在使用ffmpeg解码后的多路解码数据非常慢,还要给AI做行的加速方式是在显存处理数据,在视频拼接融合产品的产品与架构设计中,提出了比较可靠的方式是使用cuda,那么没有cuda的显卡如何处理呢
,比较好的方式是使用opencl来提高数据传输效率
核函数
在OpenCL中,将NV12格式转换为BGR格式通常涉及到对UV分量的处理,nv12 是使用ffmpeg等解码后的直接数据,注意linesize对齐
c
#define GROUP_SIZE 16
// OpenCL kernel to convert NV12 to BGR
__kernel void nv12_to_bgr(__global const uchar *nv12,
__global uchar *bgr,
int width, int height) {
int x = get_global_id(0);
int y = get_global_id(1);
// Make sure we are not out of bounds
if (x < width && y < height) {
// Calculate Y, U, and V indices
int yIndex = y * width + x;
int uvIndex = width * height + (y / 2) * (width) + (x & ~1); // Use '& ~1' to get even X indices for U/V
// Load Y, U, and V values
uchar yValue = nv12[yIndex];
uchar uValue = nv12[uvIndex];
uchar vValue = nv12[uvIndex + 1];
// Convert YUV to RGB
uchar bValue = (uchar)((yValue + 1.732446 * (uValue - 128));
uchar gValue = (uchar)((yValue - 0.344134 * (vValue - 128) - 0.714136 * (uValue - 128));
uchar rValue = (uchar)((yValue + 1.402225 * (vValue - 128));
// Pack BGR values
uchar bgrValue = (bValue << 2) | (gValue >> 4) | (rValue << 6);
// Store BGR value
bgr[yIndex] = bgrValue;
}
}
cpu上继续
注意错误处理
c
// 设置OpenCL内核参数
size_t global_work_size[2] = {width, height};
cl_kernel nv12_to_bgr_kernel = ...; // 获取你编译的内核
// 设置内核参数
clSetKernelArg(nv12_to_bgr_kernel, 0, sizeof(cl_mem), &nv12_buffer);
clSetKernelArg(nv12_to_bgr_kernel, 1, sizeof(cl_mem), &bgr_buffer);
clSetKernelArg(nv12_to_bgr_kernel, 2, sizeof(int), &width);
clSetKernelArg(nv12_to_bgr_kernel, 3, sizeof(int), &height);
// 执行内核
cl_event event;
clEnqueueNDRangeKernel(command_queue, nv12_to_bgr_kernel, 2, NULL, global_work_size, NULL, 0, NULL, &event);
// 等待命令执行完毕
clWaitForEvents(1, &event);
针对arm,非显存
用128位的寄存器进行处理。
vld1_u8 从内存中读取88位数据到寄存器
vld1q_u8 从内存中读取16 8位数据到寄存器
vld3q_u8 从内存中读取3个168位数据到寄存器中
vst3q_u8 将三个128位寄存器的数据写到内存中
vld4_u8 从内存中读取4个8 8位数据到寄存器中
vmull_u8 执行两个8*8位无符号整数的乘法操作
vshrn_n_u16 16位无符号整数右移指定的位数
vst1_u8 将128位寄存器中的8位无符号整数元素存储到内存中
vshrq_n_s16 16位整数右移指定的位数
举例
c
void bgr_to_rgb(uint8_t *bgr, uint8_t *rgb, int width, int height)
{
// Ensure BGR and BGR buffers are 16-byte aligned for NEON
uint8_t *bgr_aligned = (uint8_t *)(((uintptr_t)bgr + 15) & ~15);
uint8_t *rgb_aligned = (uint8_t *)(((uintptr_t)rgb + 15) & ~15);
for (int q = 0; q < height * width / 16; q++)
{
// Calculate the index for the current pixel
int index = q * 16 * 3;
// Load 16 BGR pixels into three vectors.
uint8x16x3_t bgr_vector = vld3q_u8(bgr_aligned + index);
// Shuffle the bytes to convert from BGR to BGR.
uint8x16_t b = bgr_vector.val[2]; // Blue
uint8x16_t g = bgr_vector.val[1]; // Green
uint8x16_t r = bgr_vector.val[0]; // Red
// Combine the shuffled bytes into a single vector.
uint8x16x3_t rgb_vector = {b, g, r};
// Store the result.
vst3q_u8(rgb_aligned + index, rgb_vector);
}
}
使用gstreamer
使用gstremaer pipeline技术写好插件,直接操作显存