[英]How to do fast memory copy from YUV_420_888 to NV21
参考这篇文章,我想写一个将 android YUV_420_888转换为 nv21 的方法。 虽然来自 camera2 API 的图像是伪装的默认 NV21,但需要更通用的实现。 它是这样的:
class NV21Image{
public byte[] y;
public byte[] uv;
}
public static void cvtYUV420ToNV21(Image image, NV21Image nv21) {
int width = image.getWidth();
int height = image.getHeight();
int ySize = width*height;
ByteBuffer yBuffer = image.getPlanes()[0].getBuffer(); // Y
ByteBuffer uBuffer = image.getPlanes()[1].getBuffer(); // U
ByteBuffer vBuffer = image.getPlanes()[2].getBuffer(); // V
int yRowStride = image.getPlanes()[0].getRowStride();
int vRowStride = image.getPlanes()[2].getRowStride();
int pixelStride = image.getPlanes()[2].getPixelStride();
assert(image.getPlanes()[0].getPixelStride() == 1);
assert(image.getPlanes()[2].getRowStride() == image.getPlanes()[1].getRowStride());
assert(image.getPlanes()[2].getPixelStride() == image.getPlanes()[1].getPixelStride());
int pos = 0;
int yBufferPos = -yRowStride; // not an actual position
for (; pos<ySize; pos+=width) {
yBufferPos += yRowStride;
yBuffer.position(yBufferPos);
yBuffer.get(nv21.y, pos, width);
}
pos = 0;
for (int row=0; row<height/2; row++) {
for (int col=0; col<vRowStride / pixelStride; col++) {
int vuPos = col*pixelStride + row * vRowStride;
nv21.uv[pos++] = vBuffer.get(vuPos);
nv21.uv[pos++] = uBuffer.get(vuPos);
}
}
}
以上代码按预期工作,但对于我的实时相机预览应用程序来说非常耗时(在 Snapdragon 865 CPU 中每帧 720p 大约 12 毫秒),因此我尝试使用 JNI 实现来加速它以利用字节访问和性能优势:
JNIEXPORT void JNICALL
Java_com_example_Utils_nFillYUVArray(JNIEnv *env, jclass clazz, jbyteArray yArr, jbyteArray uvArr,
jobject yBuf, jobject uBuf, jobject vBuf,
jint yRowStride, jint vRowStride, jint vPixelStride, jint w, jint h) {
auto ySrcPtr = (jbyte const*)env->GetDirectBufferAddress(yBuf);
auto uSrcPtr = (jbyte const*)env->GetDirectBufferAddress(uBuf);
auto vSrcPtr = (jbyte const*)env->GetDirectBufferAddress(vBuf);
for(int row = 0; row < h; row++){
env->SetByteArrayRegion(yArr, row * w, w, ySrcPtr + row * yRowStride);
}
int pos = 0;
for (int row=0; row<h/2; row++) {
for (int col=0; col<w/2; col++) {
int vuPos = col * vPixelStride + row * vRowStride;
env->SetByteArrayRegion(uvArr, pos++, 1, vSrcPtr + vuPos);
env->SetByteArrayRegion(uvArr, pos++, 1, uSrcPtr + vuPos);
}
}
}
然而,它变得比我预期的更糟(每帧大约 107 毫秒)。 而最耗时的部分是隔行扫描 memory copy for UV buffer
所以我的问题是是否有任何加速方法以及如何解决?
更新
当 U、V 平面的pixelStride
均为 1 或 2 时,我成功地加速了它(检查我的答案),我相信在大多数情况下都会发生这种情况。
正如@snachmsm 所说,libyuv 可能会有所帮助。 我找到一个可用的 API I420ToNV21
,但是它不能接收pixelStride参数,因为YUV_420_888
不保证 U、V 平面中相邻像素之间不存在间隙。
当 pixelStride 为 2(减少到每帧2.7 毫秒)时,我使用 arm 内在函数成功加速了它:
JNIEXPORT void JNICALL
Java_com_example_Utils_nFillYUVArray(JNIEnv *env, jclass clazz, jbyteArray yArr, jbyteArray uvArr,
jobject yBuf, jobject uBuf, jobject vBuf,
jint yRowStride, jint vRowStride, jint uRowStride, jint pixelStride,
jint width, jint height) {
///TODO: too time-consuming
auto ySrcPtr = (jbyte const*)env->GetDirectBufferAddress(yBuf);
auto uSrcPtr = (jbyte const*)env->GetDirectBufferAddress(uBuf);
auto vSrcPtr = (jbyte const*)env->GetDirectBufferAddress(vBuf);
for(int row = 0; row < height; row++){
env->SetByteArrayRegion(yArr, row * width, width, ySrcPtr + row * yRowStride);
}
constexpr int kStride = 8;
const size_t nGroups = width / kStride;
if(pixelStride == 2){
int8_t *line = (int8_t*)alignedAlloc(width, 64);
int8_t *mask = (int8_t*)alignedAlloc(kStride, 64);
memset(mask, 0, kStride);
for(int i=0; i < kStride / 2; i++) {
mask[i * 2] = -1;
}
int8x8_t vm = vld1_s8(mask);
for(int row = 0; row < height / 2; row ++){
size_t vrowOff = row * vRowStride;
size_t urowOff = row * uRowStride;
for(int g = 0; g < nGroups; g++) {
size_t colOff = g * kStride;
int8x8_t v0 = vld1_s8(vSrcPtr + vrowOff + colOff);
int8x8_t v1 = vld1_s8(uSrcPtr + urowOff + colOff);
int8x8_t a0 = vand_s8(v0, vm);
int16x4_t b1 = vreinterpret_s16_s8(vand_s8(v1, vm));
int8x8_t a1 = vreinterpret_s8_s16(vshl_n_s16(b1, 8));
int8x8_t r = vorr_s8(a0, a1);
vst1_s8(line + colOff, r);
}
env->SetByteArrayRegion(uvArr, row * width, width, line);
}
free(mask);
free(line);
}else if(pixelStride == 1){
int8_t *line = (int8_t*)alignedAlloc(width, 64);
for(int row = 0; row < height / 2; row ++) {
size_t vrowOff = row * vRowStride;
size_t urowOff = row * uRowStride;
for(int g = 0; g < nGroups / 2; g++){
size_t colOff = g * kStride;
int8x8_t v0 = vld1_s8(vSrcPtr + vrowOff + colOff);
int8x8_t v1 = vld1_s8(uSrcPtr + urowOff + colOff);
int8x8x2_t vz = vzip_s8(v0, v1);
vst1_s8(line + colOff, vz.val[0]);
vst1_s8(line + colOff + kStride, vz.val[1]);
}
env->SetByteArrayRegion(uvArr, row * width, width, line);
}
free(line);
}
}
pixelStride == 1
的情况未经过充分测试,但我相信它会按预期工作。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.