whisper.Android WhisperCppDemo very slow android specific transcibe times for 3s recording a res of 31227ms

[Anioz]

System info shows support on samsung note10 device for fp16_va, neon, arm_fma, release build was made with Android studio and file contents:
Default provided file contents where tested first before trying the below options, all in release build model tiny-en.bin

for Android.mk:
ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
include $(CLEAR_VARS)
LOCAL_MODULE := libwhisper_v8fp16_va
include $(LOCAL_PATH)/Whisper.mk
# Enable VFPv16, NEON, and FMA instruction sets
LOCAL_CFLAGS += -march=armv8.2-a+fp16 -mfpu=neon-v8fp16_va
include $(BUILD_SHARED_LIBRARY)
endif

for Application mk:
APP_STL := c++_static
APP_ABI := arm64-v8a

Extra manifest declarations are added:
android:largeHeap="true"
android:hardwareAccelerated="true" (this one could possibly be irrelevant because most of the time since a certain sdk level its
always been at true even without declaring this specifically)

None of these seem to bring the times within acceptable time parameters however, I would love to try some things out if you propose them and will share the results when tested.

Please put me in the right direction as I love your work and want to maximize its performance on Android devices

*Edit: Ive found that disabling R8 shrinking minifyenabled = false increases transcribe times by 50% (3040ms speech, 14762ms transcribe time coming down from 31227ms) This was done with default project repo values

[sandorkonya]

FYI Android Inference is too slow and this one
Share a simpler Cmake methd to compile and run GPU accelerated version(openBLAS and CLBlast) on android could help.