Skip to content

Commit

Permalink
* we don't use dynamic resizing, so turn it off
Browse files Browse the repository at this point in the history 
* update unused bitrate to more realistic values
* split compress_image into 3 more re-usable methods

git-svn-id: https://xpra.org/svn/Xpra/trunk@15770 3bb7dfac-3a0b-4e04-842a-767bc560f471
  • Loading branch information
@totaam
totaam committed Apr 30, 2017
1 parent 691765f commit a134cea
Showing 1 changed file with 51 additions and 37 deletions.
88 changes: 51 additions & 37 deletions src/xpra/codecs/nvenc7/encoder.pyx
View file
Expand Up @@ -1459,6 +1459,7 @@ cdef class Encoder:
self.update_bitrate()
cdef double start = monotonic_time()

#the pixel format we feed into the encoder
self.pixel_format = self.get_target_pixel_format(self.quality)
self.lossless = self.get_target_lossless(self.pixel_format, self.quality)

Expand Down Expand Up @@ -1654,8 +1655,8 @@ cdef class Encoder:
params.presetGUID = preset
params.encodeWidth = self.encoder_width
params.encodeHeight = self.encoder_height
params.maxEncodeWidth = self.encoder_width
params.maxEncodeHeight = self.encoder_height
params.maxEncodeWidth = 0
params.maxEncodeHeight = 0
params.darWidth = self.encoder_width
params.darHeight = self.encoder_height
params.enableEncodeAsync = 0 #not supported on Linux
Expand All @@ -1677,6 +1678,7 @@ cdef class Encoder:
qpmax = QP_MAX_VALUE-max(0, int(QP_MAX_VALUE*self.quality//100))
config.frameFieldMode = NV_ENC_PARAMS_FRAME_FIELD_MODE_FRAME
config.mvPrecision = NV_ENC_MV_PRECISION_QUARTER_PEL
#config.mvPrecision = NV_ENC_MV_PRECISION_FULL_PEL
if True:
#const QP:
config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CONSTQP
Expand All @@ -1685,9 +1687,9 @@ cdef class Encoder:
config.rcParams.constQP.qpInterB = qp
config.rcParams.constQP.qpIntra = qp
else:
config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR
config.rcParams.averageBitRate = 5000000
config.rcParams.maxBitRate = 10000000
config.rcParams.rateControlMode = NV_ENC_PARAMS_RC_CBR
config.rcParams.averageBitRate = 500000
config.rcParams.maxBitRate = 600000
config.rcParams.vbvBufferSize = 0
config.rcParams.vbvInitialDelay = 0
config.rcParams.enableInitialRCQP = 1
Expand All @@ -1696,8 +1698,12 @@ cdef class Encoder:
config.rcParams.initialRCQP.qpInterB = qpmin

if self.codec_name=="H264":
#3 for YUV444, 1 for NV12:
config.encodeCodecConfig.h264Config.chromaFormatIDC = 1+int(self.pixel_format=="YUV444P") * 2
if self.pixel_format=="NV12":
config.encodeCodecConfig.h264Config.chromaFormatIDC = 1
elif self.pixel_format=="YUV444P":
config.encodeCodecConfig.h264Config.chromaFormatIDC = 3
else:
raise Exception("unknown pixel format %s" % self.pixel_format)
config.encodeCodecConfig.h264Config.h264VUIParameters.colourDescriptionPresentFlag = 0
config.encodeCodecConfig.h264Config.h264VUIParameters.videoSignalTypePresentFlag = 0
config.encodeCodecConfig.h264Config.idrPeriod = config.gopLength
Expand Down Expand Up @@ -2050,23 +2056,12 @@ cdef class Encoder:
return self.convert_image(image, options, retry+1)

cdef do_compress_image(self, image, options={}):
cdef NV_ENC_PIC_PARAMS picParams #@DuplicatedSignature
cdef NV_ENC_MAP_INPUT_RESOURCE mapInputResource
cdef NV_ENC_LOCK_BITSTREAM lockOutputBuffer
cdef size_t size
cdef unsigned int x, y, stride
cdef unsigned int i
cdef NVENCSTATUS r #@DuplicatedSignature
cdef unsigned int stride, w, h
assert self.context, "context is not initialized"
cdef double start = monotonic_time()
log("compress_image(%s, %s)", image, options)
cdef unsigned int w = image.get_width()
cdef unsigned int h = image.get_height()
pixels = image.get_pixels()
cdef unsigned int image_stride = image.get_rowstride()
cdef unsigned long input_size = self.inputPitch * self.input_height
assert pixels is not None, "failed to get pixels from %s" % image
assert self.context!=NULL, "context is not initialized"
w = image.get_width()
h = image.get_height()
assert image.get_planes()==ImageWrapper.PACKED, "invalid number of planes: %s" % image.get_planes()
assert (w & WIDTH_MASK)<=self.input_width, "invalid width: %s" % w
assert (h & HEIGHT_MASK)<=self.input_height, "invalid height: %s" % h
Expand All @@ -2076,35 +2071,47 @@ cdef class Encoder:
#first frame, record pts:
self.first_frame_timestamp = image.get_timestamp()

cdef unsigned long input_size = self.inputPitch * self.input_height
stride = self.copy_image(image, self.inputBuffer, self.inputPitch)
driver.memcpy_htod(self.cudaInputBuffer, self.inputBuffer)
log("compress_image(..) uploaded %i bytes of input buffer to device", input_size)
self.exec_kernel(w, h, stride)
self.bytes_in += input_size
return self.nvenc_compress(input_size, image.get_timestamp())

cdef copy_image(self, image, target_buffer, unsigned int target_stride):
cdef unsigned int image_stride = image.get_rowstride()
cdef unsigned int h = image.get_height()
cdef unsigned int i, stride
pixels = image.get_pixels()
assert pixels is not None, "failed to get pixels from %s" % image
#copy to input buffer:
cdef object buf
if isinstance(pixels, memoryview):
#copy memoryview to inputBuffer directly:
buf = self.inputBuffer
buf = target_buffer
else:
#this is a numpy.ndarray type:
buf = self.inputBuffer.data
buf = target_buffer.data
cdef unsigned long pix_len = len(pixels)
if image_stride<=self.inputPitch:
if image_stride<=target_stride:
stride = image_stride
assert pix_len<=input_size, "too many pixels (expected %s max, got %s) image: %sx%s stride=%s, input buffer: stride=%s, height=%s" % (input_size, pix_len, w, h, stride, self.inputPitch, self.input_height)
log("copying %s pixels from %s into %s, in one shot", pix_len, type(pixels), type(self.inputBuffer))
#assert pix_len<=input_size, "too many pixels (expected %s max, got %s) image: %sx%s stride=%s, input buffer: stride=%s, height=%s" % (input_size, pix_len, w, h, stride, self.inputPitch, self.input_height)
log("copying %s pixels from %s into %s, in one shot", pix_len, type(pixels), type(target_buffer))
buf[:pix_len] = pixels
else:
#ouch, we need to copy the source pixels into the smaller buffer
#before uploading to the device... this is probably costly!
stride = self.inputPitch
log("copying %s pixels from %s into %s, %i stride at a time", stride*h, type(pixels), type(self.inputBuffer), stride)
stride = target_stride
log("copying %s pixels from %s into %s, %i stride at a time", stride*h, type(pixels), type(target_buffer), stride)
for i in range(h):
x = i*stride
y = i*image_stride
buf[x:x+stride] = pixels[y:y+stride]
return stride


#copy input buffer to CUDA buffer:
driver.memcpy_htod(self.cudaInputBuffer, self.inputBuffer)
self.bytes_in += input_size
log("compress_image(..) %i bytes of input buffer copied to device", input_size)

cdef exec_kernel(self, unsigned int w, unsigned int h, unsigned int stride):
cdef uint8_t dx, dy
if self.pixel_format=="NV12":
#(these values are derived from the kernel code - which we should know nothing about here..)
Expand Down Expand Up @@ -2144,6 +2151,13 @@ cdef class Encoder:
cdef double csc_end = monotonic_time()
log("compress_image(..) kernel %s took %.1f ms", self.kernel_name, (csc_end - csc_start)*1000.0)

cdef nvenc_compress(self, int input_size, timestamp=0):
cdef NVENCSTATUS r #@DuplicatedSignature
cdef NV_ENC_PIC_PARAMS picParams #@DuplicatedSignature
cdef NV_ENC_MAP_INPUT_RESOURCE mapInputResource
cdef NV_ENC_LOCK_BITSTREAM lockOutputBuffer
cdef double start = monotonic_time()
assert input_size>0, "invalid input size %i" % input_size
#map buffer so nvenc can access it:
memset(&mapInputResource, 0, sizeof(NV_ENC_MAP_INPUT_RESOURCE))
mapInputResource.version = NV_ENC_MAP_INPUT_RESOURCE_VER
Expand Down Expand Up @@ -2184,7 +2198,7 @@ cdef class Encoder:
#picParams.codecPicParams.h264PicParams.sliceMode = 3 #sliceModeData specifies the number of slices
#picParams.codecPicParams.h264PicParams.sliceModeData = 1 #1 slice!
picParams.frameIdx = self.frames
picParams.inputTimeStamp = image.get_timestamp()-self.first_frame_timestamp
picParams.inputTimeStamp = timestamp-self.first_frame_timestamp
#inputDuration = 0 #FIXME: use frame delay?
#picParams.rcParams.rateControlMode = NV_ENC_PARAMS_RC_VBR #FIXME: check NV_ENC_CAPS_SUPPORTED_RATECONTROL_MODES caps
#picParams.rcParams.enableMinQP = 1
Expand All @@ -2207,7 +2221,7 @@ cdef class Encoder:
r = self.functionList.nvEncEncodePicture(self.context, &picParams)
raiseNVENC(r, "error during picture encoding")
encode_end = monotonic_time()
log("compress_image(..) encoded in %.1f ms", (encode_end-csc_end)*1000.0)
log("compress_image(..) encoded in %.1f ms", (encode_end-start)*1000.0)

#lock output buffer:
lockOutputBuffer.version = NV_ENC_LOCK_BITSTREAM_VER
Expand Down Expand Up @@ -2244,7 +2258,7 @@ cdef class Encoder:

client_options = {
"frame" : int(self.frames),
"pts" : int(image.get_timestamp()-self.first_frame_timestamp),
"pts" : int(timestamp-self.first_frame_timestamp),
"speed" : self.speed,
}
if self.lossless:
Expand Down Expand Up @@ -2621,7 +2635,7 @@ def init_module():
try:
try:
test_encoder = Encoder()
test_encoder.init_context(1920, 1080, src_format, dst_formats, encoding, 50, 50, (1,1), options)
test_encoder.init_context(1920, 1080, src_format, dst_formats, encoding, 50, 100, (1,1), options)
success = True
if client_key:
log("the license key '%s' is valid", client_key)
Expand Down

0 comments on commit a134cea

Please sign in to comment.