Related
I'm new to OpenCL and I'm working on converting an existing algorithm to OpenCL.
In this process, I am experiencing a phenomenon that I cannot solve on my own, and I would like to ask some help.
Here's details.
My kernel is applied to images of different size (to be precise, each layer of the Laplacian pyramid).
I get normal results for images of larger size such as 3072 x 3072, 1536 x 1536.
But I get abnormal results for smaller images such as 12 x 12, 6 x 6, 3 x 3, 2 x 2.
At first, I suspected that clEnqueueNDRangeKernel had a bottom limit for dimensions, causing this problem. So, I added printf to the beginning of the kernel as follows. It is confirmed that all necessary kernel instances are executed.
__kernel void GetValueOfB(/* parameters */)
{
uint xB = get_global_id(0);
uint yB = get_global_id(1);
printf("(%d, %d)\n", xB, yB);
// calculation code is omitted
}
So after wandering for a while, I added the same printf to the end of the kernel. When I did this, it was confirmed that printf works only for some pixel positions. For pixel positions not output by printf, the calculated values in the resulting image are incorrect, and as a result, I concluded that some kernel instances terminate abnormally before completing the calculations.
__kernel void GetValueOfB(/* parameters */)
{
uint xB = get_global_id(0);
uint yB = get_global_id(1);
printf("(%d, %d)\n", xB, yB);
// calculation code is omitted
printf("(%d, %d, %f)\n", xB, yB, result_for_this_position);
}
It seems that there is no problem with the calculation of the kernel. If I compile the kernel turning off the optimization with the -cl-opt-disable option, I get perfectly correct results for all images regardless of their size. In addition to that, with NVIDA P4000, it works correct. Of course, in theses cases, I confirmed that the printf added at the bottom of the Kernel works for all pixels.
Below I put additional information and attach a part of the code I wrote.
Any advice is welcomed and appreciated.
Thank you.
SDK: Intel® SDK For OpenCL™ Applications 2020.3.494
Platform: Intel(R) OpenCL HD Graphics
for all images
{
...
const size_t globalSize[2] = { size_t(vtMatB_GPU_LLP[nLayerIndex].cols), size_t(vtMatB_GPU_LLP[nLayerIndex].rows) };
err = clEnqueueNDRangeKernel(_pOpenCLManager->GetCommandQueue(), kernel, 2,
NULL, globalSize, NULL, 0, NULL, NULL);
if (CL_SUCCESS != err)
return -1;
// I tried with this but it didn't make any difference
//std::this_thread::sleep_for(std::chrono::seconds(1));
err = clFinish(_pOpenCLManager->GetCommandQueue());
if (CL_SUCCESS != err)
return -1;
err = clEnqueueReadBuffer(_pOpenCLManager->GetCommandQueue(), memMatB, CL_TRUE,
0, sizeof(float) * vtMatB_GPU_LLP[nLayerIndex].cols *
vtMatB_GPU_LLP[nLayerIndex].rows, vtMatB_GPU_LLP[nLayerIndex].data, 0, nullptr, nullptr);
if (CL_SUCCESS != err)
return -1;
...
}
And I tried with event, too, but it works the same way.
for all images
{
...
const size_t globalSize[2] = { size_t(vtMatB_GPU_LLP[nLayerIndex].cols), size_t(vtMatB_GPU_LLP[nLayerIndex].rows) };
cl_event event;
err = clEnqueueNDRangeKernel(_pOpenCLManager->GetCommandQueue(), kernel, 2, NULL, globalSize, NULL, 0, NULL, &event);
if (CL_SUCCESS != err)
return -1;
err = clWaitForEvents(1, &event);
if (CL_SUCCESS != err)
return -1;
err = clFinish(_pOpenCLManager->GetCommandQueue());
if (CL_SUCCESS != err)
return -1;
err = clEnqueueReadBuffer(_pOpenCLManager->GetCommandQueue(), memMatB, CL_TRUE,
0, sizeof(float) * vtMatB_GPU_LLP[nLayerIndex].cols *
vtMatB_GPU_LLP[nLayerIndex].rows, vtMatB_GPU_LLP[nLayerIndex].data, 0, nullptr, nullptr);
if (CL_SUCCESS != err)
return -1;
...
}
/////// Added contents ////////////////////////////////////////////
Would you guys please take look at this issue in the aspect of clFinsh, or clWaitEvent. Am I missing something in this regard?
Sometimes I get less correct values and sometimes I get more correct values.
To be more specific, let's say I'm applying the kernel to 12 x 12 size image. So there're 144 pixel values.
Sometime I get correct values for 56 pixels.
Sometime I get correct values for 89 pixels.
Some other time I get correct value for n(less then 144) pixels.
If I turn off the OpenCL optimization when compiling the kernel by specifying -cl-opt-disable option, I get correct values for all 144 pixels.
The other thing that makes me think the calculation code is correct is that the same OpenCL code with no modification(other then device select code) runs perfectly correctly with NVIDIA P4000.
At first, I was really suspicious about the calculation code, but more I inspect code, more I'm confident there's nothing wrong with calculation code.
I know there's still a chance that there is an error in the calculation code so that there happen some exceptions anywhere during calculations.
I have plain C++ code for same task. I'm comparing results from those two.
/////// Another added contents ////////////////////////////////////////////
I made a minimum code(except projects template) to reproduce the phenomenon.
What's odd more is that if I install "Intel® Distribution for GDB Target" I get correct results.
https://github.com/heysweetethan/GPUOpenCLProjectforWindows
OpenCL kernels run threads in parallel on a specified global range, which in your case is the image size, with one thread per pixel.
The threads are grouped in workgroups, Workgroup size should be a multiple of 32; ideally 64 to make full use of the hardware, or 8x8 pixels in 2D. These workgroups cannot be split, so the global range must be a multiple of workgroup size.
What happens if global range is not clearly divisible by workgroup size, or smaller than workgroup size, like 3x3 pixels? Then the last workgroup is still executed with all 8x8 threads. The first 3x3 work on valid data in memory, but all the other threads read/write unallocated memory. This can cause undefined behavior or even crashes.
If you cannot have global size as a multiple of workgroup size, there is still a solution: a guard clause in the very beginning of the kernel:
if(xB>=xImage||yB>=yImage) return;
This ensures that no threads access unallocated memory.
As you don't supply a complete reproducible code sample, here's a loose collection of comments/suggestions/advice:
1. printf in kernel code
Don't rely on large amounts of printf output from kernels. It's necessarily buffered, and some implementations don't guarantee delivery of messages - often there's a fixed size buffer and when that's full, messages are dropped.
Note that your post-calculation printf increases the total amount of output, for example.
The reliable way to check or print kernel output is to write it to a global buffer and print it in host code. For example, if you want to verify each work-item reaches a specific point in the code, consider creating a zero-initialised global buffer where you can set a flag in each work-item.
2. Events
As you asked about events, flushing, etc. Your clFinish call certainly should suffice to ensure everything has executed - if anything, it's overkill, but especially while you're debugging other issues it's a good way to rule out queuing issue.
The clWaitForEvents() call preceeding it is not a great idea, as you haven't called clFlush() after queueing the kernel whose event you're waiting for. It's fairly minor, but could be a problem on some implementations.
3. Small image sizes
You've not actually posted any of the code that deals with the images themselves, so I can only guess at potential issues there. It looks like you're not using workgroups, so you shouldn't be running into the usual multiple-of-group-size pitfall.
However, are you sure you're loading the source data correctly, and you're correctly indexing into it? There could be all sorts of pitfalls here, from alignment of pixel rows in the source data, enqueueing the kernel before filling the source buffers has completed, creating source buffers with the wrong flags, etc.
So in summary, I'd suggest:
Don't believe in-kernel-printf if something strange is going on. Switch to something more reliable for observing the behaviour of your kernel code.
At minimum, post all your OpenCL API calling host code. Buffer creation, setting arguments, etc. Any fragments of kernel code accessing the buffers are probably not a bad idea either.
Thanks to a person from intel community, I could understand the phenomenon.
Briefly, if you spend to much time on a single kernel instance, 'Timeout Detection and Recovery(TDR)' stops the kernel instance.
For more information about this, you could refer to the followings.
https://learn.microsoft.com/en-us/windows-hardware/drivers/display/tdr-registry-keys
https://www.pugetsystems.com/labs/hpc/Working-around-TDR-in-Windows-for-a-better-GPU-computing-experience-777/
https://community.intel.com/t5/GPU-Compute-Software/It-s-like-OpenCL-kernel-instance-ends-abruptly/m-p/1386883#M478
I appreciate for all the people who gave me advices.
I'm currently writing a cross platform library in C which includes gamepad support. Gamepad communication on windows is handled by both raw input and xinput, depending on the specific gamepad.
While xinput facilitates force feedback on xbox360 controllers, I have not found a way to do this using raw input. I have some gamepads that can give force feedback, but I cannot find a way to do this through raw input. Is there a way to do this?
I prefer not to use the directinput api, since it's deprecated and discouraged by Microsoft.
Edit:
Since I've implemented gamepads for a large part, maybe I can narrow down the question a bit. I suspect the amount of rumble motors in a gamepad can be found by reading the NumberOutputValueCaps of a HIDP_CAPS structure. This gives the correct result for all my test gamepads.
I'm using the funtcion HidP_GetUsageValue to read axis values, which works fine. Now I suspect calling HidP_SetUsageValue should allow me to change this output value, turning on the rumble motor. Calling this function fails, however. Should this function be able to access rumble motors?
HidP_SetUsageValue() only modifies a report buffer -- you need to first prepare an appropriately-sized buffer (which may be why the function was failing; input reports and output reports won't necessarily be the same size) then send it to the device before it will have any effect. MSDN suggests you can use HidD_SetOutputReport() for that purpose, but I had better luck with WriteFile(), following the sample code at: https://code.msdn.microsoft.com/windowshardware/HClient-HID-Sample-4ec99697/sourcecode?fileId=51262&pathId=340791466
This snippet (based on the Linux driver) lets me control the motors and LED on a DualShock 4:
const char *path = /* from GetRawInputDeviceInfo(RIDI_DEVICENAME) */;
HANDLE hid_device = CreateFile(path, GENERIC_READ | GENERIC_WRITE,
FILE_SHARE_READ | FILE_SHARE_WRITE, NULL,
OPEN_EXISTING, 0, NULL);
assert(hid_device != INVALID_HANDLE_VALUE);
uint8_t buf[32];
memset(buf, 0, sizeof(buf));
buf[0] = 0x05;
buf[1] = 0xFF;
buf[4] = right_motor_strength; // 0-255
buf[5] = left_motor_strength; // 0-255
buf[6] = led_red_level; // 0-255
buf[7] = led_green_level; // 0-255
buf[8] = led_blue_level; // 0-255
DWORD bytes_written;
assert(WriteFile(hid_device, buf, sizeof(buf), &bytes_written, NULL));
assert(bytes_written == 32);
(EDIT: fixed buffer offsets)
Scatter-gather - readv()/writev()/preadv()/pwritev() - reads/writes a variable number of iovec structs in a single system call. Basically it reads/write each buffer sequentially from the 0th iovec to the Nth. However according to the documentation it can also return less on the readv/writev calls than was requested. I was wondering if there is a standard/best practice/elegant way to handle that situation.
If we are just handling a bunch of character buffers or similar this isn't a big deal. But one of the niceties is using scatter-gather for structs and/or discrete variables as the individual iovec items. How do you handle the situation where the readv/writev only reads/writes a portion of a struct or half of a long or something like that.
Below is some contrived code of what I am getting at:
int fd;
struct iovec iov[3];
long aLong = 74775767;
int aInt = 949;
char aBuff[100]; //filled from where ever
ssize_t bytesWritten = 0;
ssize_t bytesToWrite = 0;
iov[0].iov_base = &aLong;
iov[0].iov_len = sizeof(aLong);
bytesToWrite += iov[0].iov_len;
iov[1].iov_base = &aInt;
iov[1].iov_len = sizeof(aInt);
bytesToWrite += iov[1].iov_len;
iov[2].iov_base = &aBuff;
iov[2].iov_len = sizeof(aBuff);
bytesToWrite += iov[2].iov_len;
bytesWritten = writev(fd, iov, 3);
if (bytesWritten == -1)
{
//handle error
}
if (bytesWritten < bytesToWrite)
//how to gracefully continue?.........
Use a loop like the following to advance the partially-processed iov:
for (;;) {
written = writev(fd, iov+cur, count-cur);
if (written < 0) goto error;
while (cur < count && written >= iov[cur].iov_len)
written -= iov[cur++].iov_len;
if (cur == count) break;
iov[cur].iov_base = (char *)iov[cur].iov_base + written;
iov[cur].iov_len -= written;
}
Note that if you don't check for cur < count you will read past the end of iov which might contain zero.
AFAICS the vectored read/write functions work the same wrt short reads/writes as the normal ones. That is, you get back the number of bytes read/written, but this might well point into the middle of a struct, just like with read()/write(). There is no guarantee that the possible "interruption points" (for lack of a better term) coincide with the vector boundaries. So unfortunately the vectored IO functions offer no more help for dealing with short reads/writes than the normal IO functions. In fact, it's more complicated since you need to map the byte count into an IO vector element and offset within the element.
Also note that the idea of using vectored IO for individual structs or data items might not work that well; the max allowed value for the iovcnt argument (IOV_MAX) is usually quite small, something like 1024 or so. So if you data is contiguous in memory, just pass it as a single element rather than artificially splitting it up.
Vectored write will write all the data you have provided with one call to "writev" function. So byteswritten will be always be equal to total number of bytes provided as input. this is what my understanding is.
Please correct me if I am wrong
I have an SHA1 hash and I need to sign it. The CryptSignHash() method requires a HCRYPTHASH handle for signing. I create it and as I have the actual hash value already then set it:
CryptCreateHash(cryptoProvider, CALG_SHA1, 0, 0, &hash);
CryptSetHashParam(hash, HP_HASHVAL, hashBytes, 0);
The hashBytes is an array of 20 bytes.
However the problem is that the signature produced from this HCRYPTHASH handle is incorrect. I traced the problem down to the fact that CAPI actually doesn't use all 20 bytes from my hashBytes array. For some reason it thinks that SHA1 is only 4 bytes.
To verify this I wrote this small program:
HCRYPTPROV cryptoProvider;
CryptAcquireContext(&cryptoProvider, NULL, NULL, PROV_RSA_FULL, 0);
HCRYPTHASH hash;
HCRYPTKEY keyForHash;
CryptCreateHash(cryptoProvider, CALG_SHA1, keyForHash, 0, &hash);
DWORD hashLength;
CryptGetHashParam(hash, HP_HASHSIZE, NULL, &hashLength, 0);
printf("hashLength: %d\n", hashLength);
And this prints out hashLength: 4 !
Can anyone explain what I am doing wrong or why Microsoft CAPI thinks that SHA1 is 4 bytes (32 bits) instead of 20 bytes (160 bits).
There are a small error in your code. Instead of
DWORD hashLength;
CryptGetHashParam(hash, HP_HASHSIZE, NULL, &hashLength, 0);
printf("hashLength: %d\n", hashLength);
you should use
DWORD hashLength, hashSize;
hashLength = sizeof(DWORD)
CryptGetHashParam(hash, HP_HASHSIZE, (PBYTE)&hashSize, &hashLength, 0);
printf("hashSize: %d\n", hashSize);
then you will receive 20 as expected.
The Usage of CryptSignHash after CryptSetHashParam must also work. See remark at the end of the description of CryptSetHashParam function at http://msdn.microsoft.com/en-us/library/aa380270(VS.85).aspx. I suppose you just made the same error as with CryptGetHashParam(..., HP_HASHSIZE, ...) during retrieving of the result of signing. Compare your code with the code from the description of CryptSignHash function http://msdn.microsoft.com/en-us/library/aa380280(VS.85).aspx.
I don't think you can use CryptCreateHash in that way. From MSDN:
"The CryptCreateHash function
initiates the hashing of a stream of
data."
In other words, it looks like you can't instantiate a hash context in any way other than empty (and then by having it hash your input data).
How do you have the hash at present - a byte array? If so you probably just want to sign that array; I'd look into CryptSignMessage or CryptSignMessageWithKey as likely to do the job.
(I'm guessing, but what you're seeing may be explained by the output hash length not being set up until after the hash context has done some work.)
I have the very common problem of creating an index for an in-disk array of strings. In short, I need to store the position of each string in the in-disk representation. For example, a very naive solution would be an index array as follows:
uint64 idx[] = { 0, 20, 500, 1024, ..., 103434 };
Which says that the first string is at position 0, the second at position 20, the third at position 500 and the nth at position 103434.
The positions are always non-negative 64 bits integers in sequential order. Although the numbers could vary by any difference, in practice I expect the typical difference to be inside the range from 2^8 to 2^20. I expect this index to be mmap'ed in memory, and the positions will be accessed randomly (assume uniform distribution).
I was thinking about writing my own code for doing some sort of block delta encoding or other more sophisticated encoding, but there are so many different trade-offs between encoding/decoding speed and space that I would rather get a working library as a starting point and maybe even settle for something without any customizations.
Any hints? A c library would be ideal, but a c++ one would also allow me to run some initial benchmarks.
A few more details if you are still following. This will be used to build a library similar to cdb (http://cr.yp.to/cdb/cdbmake.html) on top the library cmph (http://cmph.sf.net). In short, it is for a large disk based read only associative map with a small index in memory.
Since it is a library, I don't have control over input, but the typical use case that I want to optimize have millions of hundreds of values, average value size in the few kilobytes ranges and maximum value at 2^31.
For the record, if I don't find a library ready to use I intend to implement delta encoding in blocks of 64 integers with the initial bytes specifying the block offset so far. The blocks themselves would be indexed with a tree, giving me O(log (n/64)) access time. There are way too many other options and I would prefer to not discuss them. I am really looking forward ready to use code rather than ideas on how to implement the encoding. I will be glad to share with everyone what I did once I have it working.
I appreciate your help and let me know if you have any doubts.
I use fastbit (Kesheng Wu LBL.GOV), it seems you need something good, fast and NOW, so fastbit is a highly competient improvement on Oracle's BBC (byte aligned bitmap code, berkeleydb). It's easy to setup and very good gernally.
However, given more time, you may want to look at a gray code solution, it seems optimal for your purposes.
Daniel Lemire has a number of libraries for C/++/Java released on code.google, I've read over some of his papers and they are quite nice, several advancements on fastbit and alternative approaches for column re-ordering with permutated grey codes's.
Almost forgot, I also came across Tokyo Cabinet, though I do not think it will be well suited for my current project, I may of considered it more if I had known about it before ;), it has a large degree of interoperability,
Tokyo Cabinet is written in the C
language, and provided as API of C,
Perl, Ruby, Java, and Lua. Tokyo
Cabinet is available on platforms
which have API conforming to C99 and
POSIX.
As you referred to CDB, the TC benchmark has a TC mode (TC support's several operational constraint's for varying perf) where it surpassed CDB by 10 times for read performance and 2 times for write.
With respect to your delta encoding requirement, I am quite confident in bsdiff and it's ability to out-perform any file.exe content patching system, it may also have some fundimental interfaces for your general needs.
Google's new binary compression application, courgette may be worth checking out, in case you missed the press release, 10x smaller diff's than bsdiff in the one test case I have seen published.
You have two conflicting requirements:
You want to compress very small items (8 bytes each).
You need efficient random access for each item.
The second requirement is very likely to impose a fixed length for each item.
What exactly are you trying to compress? If you are thinking about the total space of index, is it really worth the effort to save the space?
If so one thing you could try is to chop the space into half and store it into two tables. First stores (upper uint, start index, length, pointer to second table) and the second would store (index, lower uint).
For fast searching, indices would be implemented using something like B+ Tree.
I did something similar years ago for a full-text search engine. In my case, each indexed word generated a record which consisted of a record number (document id) and a word number (it could just as easily have stored word offsets) which needed to be compressed as much as possible. I used a delta-compression technique which took advantage of the fact that there would be a number of occurrences of the same word within a document, so the record number often did not need to be repeated at all. And the word offset delta would often fit within one or two bytes. Here is the code I used.
Since it's in C++, the code may is not going to be useful to you as is, but can be a good starting point for writing compressions routines.
Please excuse the hungarian notation and the magic numbers strewn within the code. Like I said, I wrote this many years ago :-)
IndexCompressor.h
//
// index compressor class
//
#pragma once
#include "File.h"
const int IC_BUFFER_SIZE = 8192;
//
// index compressor
//
class IndexCompressor
{
private :
File *m_pFile;
WA_DWORD m_dwRecNo;
WA_DWORD m_dwWordNo;
WA_DWORD m_dwRecordCount;
WA_DWORD m_dwHitCount;
WA_BYTE m_byBuffer[IC_BUFFER_SIZE];
WA_DWORD m_dwBytes;
bool m_bDebugDump;
void FlushBuffer(void);
public :
IndexCompressor(void) { m_pFile = 0; m_bDebugDump = false; }
~IndexCompressor(void) {}
void Attach(File& File) { m_pFile = &File; }
void Begin(void);
void Add(WA_DWORD dwRecNo, WA_DWORD dwWordNo);
void End(void);
WA_DWORD GetRecordCount(void) { return m_dwRecordCount; }
WA_DWORD GetHitCount(void) { return m_dwHitCount; }
void DebugDump(void) { m_bDebugDump = true; }
};
IndexCompressor.cpp
//
// index compressor class
//
#include "stdafx.h"
#include "IndexCompressor.h"
void IndexCompressor::FlushBuffer(void)
{
ASSERT(m_pFile != 0);
if (m_dwBytes > 0)
{
m_pFile->Write(m_byBuffer, m_dwBytes);
m_dwBytes = 0;
}
}
void IndexCompressor::Begin(void)
{
ASSERT(m_pFile != 0);
m_dwRecNo = m_dwWordNo = m_dwRecordCount = m_dwHitCount = 0;
m_dwBytes = 0;
}
void IndexCompressor::Add(WA_DWORD dwRecNo, WA_DWORD dwWordNo)
{
ASSERT(m_pFile != 0);
WA_BYTE buffer[16];
int nbytes = 1;
ASSERT(dwRecNo >= m_dwRecNo);
if (dwRecNo != m_dwRecNo)
m_dwWordNo = 0;
if (m_dwRecordCount == 0 || dwRecNo != m_dwRecNo)
++m_dwRecordCount;
++m_dwHitCount;
WA_DWORD dwRecNoDelta = dwRecNo - m_dwRecNo;
WA_DWORD dwWordNoDelta = dwWordNo - m_dwWordNo;
if (m_bDebugDump)
{
TRACE("%8X[%8X] %8X[%8X] : ", dwRecNo, dwRecNoDelta, dwWordNo, dwWordNoDelta);
}
// 1WWWWWWW
if (dwRecNoDelta == 0 && dwWordNoDelta < 128)
{
buffer[0] = 0x80 | WA_BYTE(dwWordNoDelta);
}
// 01WWWWWW WWWWWWWW
else if (dwRecNoDelta == 0 && dwWordNoDelta < 16384)
{
buffer[0] = 0x40 | WA_BYTE(dwWordNoDelta >> 8);
buffer[1] = WA_BYTE(dwWordNoDelta & 0x00ff);
nbytes += sizeof(WA_BYTE);
}
// 001RRRRR WWWWWWWW WWWWWWWW
else if (dwRecNoDelta < 32 && dwWordNoDelta < 65536)
{
buffer[0] = 0x20 | WA_BYTE(dwRecNoDelta);
WA_WORD *p = (WA_WORD *) (buffer+1);
*p = WA_WORD(dwWordNoDelta);
nbytes += sizeof(WA_WORD);
}
else
{
// 0001rrww
buffer[0] = 0x10;
// encode recno
if (dwRecNoDelta < 256)
{
buffer[nbytes] = WA_BYTE(dwRecNoDelta);
nbytes += sizeof(WA_BYTE);
}
else if (dwRecNoDelta < 65536)
{
buffer[0] |= 0x04;
WA_WORD *p = (WA_WORD *) (buffer+nbytes);
*p = WA_WORD(dwRecNoDelta);
nbytes += sizeof(WA_WORD);
}
else
{
buffer[0] |= 0x08;
WA_DWORD *p = (WA_DWORD *) (buffer+nbytes);
*p = dwRecNoDelta;
nbytes += sizeof(WA_DWORD);
}
// encode wordno
if (dwWordNoDelta < 256)
{
buffer[nbytes] = WA_BYTE(dwWordNoDelta);
nbytes += sizeof(WA_BYTE);
}
else if (dwWordNoDelta < 65536)
{
buffer[0] |= 0x01;
WA_WORD *p = (WA_WORD *) (buffer+nbytes);
*p = WA_WORD(dwWordNoDelta);
nbytes += sizeof(WA_WORD);
}
else
{
buffer[0] |= 0x02;
WA_DWORD *p = (WA_DWORD *) (buffer+nbytes);
*p = dwWordNoDelta;
nbytes += sizeof(WA_DWORD);
}
}
// update current setting
m_dwRecNo = dwRecNo;
m_dwWordNo = dwWordNo;
// add compressed data to buffer
ASSERT(buffer[0] != 0);
ASSERT(nbytes > 0 && nbytes < 10);
if (m_dwBytes + nbytes > IC_BUFFER_SIZE)
FlushBuffer();
CopyMemory(m_byBuffer + m_dwBytes, buffer, nbytes);
m_dwBytes += nbytes;
if (m_bDebugDump)
{
for (int i = 0; i < nbytes; ++i)
TRACE("%02X ", buffer[i]);
TRACE("\n");
}
}
void IndexCompressor::End(void)
{
FlushBuffer();
m_pFile->Write(WA_BYTE(0));
}
You've omitted critical information about the number of strings you intend to index.
But given that you say you expect the minimum length of an indexed string to be 256, storing the indices as 64% incurs at most 3% overhead. If the total length of the string file is less than 4GB, you could use 32-bit indices and incur 1.5% overhead. These numbers suggest to me that if compression matters, you're better off compressing the strings, not the indices. For that problem a variation on LZ77 seems in order.
If you want to try a wild idea, put each string in a separate file, pull them all into a zip file, and see how you can do with zziplib. This probably won't be great, but it's nearly zero work on your part.
More data on the problem would be welcome:
Number of strings
Average length of a string
Maximum length of a string
Median length of strings
Degree to which the strings file compresses with gzip
Whether you are allowed to change the order of strings to improve compression
EDIT
The comment and revised question makes the problem much clearer. I like your idea of grouping, and I would try a simple delta encoding, group the deltas, and use a variable-length code within each group. I wouldn't wire in 64 as the group size–I think you will probably want to determine that empirically.
You asked for existing libraries. For the grouping and delta encoding I doubt you will find much. For variable-length integer codes, I'm not seeing much in the way of C libraries, but you can find variable-length codings in Perl and Python. There are a ton of papers and some patents on this topic, and I suspect you're going to wind up having to roll your own. But there are some simple codes out there, and you could give UTF-8 a try—it can code unsigned integers up to 32 bits, and you can grab C code from Plan 9 and I'm sure many other sources.
Are you running on Windows? If so, I recommend creating the mmap file using naive solution your originally proposed, and then compressing the file using NTLM compression. Your application code never knows the file is compressed, and the OS does the file compression for you. You might not think this would be very performant or get good compression, but I think you'll be surprised if you try it.