G2D图像处理硬件调用和测试-基于米尔-全志T113-i开发板

电子

本篇测评由电子工程世界的优秀测评者“jf_99374259”提供。

本文将介绍基于米尔电子MYD-YT113i开发板的G2D图像处理硬件调用和测试。MYC-YT113i核心板及开发板
真正的国产核心板，100%国产物料认证

国产T113-i处理器配备2*Cortex-A7@1.2GHz ，RISC-V
外置DDR3接口、支持视频编解码器、HiFi4 DSP
接口丰富：视频采集接口、显示器接口、USB2.0 接口、CAN 接口、千兆以太网接口
工业级：-40℃~+85℃、尺寸37mm*39mm
邮票孔+LGA，140+50PIN

全志 T113-i 2D图形加速硬件支持情况

Supports layer size up to 2048 x 2048 pixels
Supports pre-multiply alpha image data
Supports color key
Supports two pipes Porter-Duff alpha blending
Supports multiple video formats 4:2:0, 4:2:2, 4:1:1 and multiple pixel formats (8/16/24/32 bits graphics
layer)
Supports memory scan order option
Supports any format convert function
Supports 1/16× to 32× resize ratio
Supports 32-phase 8-tap horizontal anti-alias filter and 32-phase 4-tap vertical anti-alias filter
Supports window clip
Supports FillRectangle, BitBlit, StretchBlit and MaskBlit
Supports horizontal and vertical flip, clockwise 0/90/180/270 degree rotate for normal buffer
Supports horizontal flip, clockwise 0/90/270 degree rotate for LBC buffer

可以看到 g2d 硬件支持相当多的2D图像处理，包括颜色空间转换，分辨率缩放，图层叠加，旋转等

开发环境配置
基础开发环境搭建参考上上上一篇
https://bbs.elecfans.com/jishu_2408808_1_1.html
除了工具链外，我们使用 opencv-mobile 加载输入图片和保存结果，用来查看颜色转换是否正常g2d硬件直接采用标准的 Linux ioctl 操纵，只需要引入相关结构体定义即可，无需链接so
https://github.com/MYIR-ALLWINNER/framework/blob/develop-yt113-framework/auto/sdk_lib/include/g2d_driver.h
此外，g2d的输入和输出数据必须在dmaion buffer上，因此还需要dmaion.h头文件，用来分配和释放dmaion buffer
https://github.com/MYIR-ALLWINNER/framework/blob/develop-yt113-framework/auto/sdk_lib/include/DmaIon.h

基于C语言实现的YUV转RGB
这里复用之前T113-i JPG解码的函数

void yuv420sp2rgb(const unsigned char* yuv420sp, int w, int h, unsigned char* rgb)
{
const unsigned char* yptr = yuv420sp;
const unsigned char* vuptr = yuv420sp + w * h;
for (int y = 0; y < h; y += 2)
{
const unsigned char* yptr0 = yptr;
const unsigned char* yptr1 = yptr + w;
unsigned char* rgb0 = rgb;
unsigned char* rgb1 = rgb + w * 3;
int remain = w;
#define SATURATE_CAST_UCHAR(X) (unsigned char)::std::min(::std::max((int)(X), 0), 255);
for (; remain > 0; remain -= 2)
{
// R = 1.164 * yy + 1.596 * vv
// G = 1.164 * yy - 0.813 * vv - 0.391 * uu
// B = 1.164 * yy + 2.018 * uu
// R = Y + (1.370705 * (V-128))
// G = Y - (0.698001 * (V-128)) - (0.337633 * (U-128))
// B = Y + (1.732446 * (U-128))
// R = ((Y << 6) + 87.72512 * (V-128)) >> 6
// G = ((Y << 6) - 44.672064 * (V-128) - 21.608512 * (U-128)) >> 6
// B = ((Y << 6) + 110.876544 * (U-128)) >> 6
// R = ((Y << 6) + 90 * (V-128)) >> 6
// G = ((Y << 6) - 46 * (V-128) - 22 * (U-128)) >> 6
// B = ((Y << 6) + 113 * (U-128)) >> 6
// R = (yy + 90 * vv) >> 6
// G = (yy - 46 * vv - 22 * uu) >> 6
// B = (yy + 113 * uu) >> 6
int v = vuptr[0] - 128;
int u = vuptr[1] - 128;
int ruv = 90 * v;
int guv = -46 * v + -22 * u;
int buv = 113 * u;
int y00 = yptr0[0] << 6;
rgb0[0] = SATURATE_CAST_UCHAR((y00 + ruv) >> 6);
rgb0[1] = SATURATE_CAST_UCHAR((y00 + guv) >> 6);
rgb0[2] = SATURATE_CAST_UCHAR((y00 + buv) >> 6);
int y01 = yptr0[1] << 6;
rgb0[3] = SATURATE_CAST_UCHAR((y01 + ruv) >> 6);
rgb0[4] = SATURATE_CAST_UCHAR((y01 + guv) >> 6);
rgb0[5] = SATURATE_CAST_UCHAR((y01 + buv) >> 6);
int y10 = yptr1[0] << 6;
rgb1[0] = SATURATE_CAST_UCHAR((y10 + ruv) >> 6);
rgb1[1] = SATURATE_CAST_UCHAR((y10 + guv) >> 6);
rgb1[2] = SATURATE_CAST_UCHAR((y10 + buv) >> 6);
int y11 = yptr1[1] << 6;
rgb1[3] = SATURATE_CAST_UCHAR((y11 + ruv) >> 6);
rgb1[4] = SATURATE_CAST_UCHAR((y11 + guv) >> 6);
rgb1[5] = SATURATE_CAST_UCHAR((y11 + buv) >> 6);
yptr0 += 2;
yptr1 += 2;
vuptr += 2;
rgb0 += 6;
rgb1 += 6;
}
#undef SATURATE_CAST_UCHAR
yptr += 2 * w;
rgb += 2 * 3 * w;
}
}

复制代码

基于ARM neon指令集优化的YUV转RGB
考虑到armv7编译器的自动neon优化能力较差，这里针对性的编写 arm neon inline assembly 实现YUV2RGB内核部分，达到最优化的性能，榨干cpu性能

void yuv420sp2rgb_neon(const unsigned char* yuv420sp, int w, int h, unsigned char* rgb)
{
const unsigned char* yptr = yuv420sp;
const unsigned char* vuptr = yuv420sp + w * h;

#if __ARM_NEON
uint8x8_t _v128 = vdup_n_u8(128);
int8x8_t _v90 = vdup_n_s8(90);
int8x8_t _v46 = vdup_n_s8(46);
int8x8_t _v22 = vdup_n_s8(22);
int8x8_t _v113 = vdup_n_s8(113);
#endif // __ARM_NEON

for (int y = 0; y < h; y += 2)

{

const unsigned char* yptr0 = yptr;

const unsigned char* yptr1 = yptr + w;

unsigned char* rgb0 = rgb;

unsigned char* rgb1 = rgb + w * 3;

#if __ARM_NEON

int nn = w >> 3;

int remain = w - (nn << 3);

#else

int remain = w;

#endif // __ARM_NEON

#if __ARM_NEON

#if __aarch64__

for (; nn > 0; nn--)

{

int16x8_t _yy0 = vreinterpretq_s16_u16(vshll_n_u8(vld1_u8(yptr0), 6));

int16x8_t _yy1 = vreinterpretq_s16_u16(vshll_n_u8(vld1_u8(yptr1), 6));

int8x8_t _vvuu = vreinterpret_s8_u8(vsub_u8(vld1_u8(vuptr), _v128));

int8x8x2_t _vvvvuuuu = vtrn_s8(_vvuu, _vvuu);

int8x8_t _vv = _vvvvuuuu.val[0];

int8x8_t _uu = _vvvvuuuu.val[1];

int16x8_t _r0 = vmlal_s8(_yy0, _vv, _v90);

int16x8_t _g0 = vmlsl_s8(_yy0, _vv, _v46);

_g0 = vmlsl_s8(_g0, _uu, _v22);

int16x8_t _b0 = vmlal_s8(_yy0, _uu, _v113);

int16x8_t _r1 = vmlal_s8(_yy1, _vv, _v90);

int16x8_t _g1 = vmlsl_s8(_yy1, _vv, _v46);

_g1 = vmlsl_s8(_g1, _uu, _v22);

int16x8_t _b1 = vmlal_s8(_yy1, _uu, _v113);

uint8x8x3_t _rgb0;

_rgb0.val[0] = vqshrun_n_s16(_r0, 6);

_rgb0.val[1] = vqshrun_n_s16(_g0, 6);

_rgb0.val[2] = vqshrun_n_s16(_b0, 6);

uint8x8x3_t _rgb1;

_rgb1.val[0] = vqshrun_n_s16(_r1, 6);

_rgb1.val[1] = vqshrun_n_s16(_g1, 6);

_rgb1.val[2] = vqshrun_n_s16(_b1, 6);

vst3_u8(rgb0, _rgb0);

vst3_u8(rgb1, _rgb1);

yptr0 += 8;

yptr1 += 8;

vuptr += 8;

rgb0 += 24;

rgb1 += 24;

}

#else

if (nn > 0)

{

asm volatile(

"0: n"

"pld [%3, #128] n"

"vld1.u8 {d2}, [%3]! n"

"vsub.s8 d2, d2, %12 n"

"pld [%1, #128] n"

"vld1.u8 {d0}, [%1]! n"

"pld [%2, #128] n"

"vld1.u8 {d1}, [%2]! n"

"vshll.u8 q2, d0, #6 n"

"vorr d3, d2, d2 n"

"vshll.u8 q3, d1, #6 n"

"vorr q9, q2, q2 n"

"vtrn.s8 d2, d3 n"

"vorr q11, q3, q3 n"

"vmlsl.s8 q9, d2, %14 n"

"vorr q8, q2, q2 n"

"vmlsl.s8 q11, d2, %14 n"

"vorr q10, q3, q3 n"

"vmlal.s8 q8, d2, %13 n"

"vmlal.s8 q2, d3, %16 n"

"vmlal.s8 q10, d2, %13 n"

"vmlsl.s8 q9, d3, %15 n"

"vmlal.s8 q3, d3, %16 n"

"vmlsl.s8 q11, d3, %15 n"

"vqshrun.s16 d24, q8, #6 n"

"vqshrun.s16 d26, q2, #6 n"

"vqshrun.s16 d4, q10, #6 n"

"vqshrun.s16 d25, q9, #6 n"

"vqshrun.s16 d6, q3, #6 n"

"vqshrun.s16 d5, q11, #6 n"

"subs %0, #1 n"

"vst3.u8 {d24-d26}, [%4]! n"

"vst3.u8 {d4-d6}, [%5]! n"

"bne 0b n"

: "=r"(nn), // %0

"=r"(yptr0), // %1

"=r"(yptr1), // %2

"=r"(vuptr), // %3

"=r"(rgb0), // %4

"=r"(rgb1) // %5

: "0"(nn),

"1"(yptr0),

"2"(yptr1),

"3"(vuptr),

"4"(rgb0),

"5"(rgb1),

"w"(_v128), // %12

"w"(_v90), // %13

"w"(_v46), // %14

"w"(_v22), // %15

"w"(_v113) // %16

: "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "d26");

}

#endif // __aarch64__

#endif // __ARM_NEON

#define SATURATE_CAST_UCHAR(X) (unsigned char)::std::min(::std::max((int)(X), 0), 255);

for (; remain > 0; remain -= 2)

{

// R = 1.164 * yy + 1.596 * vv

// G = 1.164 * yy - 0.813 * vv - 0.391 * uu

// B = 1.164 * yy + 2.018 * uu

// R = Y + (1.370705 * (V-128))

// G = Y - (0.698001 * (V-128)) - (0.337633 * (U-128))

// B = Y + (1.732446 * (U-128))

// R = ((Y << 6) + 87.72512 * (V-128)) >> 6

// G = ((Y << 6) - 44.672064 * (V-128) - 21.608512 * (U-128)) >> 6

// B = ((Y << 6) + 110.876544 * (U-128)) >> 6

// R = ((Y << 6) + 90 * (V-128)) >> 6

// G = ((Y << 6) - 46 * (V-128) - 22 * (U-128)) >> 6

// B = ((Y << 6) + 113 * (U-128)) >> 6

// R = (yy + 90 * vv) >> 6

// G = (yy - 46 * vv - 22 * uu) >> 6

// B = (yy + 113 * uu) >> 6

int v = vuptr[0] - 128;

int u = vuptr[1] - 128;

int ruv = 90 * v;

int guv = -46 * v + -22 * u;

int buv = 113 * u;

int y00 = yptr0[0] << 6;

rgb0[0] = SATURATE_CAST_UCHAR((y00 + ruv) >> 6);

rgb0[1] = SATURATE_CAST_UCHAR((y00 + guv) >> 6);

rgb0[2] = SATURATE_CAST_UCHAR((y00 + buv) >> 6);

int y01 = yptr0[1] << 6;

rgb0[3] = SATURATE_CAST_UCHAR((y01 + ruv) >> 6);

rgb0[4] = SATURATE_CAST_UCHAR((y01 + guv) >> 6);

rgb0[5] = SATURATE_CAST_UCHAR((y01 + buv) >> 6);

int y10 = yptr1[0] << 6;

rgb1[0] = SATURATE_CAST_UCHAR((y10 + ruv) >> 6);

rgb1[1] = SATURATE_CAST_UCHAR((y10 + guv) >> 6);

rgb1[2] = SATURATE_CAST_UCHAR((y10 + buv) >> 6);

int y11 = yptr1[1] << 6;

rgb1[3] = SATURATE_CAST_UCHAR((y11 + ruv) >> 6);

rgb1[4] = SATURATE_CAST_UCHAR((y11 + guv) >> 6);

rgb1[5] = SATURATE_CAST_UCHAR((y11 + buv) >> 6);

yptr0 += 2;

yptr1 += 2;

vuptr += 2;

rgb0 += 6;

rgb1 += 6;

}

#undef SATURATE_CAST_UCHAR

yptr += 2 * w;

rgb += 2 * 3 * w;

}

复制代码

基于G2D图形硬件的YUV转RGB
我们先实现 dmaion buffer 管理器，参考
https://github.com/MYIR-ALLWINNER/framework/blob/develop-yt113-framework/auto/sdk_lib/sdk_memory/DmaIon.cpp
这里贴的代码省略了异常错误处理的逻辑，有个坑是 linux-4.9 和 linux-5.4 用法不一样，米尔电子的这个T113-i系统是linux-5.4，所以不兼容4.9内核的ioctl用法习惯

[size=18px]struct ion_memory[/size]
[size=18px]{[/size]
[size=18px] size_t size;[/size]
[size=18px] int fd;[/size]
[size=18px] void* virt_addr;[/size]
[size=18px] unsigned int phy_addr;[/size]
[size=18px]};[/size]
[size=18px]class ion_allocator[/size]
[size=18px]{[/size]
[size=18px]public:[/size]
[size=18px] ion_allocator();[/size]
[size=18px] ~ion_allocator();[/size]
[size=18px] [/size]
[size=18px] int open();[/size]
[size=18px] void close();[/size]
[size=18px] [/size]
[size=18px] int alloc(size_t size, struct ion_memory* mem);[/size]
[size=18px] int free(struct ion_memory* mem);[/size]
[size=18px] [/size]
[size=18px] int flush(struct ion_memory* mem);[/size]
[size=18px] [/size]
[size=18px]public:[/size]
[size=18px] int ion_fd;[/size]
[size=18px] int cedar_fd;[/size]
[size=18px]};[/size]
[size=18px] ion_allocator::ion_allocator()[/size]
[size=18px]{[/size]
[size=18px] ion_fd = -1;[/size]
[size=18px] cedar_fd = -1;[/size]
[size=18px]}[/size]
[size=18px]ion_allocator::~ion_allocator()[/size]
[size=18px]{[/size]
[size=18px] close();[/size]
[size=18px]}[/size]
[size=18px]int ion_allocator::open()[/size]
[size=18px]{[/size]
[size=18px] close();[/size]
[size=18px] [/size]
[size=18px] ion_fd = ::open("/dev/ion", O_RDWR);[/size]
[size=18px] cedar_fd = ::open("/dev/cedar_dev", O_RDONLY);[/size]
[size=18px] [/size]
[size=18px] ioctl(cedar_fd, IOCTL_ENGINE_REQ, 0);[/size]
[size=18px] [/size]
[size=18px] return 0;[/size]
[size=18px]}[/size]
[size=18px]void ion_allocator::close()[/size]
[size=18px]{[/size]
[size=18px] if (cedar_fd != -1)[/size]
[size=18px] {[/size]
[size=18px] ioctl(cedar_fd, IOCTL_ENGINE_REL, 0);[/size]
[size=18px] ::close(cedar_fd);[/size]
[size=18px] cedar_fd = -1;[/size]
[size=18px] }[/size]
[size=18px] [/size]
[size=18px] if (ion_fd != -1)[/size]
[size=18px] {[/size]
[size=18px] ::close(ion_fd);[/size]
[size=18px] ion_fd = -1;[/size]
[size=18px] }[/size]
[size=18px]}[/size]
[size=18px]int ion_allocator::alloc(size_t size, struct ion_memory* mem)[/size]
[size=18px]{[/size]
[size=18px] struct aw_ion_new_alloc_data alloc_data;[/size]
[size=18px] alloc_data.len = size;[/size]
[size=18px] alloc_data.heap_id_mask = AW_ION_SYSTEM_HEAP_MASK;[/size]
[size=18px] alloc_data.flags = AW_ION_CACHED_FLAG | AW_ION_CACHED_NEEDS_SYNC_FLAG;[/size]
[size=18px] alloc_data.fd = 0;[/size]
[size=18px] alloc_data.unused = 0;[/size]
[size=18px] ioctl(ion_fd, AW_ION_IOC_NEW_ALLOC, &alloc_data);[/size]
[size=18px] [/size]
[size=18px] void* virt_addr = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_SHARED, alloc_data.fd, 0);[/size]
[size=18px] [/size]
[size=18px] struct aw_user_iommu_param iommu_param;[/size]
[size=18px] iommu_param.fd = alloc_data.fd;[/size]
[size=18px] iommu_param.iommu_addr = 0;[/size]
[size=18px] ioctl(cedar_fd, IOCTL_GET_IOMMU_ADDR, &iommu_param);[/size]
[size=18px] [/size]
[size=18px] mem->size = size;[/size]
[size=18px] mem->fd = alloc_data.fd;[/size]
[size=18px] mem->virt_addr = virt_addr;[/size]
[size=18px] mem->phy_addr = iommu_param.iommu_addr;[/size]
[size=18px] [/size]
[size=18px] return 0;[/size]
[size=18px]}[/size]
[size=18px]int ion_allocator::free(struct ion_memory* mem)[/size]
[size=18px]{[/size]
[size=18px] if (mem->fd == -1)[/size]
[size=18px] return 0;[/size]
[size=18px] [/size]
[size=18px] struct aw_user_iommu_param iommu_param;[/size]
[size=18px] iommu_param.fd = mem->fd;[/size]
[size=18px] ioctl(cedar_fd, IOCTL_FREE_IOMMU_ADDR, &iommu_param);[/size]
[size=18px] [/size]
[size=18px] munmap(mem->virt_addr, mem->size);[/size]
[size=18px] [/size]
[size=18px] ::close(mem->fd);[/size]
[size=18px] [/size]
[size=18px] mem->size = 0;[/size]
[size=18px] mem->fd = -1;[/size]
[size=18px] mem->virt_addr = 0;[/size]
[size=18px] mem->phy_addr = 0;[/size]
[size=18px] [/size]
[size=18px] return 0;[/size]
[size=18px]}[/size]
[size=18px]int ion_allocator::flush(struct ion_memory* mem)[/size]
[size=18px]{[/size]
[size=18px] struct dma_buf_sync sync;[/size]
[size=18px] sync.flags = DMA_BUF_SYNC_END | DMA_BUF_SYNC_RW;[/size]
[size=18px] ioctl(mem->fd, DMA_BUF_IOCTL_SYNC, &sync);[/size]
[size=18px] [/size]
[size=18px] return 0;[/size]
[size=18px]}[/size]

复制代码

然后再实现 G2D图形硬件 YUV转RGB 的转换器

提前分配好YUV和RGB的dmaion buffer
将YUV数据拷贝到dmaion buffer，flush cache完成同步
配置转换参数，ioctl调用G2D_CMD_BITBLT_H完成转换
flush cache完成同步，从dmaion buffer拷贝出RGB数据
释放dmaion buffer

[size=18px]// 步骤1[/size]
[size=18px]ion_allocator ion;[/size]
[size=18px]ion.open();[/size]
[size=18px]struct ion_memory yuv_ion;[/size]
[size=18px]ion.alloc(rgb_size, &rgb_ion);[/size]
[size=18px]struct ion_memory rgb_ion;[/size]
[size=18px]ion.alloc(yuv_size, &yuv_ion);[/size]
[size=18px]int g2d_fd = ::open("/dev/g2d", O_RDWR);[/size]
[size=18px]// 步骤2[/size]
[size=18px]memcpy((unsigned char*)yuv_ion.virt_addr, yuv420sp, yuv_size);[/size]
[size=18px]ion.flush(&yuv_ion);[/size]
[size=18px]// 步骤3[/size]
[size=18px]g2d_blt_h blit;[/size]
[size=18px]memset(&blit, 0, sizeof(blit));[/size]
[size=18px]blit.flag_h = G2D_BLT_NONE_H;[/size]
[size=18px]blit.src_image_h.format = G2D_FORMAT_YUV420UVC_V1U1V0U0;[/size]
[size=18px]blit.src_image_h.width = width;[/size]
[size=18px]blit.src_image_h.height = height;[/size]
[size=18px]blit.src_image_h.align[0] = 0;[/size]
[size=18px]blit.src_image_h.align[1] = 0;[/size]
[size=18px]blit.src_image_h.clip_rect.x = 0;[/size]
[size=18px]blit.src_image_h.clip_rect.y = 0;[/size]
[size=18px]blit.src_image_h.clip_rect.w = width;[/size]
[size=18px]blit.src_image_h.clip_rect.h = height;[/size]
[size=18px]blit.src_image_h.gamut = G2D_BT601;[/size]
[size=18px]blit.src_image_h.bpremul = 0;[/size]
[size=18px]blit.src_image_h.mode = G2D_PIXEL_ALPHA;[/size]
[size=18px]blit.src_image_h.use_phy_addr = 0;[/size]
[size=18px]blit.src_image_h.fd = yuv_ion.fd;[/size]
[size=18px]blit.dst_image_h.format = G2D_FORMAT_RGB888;[/size]
[size=18px]blit.dst_image_h.width = width;[/size]
[size=18px]blit.dst_image_h.height = height;[/size]
[size=18px]blit.dst_image_h.align[0] = 0;[/size]
[size=18px]blit.dst_image_h.clip_rect.x = 0;[/size]
[size=18px]blit.dst_image_h.clip_rect.y = 0;[/size]
[size=18px]blit.dst_image_h.clip_rect.w = width;[/size]
[size=18px]blit.dst_image_h.clip_rect.h = height;[/size]
[size=18px]blit.dst_image_h.gamut = G2D_BT601;[/size]
[size=18px]blit.dst_image_h.bpremul = 0;[/size]
[size=18px]blit.dst_image_h.mode = G2D_PIXEL_ALPHA;[/size]
[size=18px]blit.dst_image_h.use_phy_addr = 0;[/size]
[size=18px]blit.dst_image_h.fd = rgb_ion.fd;[/size]
[size=18px]ioctl(g2d_fd, G2D_CMD_BITBLT_H, &blit);[/size]
[size=18px]// 步骤4[/size]
[size=18px]ion.flush(&rgb_ion);[/size]
[size=18px]memcpy(rgb, (const unsigned char*)rgb_ion.virt_addr, rgb_size);[/size]
[size=18px]// 步骤5[/size]
[size=18px]ion.free(&rgb_ion);[/size]
[size=18px]ion.free(&yuv_ion);[/size]
[size=18px]ion.close();[/size]
[size=18px]::close(g2d_fd);[/size]

复制代码

G2D图像硬件YUV转RGB测试
考虑到dmaion buffer分配和释放都比较耗时，我们提前做好，循环调用步骤3的G2D转换，统计耗时，并在top工具中查看CPU占用率
[code]sh-4.4# LD_LIBRARY_PATH=. ./g2dtest
INFO : cedarc

搜索历史

G2D图像处理硬件调用和测试-基于米尔-全志T113-i开发板

相关推荐

评论

精选推荐

站长推荐 /6