350316 - use JDCT_FLOAT

Status: RESOLVED INVALID

(Core :: Graphics: ImageLib, enhancement)

Description

[wbrana]

User-Agent:       Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.8.1b2) Gecko/20060826 BonEcho/2.0b2
Build Identifier: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.8.1b2) Gecko/20060826 BonEcho/2.0b2

JDCT_FLOAT could be used to speedup jpeg decoding

Reproducible: Always

Actual Results:  
JDCT_ISLOW is used

Expected Results:  
JDCT_FLOAT could be used

Comment 1

[wbrana]

Attachment: jpeg-float.patch

Comment 2

[:Gavin Sharp [email: [e-mail]]]

You should probably ask for review from [e-mail] or [e-mail].

Comment 3

[Michael Moy]

Have you timing tests to show that FLOAT performs better than ISLOW on Linux?

Most floating point instructions take five or more cycles per operation while integer operations can take from one (addition, subtraction) to four (multiplication). This is why scaled integer math is used to simulate floating point operations.

On Windows we use an SIMD IDCT routine which can do far more operations in a given number of cycles than either FLOAT or ISLOW scaler routines can. There are patches to extend the use of SIMD IDCT code to other x86 platforms and hopefully these will be added in the future.

Comment 4

[wbrana]

I tested it with djpeg. ISLOW and FLOAT have same speed.