Microsoft introduced a new remote desktop (RDP) technology called RemoteFX in Windows 7 SP1 and Windows Server 2008 R2 SP1. Remote FX shifts most of the graphics processing to server from client. As such, thin clients can display 3D images and stunning visual details that were not possible before. Since most of the graphics processing takes place on server side, there are some requirement on the server hardware. For details, please bing "RemoteFX". For brevity, I'll use RFX for RemoteFX in the rest of this document.
To achieve the graphics performance offered by RFX, a new codec is introduced for RDP. The official specification for this codec is named “Remote Desktop Protocol: RemoteFX Codec Extension” and will be mentioned as [MSRDPRFX] in the rest of this blog.
RFX codec uses modified version of Adaptive RunLength GolombRice (ARLGR) encoding to perform lossless compression on graphic data that is sent to a client. Although RLGR encoding compresses data losslessly, RFX codec performs a lossy compression on graphic data. The lossy part of compression is performed at quantization stage. The details of encoding and decoding are documented in [MSRDPRFX]. This blogs focuses on RLGR decoding and intends to provide test data for developers implementing RLGR decoding on the client side.
TS_RFX_TILESSET message contains number of tiles. Each tile contains RLGR encoded data for Y, Cb and Cr channels. The first step is to RLGR decode the three channels which results in three vectors of 4096 (64x64) elements each. We will decode only Y channel in this blog. Decoding for Cb and Cr is exactly the same. We will start from the data for Y channel as received in a tile of TS_RFX_TILESET message and will perform the RLGR3 decoding. After the decoding as described in this document, data is ready for “Subband Reconstruction” as described in section 3.1.8.2 of [MSRDPRFX].
I have assumed that the reader of this blog has read and somewhat understood section 2 and 3 of [ARLGR]. Please keep a copy handy while reading this blog. This blog’s mission is to explain the algorithm presented in section 3.1.8.1.7.3.1 of MSRDPRFX and I’ll use the terminology presented there without explaining it. So, it will definitely help to keep a browser window open pointed to [MSRDPRFX].
Here is the data for Y channel as received in one of the tiles of TS_RFX_TILESET message from an actual Windows 7 SP1 virtual machine running under hypervisor on Windows 2008 R2 SP1 server:
0x06, 0x20, 0xda, 0x17, 0x42, 0xe8, 0xfa, 0x00, 0x1f, 0xfc, 0x80, 0x64, 0x06, 0x40, 0xc8, 0x32,
0x0c, 0x86, 0x46, 0x4c, 0x99, 0x67, 0xc5, 0xf8, 0xba, 0x5d, 0x2e, 0x96, 0x4b, 0x45, 0x00, 0x00,
0x01, 0xba, 0x44, 0x03, 0x30, 0xe8, 0x80, 0xcc, 0xe8, 0x83, 0x97, 0x80, 0x39, 0x88, 0x02, 0x01,
0x01, 0x04, 0x10, 0x8f, 0x2f, 0x29, 0x1f, 0xff, 0xff, 0xff, 0x9f, 0x73, 0xee, 0x7d, 0xcf, 0xb9,
0xf7, 0x3e, 0x17, 0x80, 0x00, 0x02, 0x00, 0x03, 0x67, 0x8c, 0x7a, 0xa7, 0x0a, 0x91, 0xc0, 0x70,
0x1c, 0x1c, 0x38, 0xe2, 0xff, 0xfc, 0x0f, 0xec, 0xdf, 0x33, 0x7c, 0x4d, 0x89, 0x00, 0x00, 0x2f,
0x07, 0x86, 0x10, 0x96, 0x90, 0x8b, 0xcf, 0xf2, 0x8f, 0xa1, 0x64, 0xb8, 0xc7, 0x81, 0x00, 0x8c,
0x30, 0x03, 0x10, 0x00
The size of this block is 116 bytes. The data is always a multiple of 4 bytes or in other words contains complete 32 bit blocks. This is due to the fact that at the time of encoding, data is padded with zeros to make output a multiple of 32 bits. This has some implication at the time of decoding as we will see later.
After RLGR3 decoding, the output vector has 4096 elements. Most of the elements are zero, so I am reproducing here the elements that are not zero. Please note that index starts from 0. These elements are called coefficients in {MSRDPRFX] and I'll use that for these elements for the remainder of this blog.
[0024]= 3 
[0056]= 3 
[0088]= 3 
[0120]= 3 
[0152]= 3 
[1152]= 3 
[1153]= 3 
[1154]= 3 
[1155]= 3 
[1156]= 3 
[1157]= 3 
[1158]= 3 
[1159]= 3 
[1160]= 3 
[1161]= 3 
[1162]= 3 
[1163]= 3 
[1164]= 3 
[1165]= 3 
[1166]= 3 
[1167]= 3 
[1168]= 3 
[1169]= 3 
[1170]= 3 
[1171]= 3 
[1172]= 3 
[1173]= 3 
[1174]= 3 
[1175]= 3 
[1176]= 4 
[3083]= 1 
[3084]= 6 
[3099]= 1 
[3100]= 6 
[3115]= 1 
[3116]= 5 
[3132]= 1 
[3344]= 1 
[3345]= 1 
[3346]= 1 
[3347]= 1 
[3348]= 1 
[3349]= 1 
[3350]= 1 
[3351]= 1 
[3352]= 1 
[3353]= 1 
[3354]= 1 
[3355]= 1 
[3356]= 1 
[3360]= 8 
[3361]= 8 
[3362]= 8 
[3363]= 8 
[3364]= 8 
[3365]= 8 
[3366]= 8 
[3367]= 8 
[3368]= 8 
[3369]= 8 
[3370]= 8 
[3371]= 8 
[3372]= 8 
[3373]= 1 
[3628]= 1 
[3846]=18 
[3854]=11 
[3862]= 2 
[3904]= 4 
[3905]= 4 
[3906]= 4 
[3907]= 4 
[3908]= 4 
[3909]= 4 
[3910]= 3 
[3911]= 1 
[3912]=14 
[3913]=14 
[3914]=14 
[3915]=14 
[3916]=14 
[3917]=14 
[3918]=12 
[3919]= 2 
[3974]= 1 
[3982]= 4 
[4032]= 6 
[4037]= 1 
[4038]= 8 
[4039]=59 
[4040]= 45 
[4046]= 5 
[4047]=36 
[4048]= 2 
[4054]= 1 
[4055]= 7 
[4056]= 1 





The algorithm to encode/decode is a modified form of [ARLGR]. There is one very important difference that is captured in the pseudo code for encoding and decoding algorithm but is not explicitly mentioned in [MSRDPRFX].
In run mode, the input to GR(u,kr) is signed numbers i.e. no folding and interleaving mapping is performed. On the other hand, the input to “no run” mode is folding and interleaving mapped. Folding and interleaving mapping is defined in [ARLGR].
In the following, I’ll present the steps taken to decode the data I presented earlier in this blog. At the encoding stage, the output of encoding starts from the MSB (most significant bit) of the byte. So, at the time of decoding, we start from the MSB of the first byte in the data. I’ll decode in chunks of 4 bytes at a time and refill from the input data as needed.
The first four bytes are 0x06, 0x20, 0xda, 0x17 which in binary would be 00000110001000001101101000010111.
The initial values for decoding parameters are:
k=1 kp=8 kr=1 krp=8
Since k=1, we are in run mode, we remove first bit from the input which is a zero, the output is two zeros, based on the fact that 2^{k} is 2 so we produce two coefficients each equal to zero (see Table 2 in [ARLGR]).
[0000] = 0, [0001]=0
Next we increment kp=kp+4=12, then we right shift kp, LSGR times (3) to get new value of k, which is 1. We continue in this fashion till we hit the first nonzero bit in the input. Here is the output before we hit the first nonzero bit.
Input now = 0000110001000001101101000010111
Output = two coefficients each equal to zero ([0002]=0, [0003]=0)
k=2 kp=16
Remove one bit
Input now = 000110001000001101101000010111
Output = 4 coefficients each equal to zero ([0004]=0, [0005]=0, [0006]=0, [0007]=0)
k=2 kp=20
Remove one bit
Input now = 00110001000001101101000010111
Output = 4 coefficients each equal to zero ([0008]=0, [0009]=0, [0010]=0, [0011]=0)
k=3 kp=24
Remove one bit
Input now = 0110001000001101101000010111
Output = 8 coefficients each equal to zero ([0012]=0, [0013]=0, [0014]=0, [0015]=0, [0016]=0, [0017]=0, [0018]=0, [0019]=0)
k=3 kp=28
Remove one bit
Input now = 110001000001101101000010111
Now we have a nonzero bit and as we are in run mode, this is of the form 1 + bin(m, k) + GR(u, kR) (see [ARLGR], Table 2) which means that the first 1 is to mark partial run, the next three bits (since k=3) tell us how many zeros to output and they are 100 so we output four coefficients of zero each ([0020]=0, [0021]=0, [0022]=0, [0023]=0). So far the number of coefficients produced is 4+8+4+4+2+2=24. The twenty fifth coefficient is encoded with GR(u, kR). As I explained earlier, this GR is different from what is described in [ARLGR]. The first bit is the sign bit. If it is zero, the coefficient is a positive integer, otherwise negative. Since in this case first bit is 0, the coefficient is positive. The bits that follow represent the GR code of the (magnitude1) of the coefficient. So the number of 1’s before we hit a zero is equal to u/2^{kr}. Since there is only one 1, u is equal to 1X2 = 2 because kr=1. The next zero in the input is the delimiter 0 to mark the end of 1’s. The next kr bits represent remainder. So the remainder is 0. That means that (magnitude1) is 2+0=2 and magnitude is +3. This is the 25^{th} coefficient that we produced ([0024=3). The number of bit from input used=4+4=8.
Since we decoded a GR(u,kr), we need to decrement kp and adjust krp based on number of 1’s we decoded (p). Since p = 1 (see [ARLGR], Table 3), no change in kr and krp. Kp is reduced by DN_GR (6) and becomes 22 and k=kp>>LSGR =22>>3=2. After removing 8 bits from input:
Input now = 0001101101000010111
We will continue in “run mode” in similar fashion. When we produce a series of nonzero coefficients without having zero coefficients in “run mode”, it results in decrement of kp and eventually results in k becoming zero. This happens when coefficients from index 1152 to 1162 are produced. All these coefficients are nonzero (3) and resulted in k going all the way to zero. This changes the mode of operation from “run mode” to “no run” or GR mode. When we enter GR mode we have the input as
Input now = 1111100010111111000
k=0 kp=4 kr=1 krp=8
So looking at the input, we have five 1’s. So p=5. Since kr=1, u=p*2^{kr} =5*2=10. The remainder is zero (since next bit is zero and kr=1). We remove 7 bits from the input.
Input now= 010111111000
The decoding we are performing here is RLGR3. In RLGR3, GR mode operates a little different. For details please see section 3.1.8.1.7.2 of [MSRDPRFX]. So, 10 is sum of two coefficients. The first coefficient is represented in the least number of bits needed to represent the sum (10).
Now log_{2} 10 > 3 so the least number of bits needed to represent next coefficient is 4. The next 4 bits are 0101. That is 5 in decimal. The second coefficient is 105=5. These two numbers are folding and interleaving mapped. So the actual coefficients are 2y1=5 => 2y=6 => y=3. So, both are equal to 3 ([1163]=3, [1164]=3). We remove 4 bits from the input.
Input now = 11111000
The adaption parameters are adjusted differently in GR mode. Kr and krp are adjusted for the sum. K and kp are adjusted based on the two coefficients produced. Since p=5, krp = krp+5= 8+5=13. Kr=krp>>LSGR=13>>3=1. Kp (and subsequently k) is adjusted as per the pseudo code in section 3.1.8.1.7.3.2 of [MSRDPRFX]. When we entered Gr mode, k=0 and kp=4. The coefficients produced are nonzero so we decrement kp by 2*DQ_GR. 2*DQ_GR is 6 and kp=46=2. But the minimum kp can go zero, so kp=0 and therefore k=0.
So far we have covered all the scenarios of RLGR3 decoding. The only thing left is to discuss a little bit how the end of the input is processed.
Everything is dealt as explained earlier but algorithm stops when all 4096 coefficients are generated. As explained earlier, the data is padded to make number of bits a multiple of 32. So it is not necessary to exhaust all the bits in the input.

References
[MSRDPRFX]: Remote Desktop Protocol: RemoteFX Codec Extension
http://msdn.microsoft.com/enus/library/ff635423(PROT.10).aspx
[ARLGR] Malvar, H.S., "Adaptive RunLength / GolombRice Encoding of Quantized Generalized Gaussian Sources with Unknown Statistics", Proceedings of the Data Compression Conference, 2006 (DCC 2006) pp. 23  32, March 2006,
http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=01607237
Hiya,
where can I find out more details about how RemoteFX works. I'm particularly interested in how & why MS made a decisions to assign x GPU ram per monitor. Let us say I want to assign only two VDI users with RFX and I want them to have 512mb GPU ram each on a 1024+mb card? Can't do it today so have to go physical. Would love to see more in depth content.
I read that the FRX encoding is done on the GPU, so the 1024mb card would need to leave some ram for encoding.