141
+ 0x00002580..0x000025f3 ( 116 items) : ../Release/0_0_reloadable5.o::_ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 0)
142
+ 0x00002600..0x00002649 ( 74 items) : ../Release/0_0_reloadable5.o::_ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
143
+
144
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv
145
+
146
+ 0x00002650..0x00002865 ( 534 items) : ../Release/0_0_reloadable5.o::_ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE (Function, Local, .text) (stack frame size = 128)
147
+
148
+ Referenced symbols: _ZN12me_primitive11control_rndE
149
+
150
+ 0x00002870..0x00002905 ( 150 items) : ../Release/0_0_reloadable5.o::_ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 64)
151
+
152
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
153
+
154
+ 0x00002910..0x00002af7 ( 488 items) : ../Release/0_0_reloadable5.o::_Z40superkernel_mul1d_attribute_broadcastingRN3adf9io_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)
155
+
156
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv
157
+ _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_R27elementwise_binary_params_tIS4_E
158
+
159
+ Referenced symbols: _ZL9curr_iter
160
+ _ZL8core_row
161
+ mul1d_attribute_broadcasting_params
162
+ _ZN12me_primitive11control_rndE
163
+ _ZN12me_primitive11control_satE
164
+ _ZL11ifm1_offset
165
+ _ZL8num_iter
166
+
167
+ 0x00002b00..0x00002b73 ( 116 items) : ../Release/0_0_reloadable5.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 0)
168
+ 0x00002b80..0x00002be1 ( 98 items) : ../Release/0_0_reloadable5.o::_ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv (Function, Weak, .text) (stack frame size = 64)
169
+
170
+ Called functions : _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv
171
+
172
+ 0x00002bf0..0x00002bff ( 16 items) : ../Release/0_0_reloadable5.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)
173
+
174
+ Called functions : _ZL19shared_run_backboneI8bfloat16L5act_t0EEKvPT_S4_S4_R27elementwise_binary_params_tI15shared_params_tIS3_EE
175
+
176
+ 0x00002c00..0x00002c17 ( 24 items) : ../Release/0_0_reloadable5.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
177
+ 0x00002c20..0x00002ca9 ( 138 items) : ../Release/0_0_reloadable5.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv (Function, Weak, .text) (stack frame size = 64)
178
+
179
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E
180
+
181
+ 0x00002cb0..0x00002dd3 ( 292 items) : ../Release/0_0_reloadable5.o::_ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E (Function, Weak, .text) (stack frame size = 0)
182
+
183
+ Referenced symbols: _ZN12me_primitive11control_rndE
184
+
185
+ 0x00002de0..0x00003039 ( 602 items) : ../Release/0_0_reloadable5.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)
186
+
187
+ Called functions : _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
188
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
189
+
190
+ Referenced symbols: _ZL9curr_iter
191
+ _ZL8core_row
192
+ mul1d_params
193
+ _ZN12me_primitive11control_rndE
194
+ _ZN12me_primitive11control_satE
195
+ _ZL11ifm1_offset
196
+ _ZL11ifm2_offset
197
+ _ZL8num_iter
198
+
199
+ 0x00003040..0x000032df ( 672 items) : ../Release/0_0_reloadable5.o::_ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh (Function, Local, .text) (stack frame size = 64)
200
+
201
+ Called functions : _ZN12me_primitive10udiv_dstepEjjRjS0_
202
+
203
+ Referenced symbols: conv2d_dw_params
204
+ _ZN12me_primitive11control_rndE
205
+
206
+ 0x000032e0..0x000035b1 ( 722 items) : ../Release/0_0_reloadable5.o::_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 (Function, Weak, .text) (stack frame size = 0)
207
+
208
+ Referenced symbols: _ZN12me_primitive11control_rndE
209
+
210
+ 0x000035c0..0x0000379d ( 478 items) : ../Release/0_0_reloadable5.o::_Z22superkernel_conv2d_dwcRN3adf9io_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 = 128)
211
+
212
+ Called functions : _ZL27setup_conv2d_dw_params_bf16PKjR21conv2d_dw_bf16_paramsh
213
+ _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
214
+
215
+ Referenced symbols: _ZL9curr_iter
216
+ _ZL8core_row
217
+ _ZN12me_primitive11control_rndE
218
+ _ZN12me_primitive11control_satE
219
+ _ZL8num_iter
220
+ _ZL10ifmsv_size
221
+ conv2d_dw_params
222
+
223
+ 0x000037a0..0x00003c05 ( 1126 items) : ../Release/0_0_reloadable5.o::_Z26superkernel_conv_eltbinaryRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERA17_KjRNS0_IS1_NS2_3outESH_EE (Function, Global, .text) (stack frame size = 64)
224
+
225
+ Called functions : _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh
226
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv
227
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv
228
+ _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
229
+ _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E
230
+ _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E
231
+
232
+ Referenced symbols: _ZL9curr_iter
233
+ _ZL8core_row
234
+ _ZN12me_primitive11control_rndE
235
+ _ZN12me_primitive11control_satE
236
+ conv2d_params
237
+ add1d_params
238
+ mul1d_params
239
+ _ZL14num_depth_iter
240
+ _ZL11ifm2_offset
241
+ _ZL8num_iter
242
+ _ZL10depth_iter
243
+ _ZL11total_iters
244
+
245
+ 0x00003c10..0x00003c33 ( 36 items) : ../Release/0_0_reloadable5.o::_Z13_b896_wrapperPPv (Function, Global, .text) (stack frame size = 0)
246
+
247
+ Called functions : _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
248
+
249
+ 0x00003c40..0x00003c5f ( 32 items) : ../Release/0_0_reloadable5.o::_Z13_b901_wrapperPPv (Function, Global, .text) (stack frame size = 0)
250
+
251
+ Called functions : _Z40superkernel_add1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
252
+
253
+ 0x00003c60..0x00003c7f ( 32 items) : ../Release/0_0_reloadable5.o::_Z13_b906_wrapperPPv (Function, Global, .text) (stack frame size = 0)
254
+
255
+ Called functions : _Z18superkernel_clip1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
256
+
257
+ 0x00003c80..0x00003c9f ( 32 items) : ../Release/0_0_reloadable5.o::_Z13_b881_wrapperPPv (Function, Global, .text) (stack frame size = 0)
258
+
259
+ Called functions : _Z40superkernel_mul1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
260
+
261
+ 0x00003ca0..0x00003cc3 ( 36 items) : ../Release/0_0_reloadable5.o::_Z13_b891_wrapperPPv (Function, Global, .text) (stack frame size = 0)
262
+
263
+ 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
264
+
265
+ 0x00003cd0..0x00003cf7 ( 40 items) : ../Release/0_0_reloadable5.o::_Z13_b924_wrapperPPv (Function, Global, .text) (stack frame size = 0)
266
+
267
+ Called functions : _Z26superkernel_conv_eltbinaryRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERA17_KjRNS0_IS1_NS2_3outESH_EE
268
+
269
+ 0x00003d00..0x00003d23 ( 36 items) : ../Release/0_0_reloadable5.o::_Z13_b919_wrapperPPv (Function, Global, .text) (stack frame size = 0)
270
+
271
+ Called functions : _Z22superkernel_conv2d_dwcRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
272
+
273
+ 0x00003d30..0x00003dbd ( 142 items) : me_div.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_ZN12me_primitive10udiv_dstepEjjRjS0_ (Function, Global, .text) (stack frame size = 0)
274
+
275
+ External symbols:
276
+
277
+ __dso_handle = 0x0
278
+ _ctors_end = 0x0
279
+ _ctors_start = 0x0
280
+ _dtors_end = 0x0
281
+ _dtors_start = 0x0
282
+ _pc_end = 0x3dbe
283
+ _pc_start = 0x930
284
+ _sp_end_DM_stack = 0x7c400
285
+ _sp_start_DM_stack = 0x7bac0
286
+
287
+ Section summary for memory 'DM_stack':
288
+
289
+ .stack File
290
+ ---------- ----------
291
+ 2368
292
+ ---------- ----------
293
+ 2368 Total
294
+
295
+ Section summary for memory 'DMb':
296
+
297
+ .bss .data .rodata File
298
+ ---------- ---------- ---------- ----------
299
+ 1056 4 28 ../Release/0_0_reloadable5.o
300
+ 5 0 0 me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
301
+ ---------- ---------- ---------- ----------
302
+ 1061 4 28 Total
303
+
304
+ Section summary for memory 'PM':
305
+
306
+ .text File
307
+ ---------- ----------
308
+ 13008 ../Release/0_0_reloadable5.o
309
+ 142 me_div.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
310
+ ---------- ----------
311
+ 13150 Total
312
+
313
+ File summary:
314
+
315
+ ../Release/0_0_reloadable5.o
316
+ DMb 1088
317
+ PM 13008
318
+
319
+ me_defs.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
320
+ DMb 5
321
+
322
+ me_div.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
323
+ PM 142
324
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.sdr ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:49:21 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable5 ../Release/0_0_reloadable5.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_reloadable5.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/chesswork4008 -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 0x0007c464
22
+ _symbol _ZN12me_primitive11control_rndE 0x0007c468
23
+ _symbol add1d_attribute_broadcasting_params 0x0007c4c0
24
+ _symbol mul1d_attribute_broadcasting_params 0x0007c500
25
+ _symbol add1d_params 0x0007c540
26
+ _symbol mul1d_params 0x0007c580
27
+ _symbol clip1d_params 0x0007c5c0
28
+ _symbol conv2d_params 0x0007c600
29
+ _symbol conv2d_dw_params 0x0007c7c0
30
+ // Symbols in memory 'DMh':
31
+ // Symbols in memory 'DMh_bankA':
32
+ // Symbols in memory 'DMh_bankAB':
33
+ // Symbols in memory 'DMh_bankAC':
34
+ // Symbols in memory 'DMh_bankAD':
35
+ // Symbols in memory 'DMh_bankB':
36
+ // Symbols in memory 'DMh_bankBC':
37
+ // Symbols in memory 'DMh_bankBD':
38
+ // Symbols in memory 'DMh_bankC':
39
+ // Symbols in memory 'DMh_bankCD':
40
+ // Symbols in memory 'DMh_bankD':
41
+ // Symbols in memory 'DMh_stack':
42
+ // Symbols in memory 'DMs':
43
+ // Symbols in memory 'DMs_bankA':
44
+ // Symbols in memory 'DMs_bankAB':
45
+ // Symbols in memory 'DMs_bankAC':
46
+ // Symbols in memory 'DMs_bankAD':
47
+ // Symbols in memory 'DMs_bankB':
48
+ // Symbols in memory 'DMs_bankBC':
49
+ // Symbols in memory 'DMs_bankBD':
50
+ // Symbols in memory 'DMs_bankC':
51
+ // Symbols in memory 'DMs_bankCD':
52
+ // Symbols in memory 'DMs_bankD':
53
+ // Symbols in memory 'DMs_stack':
54
+ // Symbols in memory 'DMv':
55
+ // Symbols in memory 'DMv_bankA':
56
+ // Symbols in memory 'DMv_bankAB':
57
+ // Symbols in memory 'DMv_bankAC':
58
+ // Symbols in memory 'DMv_bankAD':
59
+ // Symbols in memory 'DMv_bankB':
60
+ // Symbols in memory 'DMv_bankBC':
61
+ // Symbols in memory 'DMv_bankBD':
62
+ // Symbols in memory 'DMv_bankC':
63
+ // Symbols in memory 'DMv_bankCD':
64
+ // Symbols in memory 'DMv_bankD':
65
+ // Symbols in memory 'DMv_stack':
66
+ // Symbols in memory 'DMw':
67
+ // Symbols in memory 'DMw_bankA':
68
+ // Symbols in memory 'DMw_bankAB':
69
+ // Symbols in memory 'DMw_bankAC':
70
+ // Symbols in memory 'DMw_bankAD':
71
+ // Symbols in memory 'DMw_bankB':
72
+ // Symbols in memory 'DMw_bankBC':
73
+ // Symbols in memory 'DMw_bankBD':
74
+ // Symbols in memory 'DMw_bankC':
75
+ // Symbols in memory 'DMw_bankCD':
76
+ // Symbols in memory 'DMw_bankD':
77
+ // Symbols in memory 'DMw_stack':
78
+ // Symbols in memory 'DMx':
79
+ // Symbols in memory 'DMx_bankA':
80
+ // Symbols in memory 'DMx_bankAB':
81
+ // Symbols in memory 'DMx_bankAC':
82
+ // Symbols in memory 'DMx_bankAD':
83
+ // Symbols in memory 'DMx_bankB':
84
+ // Symbols in memory 'DMx_bankBC':
85
+ // Symbols in memory 'DMx_bankBD':
86
+ // Symbols in memory 'DMx_bankC':
87
+ // Symbols in memory 'DMx_bankCD':
88
+ // Symbols in memory 'DMx_bankD':
89
+ // Symbols in memory 'DMx_stack':
90
+ // Symbols in memory 'PM':
91
+ _symbol _Z13kernelWrapperPPvjjjj 0x00000930
92
+ _symbol _Z24setup_conv2d_bf16_paramsILb1ELb0EEvPKjR18conv2d_bf16_paramshh 0x00000ac0
93
+ _symbol _Z21convert_bf16_to_bfp16I8bfloat16Lb0EEvPT_PS0_RK13BfToBfpParams 0x00001060
94
+ _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 0x00001170
95
+ _symbol _Z14conv2d_maxpoolRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00001ae0
96
+ _symbol _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_E 0x00001d20
97
+ _symbol _ZN18elementwise_binaryIJ8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS5_EPKv 0x00001d40
98
+ _symbol _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_E 0x00001df0
99
+ _symbol _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE5setupER27elementwise_binary_params_tIS5_EPKv 0x00001e30
100
+ _symbol _ZN31elementwise_binary_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E 0x00001e70
101
+ _symbol _ZN41elementwise_binary_attribute_broadcastingI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_EE3runEPS0_S7_R27elementwise_binary_params_tIS5_E 0x00001fb0
102
+ _symbol _Z40superkernel_add1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00002030
103
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE5setupER26elementwise_unary_params_tIS4_EPKv 0x00002220
104
+ _symbol _ZN17elementwise_unaryI8bfloat1616elementwise_clipIS0_E20clip_internal_paramsIS0_EE3runEPS0_S6_R26elementwise_unary_params_tIS4_E 0x00002290
105
+ _symbol _Z18superkernel_clip1dRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00002390
106
+ _symbol _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE21shared_setup_backboneER27elementwise_binary_params_tIS4_EPKv 0x00002580
107
+ _symbol _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE5setupER27elementwise_binary_params_tIS4_EPKv 0x00002600
108
+ _symbol _ZN25elementwise_binary_sharedI8bfloat1626mul_impl_broadcasting_attrIS0_E15shared_params_tIS0_EL5act_t0EE3runEPS0_S7_R27elementwise_binary_params_tIS4_E 0x00002870
109
+ _symbol _Z40superkernel_mul1d_attribute_broadcastingRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x00002910
110
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE21shared_setup_backboneER27elementwise_binary_params_tIS5_EPKv 0x00002b00
111
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE5setupER27elementwise_binary_params_tIS5_EPKv 0x00002b80
112
+ _symbol _ZN25elementwise_binary_sharedI8bfloat168add_implIS0_L5act_t0EE15shared_params_tIS0_ELS2_0EE3runEPS0_S7_S7_R27elementwise_binary_params_tIS5_E 0x00002bf0
113
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_E 0x00002c00
114
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE5setupER27elementwise_binary_params_tIS4_EPKv 0x00002c20
115
+ _symbol _ZN18elementwise_binaryIJ8bfloat168mul_implIS0_E15shared_params_tIS0_EEE3runEPS0_S6_S6_R27elementwise_binary_params_tIS4_E 0x00002cb0
116
+ _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 0x00002de0
117
+ _symbol _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 0x000032e0
118
+ _symbol _Z22superkernel_conv2d_dwcRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE 0x000035c0
119
+ _symbol _Z26superkernel_conv_eltbinaryRN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEESF_RNS0_IS1_S3_NS4_IS6_NS7_5asyncESA_SC_EEEERA17_KjRNS0_IS1_NS2_3outESH_EE 0x000037a0
120
+ _symbol _Z13_b896_wrapperPPv 0x00003c10
121
+ _symbol _Z13_b901_wrapperPPv 0x00003c40
122
+ _symbol _Z13_b906_wrapperPPv 0x00003c60
123
+ _symbol _Z13_b881_wrapperPPv 0x00003c80
124
+ _symbol _Z13_b891_wrapperPPv 0x00003ca0
125
+ _symbol _Z13_b924_wrapperPPv 0x00003cd0
126
+ _symbol _Z13_b919_wrapperPPv 0x00003d00
127
+ _symbol _ZN12me_primitive10udiv_dstepEjjRjS0_ 0x00003d30
128
+ // Symbols in memory 'PMw':
129
+ // Symbols in memory 'TM4':
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable5/Release/0_2_reloadable5.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_2_reloadable5/Release/0_2_reloadable5.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_2_reloadable5/scripts/0_2_reloadable5.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_2_reloadable5/scripts/0_2_reloadable5.prx ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <project name="Project" processor="me">
2
+ <issinit/>
3
+ <option id="cpp.define" value="__AIENGINE__ __AIE_ARCH__=21" inherit = "1"/>
4
+ <option id="llvm.xargs" value="-fno-jump-tables -fno-discard-value-names" inherit = "1"/>
5
+ <option id="llvm.lang" value="Follow file extension"/>
6
+ <option id="bridge.cfg" value="0_2_reloadable5.bcf"/>
7
+ <option id="bridge.alibs" value=" ../Release/0_2_reloadable5.o" inherit="1" cfg="Release_LLVM"/>
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_2_reloadable5"/>
10
+ <option id="project.name" value="0_2_reloadable5"/>
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_2_reloadable5/src/0_2_reloadable5.cc ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 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);
10
+ void superkernel_add1d_attribute_broadcasting(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_clip1d(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_mul1d_attribute_broadcasting(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);
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_conv_eltbinary(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,adf::io_buffer<bfloat16, adf::direction::in, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict,const unsigned int (&)[17],adf::io_buffer<bfloat16, adf::direction::out, adf::io_buffer_config<adf::extents<>, adf::locking::async>> &__restrict);
15
+ void superkernel_conv2d_dwc(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);
16
+
17
+ // Declare Kernel objects and external arrays
18
+
19
+
20
+ void _b896_wrapper(void* args[])
21
+ {
22
+ conv2d_maxpool(
23
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
24
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
25
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
26
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
27
+ }
28
+
29
+ void _b901_wrapper(void* args[])
30
+ {
31
+ superkernel_add1d_attribute_broadcasting(
32
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
33
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
34
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
35
+ }
36
+
37
+ void _b906_wrapper(void* args[])
38
+ {
39
+ superkernel_clip1d(
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 _b881_wrapper(void* args[])
46
+ {
47
+ superkernel_mul1d_attribute_broadcasting(
48
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
49
+ *reinterpret_cast<uint32 const(*)[16]>(args[2]),
50
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]));
51
+ }
52
+
53
+ void _b891_wrapper(void* args[])
54
+ {
55
+ superkernel_mul1d(
56
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
57
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
58
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[1]),
59
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
60
+ }
61
+
62
+ void _b924_wrapper(void* args[])
63
+ {
64
+ superkernel_conv_eltbinary(
65
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
66
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
67
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]),
68
+ *reinterpret_cast<uint32 const(*)[17]>(args[4]),
69
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[3]));
70
+ }
71
+
72
+ void _b919_wrapper(void* args[])
73
+ {
74
+ superkernel_conv2d_dwc(
75
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[0]),
76
+ *reinterpret_cast<adf::input_buffer_conf<bfloat16, adf::bpc_sync_0d>*>(args[1]),
77
+ *reinterpret_cast<uint32 const(*)[16]>(args[3]),
78
+ *reinterpret_cast<adf::output_buffer_conf<bfloat16, adf::bpc_async_0d>*>(args[2]));
79
+ }
80
+
81
+ using UniformKernelFunc = void (*)(void **);
82
+
83
+ static UniformKernelFunc g_uniformKernelFuncs[7] = {
84
+ _b896_wrapper,
85
+ _b901_wrapper,
86
+ _b906_wrapper,
87
+ _b881_wrapper,
88
+ _b891_wrapper,
89
+ _b924_wrapper,
90
+ _b919_wrapper
91
+ };
92
+
93
+ __attribute__((always_inline)) void kernelWrapper(void* args[], uint32 kernelId, uint32 numSyncIn, uint32 numAsyncIn, uint32 numSyncOut)
94
+ {
95
+ uint32 idx = 0;
96
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 0);
97
+ idx += (numSyncIn > 0) ? 1 : 0;
98
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->acquire<adf::direction::in>(numSyncIn > 1);
99
+ idx += (numSyncIn > 1) ? 1 : 0;
100
+ idx += numAsyncIn;
101
+
102
+ (*(g_uniformKernelFuncs[kernelId]))(args);
103
+
104
+ idx = 0;
105
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 0);
106
+ idx += (numSyncIn > 0) ? 1 : 0;
107
+ reinterpret_cast<adf::_io_buffer_light_weight_generic*>(args[idx])->release<adf::direction::in>(numSyncIn > 1);
108
+ idx += (numSyncIn > 1) ? 1 : 0;
109
+ idx += numAsyncIn;
110
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_2_reloadable6/Release/0_2_reloadable6.calltree ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:46:39 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0_reloadable2 ../Release/0_0_reloadable2.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_reloadable2.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/chesswork1731 -pme
5
+
6
+
7
+ // Release: ipp V-2024.06-TGT-241219
8
+
9
+ _Z13kernelWrapperPPvjjjj
10
+ _Z26superkernel_reduce_mean_c8RN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
11
+ _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE5setupER18reduce_c8_params_tIS4_EPKv
12
+ _Z6pad_3dIL11pad_3d_mode0E8bfloat16Li1EEvPT0_S3_R15pad_3d_params_t
13
+ _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE3runEPS0_S6_R18reduce_c8_params_tIS4_E
14
+ _ZN12me_primitive10udiv_dstepEjjRjS0_
15
+ _ZN12me_primitive10udiv_dstepEjjRjS0_ (*)
16
+ int32_to_float32
17
+ _ZL28normalizeRoundAndPackFloat32iij
18
+ _ZL19roundAndPackFloat32iij
19
+ float32_add
20
+ _ZL14addFloat32Sigsjji
21
+ _ZL19propagateFloat32NaNjj
22
+ _ZL19roundAndPackFloat32iij (*)
23
+ _ZL14subFloat32Sigsjji
24
+ _ZL19propagateFloat32NaNjj (*)
25
+ _ZL28normalizeRoundAndPackFloat32iij (*)
26
+
27
+
28
+ Call tree stack and functions sizes:
29
+
30
+ stack stack stack call func func function name
31
+ desc level level desc
32
+ ----- ----- ----- ----- ----- ----- --------------------------------------------------------------
33
+ 64 448 0 0 220 10058 _Z13kernelWrapperPPvjjjj
34
+ 128 384 1 1 2676 9838 _Z26superkernel_reduce_mean_c8RN3adf9io_bufferI8bfloat16NS_9direction2inENS_16io_buffer_configINS_7extentsIJEEENS_7locking4syncENS_10addressing6linearENS_6marginILj0EEEEEEERA16_KjRNS0_IS1_NS2_3outENS4_IS6_NS7_5asyncESA_SC_EEEE
35
+ 64 64 2 2 1588 1588 _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE5setupER18reduce_c8_params_tIS4_EPKv
36
+ 0 0 2 2 670 670 _Z6pad_3dIL11pad_3d_mode0E8bfloat16Li1EEvPT0_S3_R15pad_3d_params_t
37
+ 256 256 2 2 2680 2822 _ZN18reduce_skeleton_c8I8bfloat1619reduce_mean_c8_implIS0_E23reduce_mean_c8_params_tIS0_EE3runEPS0_S6_R18reduce_c8_params_tIS4_E
38
+ 0 0 3 3 142 142 _ZN12me_primitive10udiv_dstepEjjRjS0_
39
+ 0 0 2 2 142 142 _ZN12me_primitive10udiv_dstepEjjRjS0_
40
+ 0 0 2 2 114 530 int32_to_float32
41
+ 0 0 2 3 24 416 _ZL28normalizeRoundAndPackFloat32iij <tail call>
42
+ 0 0 2 4 392 392 _ZL19roundAndPackFloat32iij <tail call>
43
+ 0 0 2 2 64 1968 float32_add
44
+ 0 0 3 3 624 1128 _ZL14addFloat32Sigsjji
45
+ 0 0 4 4 112 112 _ZL19propagateFloat32NaNjj
46
+ 0 0 3 4 392 392 _ZL19roundAndPackFloat32iij <tail call>
47
+ 0 0 2 3 752 1280 _ZL14subFloat32Sigsjji <tail call>
48
+ 0 0 3 4 112 112 _ZL19propagateFloat32NaNjj
49
+ 0 0 2 4 24 416 _ZL28normalizeRoundAndPackFloat32iij <tail call> (*)
50
+
51
+
52
+ Maximum call level : 4
53
+ Maximum stack level: 4
54
+ Maximum stack size : 448