neon instruction

[kumarchaitanyab]

Hi Team,
I have compiled and executed the Compute Library and the neon_cnn example code on an RP5 board with an ARM Cortex-A76 core. Although the test appears to have passed, I have a few questions regarding the generated assembly code. Specifically, I did not find any NEON instructions in the generated list file, and I am unsure if I might have missed something.

Could you confirm whether the example actually utilizes NEON SIMD instructions?

I have attached the generated list file for your reference.

commands:

to generate compute library:
scons Werror=1 -j8 debug=1 neon=1 opencl=0 os=linux arch=arm64-v8.2-a build=native

neon_cnn example:
g++ examples/neon_cnn.cpp utils/Utils.cpp -I. -Iinclude -L$PWD/build_debug_1 -larm_compute -larm_compute_graph -std=c++14 -g -o neon_cnn

**

neon_cnn_list.docx

**
objdump -d -S neon_cnn > neon_cnn.lst

[morgolock]

Hi @kumarchaitanyab

The reason why you don't see neon op codes in neon_cnn is because you linked dynamically with libarm_compute.so. All the neon code should be in the .so file and this gets called from the test.

You could alternatively link statically with libarm_compute-static.a and then your binary will have the neon op codes.

Hope this helps

[kumarchaitanyab]

Hi @kumarchaitanyab

The reason why you don't see neon op codes in neon_cnn is because you linked dynamically with libarm_compute.so. All the neon code should be in the .so file and this gets called from the test.

You could alternatively link statically with libarm_compute-static.a and then your binary will have the neon op codes.

Hope this helps
Thanks for your reply.

I used below command and even then I am unable to see neon opcodes.

g++ examples/neon_cnn.cpp utils/Utils.cpp -I. -Iinclude -L$PWD/build -larm_compute-static -larm_compute_graph-static -std=c++14 -g -o neon_cnn

[morgolock]

Can you please share the command used to build the library?

[kumarchaitanyab]

command for library is "scons Werror=1 -j8 debug=1 neon=1 opencl=0 os=linux arch=arm64-v8.2-a build=native"

[morgolock]

Hi @kumarchaitanyab

In ACL the SIMD code for the CPU backend is in libarm_compute.so and libarm_compute-static.a.

You can easily see all the neon function in ACL are listed in the binary's symbols:
nm -C build/libarm_compute-static.a | grep neon

You can also see the sdot instruction present in the binary with

~/work/acl/ComputeLibrary$ ../../toolchains/android-ndk-r26d/toolchains/llvm/prebuilt/linux-x86_64/bin/llvm-objdump -d build/libarm_compute.so | grep -E sdot  | tail -n 5
  75edec: 4f81e0ef     	sdot	v15.4s, v7.16b, v1.4b[0]
  75edf0: 4f82e0f3     	sdot	v19.4s, v7.16b, v2.4b[0]
  75edf4: 4f83e0f7     	sdot	v23.4s, v7.16b, v3.4b[0]
  75edf8: 4f84e0fb     	sdot	v27.4s, v7.16b, v4.4b[0]
  75edfc: 4f85e0ff     	sdot	v31.4s, v7.16b, v5.4b[0]

Hope this helps

[kumarchaitanyab]

Hi @morgolock , I am unable to see the neon instructions in the list file generated from ELF file. Weather I use *.so or *.a file for linking. Can youi please help here?

[morgolock]

Hi @kumarchaitanyab

How are you generating the list? If you build the library for arch=arm64-v8.2-a and debug=0 neon=1 and then run
llvm-objdump -d build/libarm_compute.so | grep -E sdot | tail -n 5 you should see the sdot opcodes as shown above.

See these other commands the commands below confirming the presence of neon instructions fmla and add.

ComputeLibrary$ ../../toolchains/android-ndk-r26d/toolchains/llvm/prebuilt/linux-x86_64/bin/llvm-objdump -C -d build/libarm_compute.so | grep -E 'fmla[[:space:]]+v' | tail -n 10
  77d4d0: 4f81189e     	fmla	v30.4s, v4.4s, v1.s[2]
  77d4d8: 4fa1184f     	fmla	v15.4s, v2.4s, v1.s[3]
  77d4e0: 4fa11877     	fmla	v23.4s, v3.4s, v1.s[3]
  77d4e8: 4fa1189f     	fmla	v31.4s, v4.4s, v1.s[3]
  79037c: 4e34cc15     	fmla	v21.4s, v0.4s, v20.4s
  790388: 4e33cc19     	fmla	v25.4s, v0.4s, v19.4s
  790394: 4e34ccfb     	fmla	v27.4s, v7.4s, v20.4s
  79039c: 4e34ccdc     	fmla	v28.4s, v6.4s, v20.4s
  7903a4: 4e34cc99     	fmla	v25.4s, v4.4s, v20.4s
  7903a8: 4e34cc7a     	fmla	v26.4s, v3.4s, v20.4s

ComputeLibrary$ ../../toolchains/android-ndk-r26d/toolchains/llvm/prebuilt/linux-x86_64/bin/llvm-objdump -C -d build/libarm_compute.so | grep -E 'add[[:space:]]+v' | tail -n 10
  78fc58: 4ea487de     	add	v30.4s, v30.4s, v4.4s
  78fc5c: 4ea48718     	add	v24.4s, v24.4s, v4.4s
  78fc60: 4ea48739     	add	v25.4s, v25.4s, v4.4s
  78fc64: 4ea4875a     	add	v26.4s, v26.4s, v4.4s
  78fc68: 4ea4877b     	add	v27.4s, v27.4s, v4.4s
  790384: 4e34d678     	fadd	v24.4s, v19.4s, v20.4s
  79038c: 4e35d6f5     	fadd	v21.4s, v23.4s, v21.4s
  790398: 4e38d6f8     	fadd	v24.4s, v23.4s, v24.4s
  7903b0: 4e3cd67c     	fadd	v28.4s, v19.4s, v28.4s
  7903b4: 4e3bd673     	fadd	v19.4s, v19.4s, v27.4s