Enabling neon on imx8m plus

[matheus-nogueira]

Hello, we are using the imx8m plus to do ultrasound signal processing. At first we tried openCL, but results are too slow for our application so now we are trying it with neon. The problem is that the cross-compile gcc generated with yocto don't recognize -mfloat-abi neither -mfpu flags, and when we execute commands to detect neon on linux it doesn't show up (https://developer.arm.com/documentation/den0018/a/Compiling-NEON-Instructions/Detecting-presence-of-...).

I wrote some code using neon intrinsics and I can compile using the make file below, but I see in assembly code that it isn't using neon instructions like vadd, vmul etc.

What should I do to enable neon? Should I do something on yocto?

Makefile:

PROJ_NAME = frame_processing

# Diretório do sysroot e do Yocto SDK

ROOTFS_DIR = /opt/poky/4.0.6/sysroots/cortexa53-crypto-poky-linux

TARGET_PATH_LIB = $(ROOTFS_DIR)/usr/lib

TARGET_PATH_INCLUDE = $(ROOTFS_DIR)/usr/include

# Diretório da toolchain Yocto

TOOLCHAIN_DIR = /opt/poky/4.0.6/sysroots/x86_64-pokysdk-linux/usr/bin/aarch64-poky-linux

# Fontes C e Headers

C_SOURCE = $(wildcard *.c)

H_SOURCE = $(wildcard *.h)

OBJ = $(C_SOURCE:.c=.o)

# Compilador correto do Yocto SDK

CC = $(TOOLCHAIN_DIR)/aarch64-poky-linux-gcc

# Flags de compilação e linkagem

CPU_CFLAGS = --sysroot=$(ROOTFS_DIR) -Wall -Wextra -std=c11 -O3 -Ofast -I$(TARGET_PATH_INCLUDE) -g -mcpu=cortex-a53 -march=armv8-a+simd -std=gnu11 -ftree-vectorize -fopt-info-vec-missed

CPU_LFLAGS = --sysroot=$(ROOTFS_DIR) -L$(TARGET_PATH_LIB) -lm

# Regra principal

all: $(PROJ_NAME)

# Regra para gerar o binário final

$(PROJ_NAME): $(OBJ)

$(CC) $(CPU_CFLAGS) $(CPU_LFLAGS) -o $@ $^

# Compilação dos arquivos .o

%.o: %.c

$(CC) $(CPU_CFLAGS) -c -o $@ $<

# Regra para gerar código assembly (.s) a partir de .c

%.s: %.c

$(CC) $(CPU_CFLAGS) -S -o $@ $< $(CPU_LFLAGS)

# Gera os arquivos .s (assembly) para todos os arquivos .c

assembly: $(C_SOURCE:.c=.s)

# Limpeza dos arquivos compilados

clean:

rm -f $(OBJ) $(PROJ_NAME) *.s

[Bio_TICFSL]

Hello,

You have to upgrade your gcc or toolchain, current yocto BSP has neon and fpu support on cortex A53, also make sure neon its enable in kernel configuration file.

Regards

[matheus-nogueira]

Thanks for responding. We tried modifying our yocto configuration file, generated a new cross-compile toolchain and compiled a simple .c file using makefile. But as we understand it, we have not yet compiled vector instructions nor have we seen neon listed when using cat /proc/cpuinfo. Could you tell me what and how we should modify our files? I attached the files as .txt.