1 |
2 |
kdv |
[Developer Home][Contents][Search][Contact Us][Support][Intel(r)][Image]
|
2 |
|
|
|
3 |
|
|
[Application Note]
|
4 |
|
|
|
5 |
|
|
Using MMX? Instructions in a Fast iDCT
|
6 |
|
|
Algorithm for MPEG Decoding
|
7 |
|
|
|
8 |
|
|
Disclaimer
|
9 |
|
|
Information in this document is provided in connection with Intel
|
10 |
|
|
products. No license under any patent or copyright is granted expressly
|
11 |
|
|
or implied by this publication. Intel assumes no liability whatsoever,
|
12 |
|
|
including infringement of any patent or copyright, for sale and use of
|
13 |
|
|
Intel products except as provided in Intel's Terms and Conditions of Sale
|
14 |
|
|
for such products.
|
15 |
|
|
|
16 |
|
|
Intel retains the right to make changes to these specifications at any
|
17 |
|
|
time, without notice. Microcomputer Products may have minor variations to
|
18 |
|
|
their specifications known as errata.
|
19 |
|
|
|
20 |
|
|
MPEG is an international standard for audio and video compression and
|
21 |
|
|
decompression promoted by ISO. Implementations of MPEG CODEC?s or MPEG
|
22 |
|
|
enabled platforms may require licenses from various entities, including
|
23 |
|
|
Intel Corporation. Intel makes no representation as to the need for
|
24 |
|
|
licenses from any entity. No licenses, either express, implied or by
|
25 |
|
|
estoppel are granted herein. For information on licensing Intel patents,
|
26 |
|
|
please contact Intel.
|
27 |
|
|
|
28 |
|
|
1.0. INTRODUCTION
|
29 |
|
|
* 1.1. MPEG Compression Method
|
30 |
|
|
|
31 |
|
|
2.0. iDCT ALGORITHM
|
32 |
|
|
* 2.1. Selecting a Fast iDCT Algorithm
|
33 |
|
|
* 2.2. AAN Algorithm
|
34 |
|
|
|
35 |
|
|
3.0. AAN ALGORITHM IMPLEMENTED WITH MMX[tm] INSTRUCTIONS
|
36 |
|
|
* 3.1. iDCT Routine Interface
|
37 |
|
|
* 3.2. Optimization Considerations
|
38 |
|
|
|
39 |
|
|
4.0. PERFORMANCE GAINS
|
40 |
|
|
|
41 |
|
|
5.0. REFERENCES
|
42 |
|
|
|
43 |
|
|
6.0. TWO-DIMENSIONAL IDCT CODE LISTING
|
44 |
|
|
|
45 |
|
|
1.0. INTRODUCTION
|
46 |
|
|
|
47 |
|
|
The Intel Architecture (IA) media extensions include single-instruction,
|
48 |
|
|
multi-data (SIMD) instructions. This application note presents examples of
|
49 |
|
|
code that exploit these instructions. Specifically, this document describes
|
50 |
|
|
an implementation of a two-dimensional (2D) inverse Discrete Cosine Transfer
|
51 |
|
|
(iDCT) using MMX[tm] instructions. This transformation is widely used in
|
52 |
|
|
image compression algorithms; most notably in the Joint Photographic Experts
|
53 |
|
|
Group (JPEG) and Motion Picture Experts Group (MPEG) standards.
|
54 |
|
|
|
55 |
|
|
This document focuses on one iDCT algorithm that provides efficient MPEG
|
56 |
|
|
decoding. The implementation of this algorithm in MMX code, listed in
|
57 |
|
|
Section 6.0, can be used "as is" according to the guidelines presented in
|
58 |
|
|
Section 3.1. However, many iDCT algorithms exist. The reader is encouraged
|
59 |
|
|
to consider the alternative ideas and issues presented in this document,
|
60 |
|
|
since they have implications for other iDCT algorithms.
|
61 |
|
|
|
62 |
|
|
MPEG Compression Method
|
63 |
|
|
|
64 |
|
|
The MPEG compression method has two phases, encoding and decoding.
|
65 |
|
|
Multimedia images are first encoded, then transmitted, and finally decoded
|
66 |
|
|
back into a sequence of images by the MPEG player.
|
67 |
|
|
|
68 |
|
|
The encoding process follows these steps:
|
69 |
|
|
|
70 |
|
|
1. The input data is transformed using a Discrete Cosine Transform (DCT).
|
71 |
|
|
2. The resulting DCT coefficients are quantized.
|
72 |
|
|
3. The quantized coefficients are packed efficiently using Huffman tables.
|
73 |
|
|
|
74 |
|
|
During the decoding process, the MPEG player reverses these steps by
|
75 |
|
|
unpacking the coefficients, dequantizing them, and applying a 2D iDCT. To
|
76 |
|
|
achieve a high number of frames per second, this decoding process must be
|
77 |
|
|
very fast. This document concentrates on the iDCT component of the decoding
|
78 |
|
|
process.
|
79 |
|
|
|
80 |
|
|
2.0. iDCT ALGORITHM
|
81 |
|
|
|
82 |
|
|
The 8x8 two-dimensional iDCT used in MPEG decompression is defined as:
|
83 |
|
|
|
84 |
|
|
[Image]
|
85 |
|
|
|
86 |
|
|
where the normalization factors, and are defined as:
|
87 |
|
|
|
88 |
|
|
alpha () = [Image] = [Image] for = 0
|
89 |
|
|
|
90 |
|
|
= 1/2 for > 0
|
91 |
|
|
|
92 |
|
|
Where is either u or v and [Image] is the coefficient matrix.
|
93 |
|
|
|
94 |
|
|
The solution to this equation contains 64 multiplications and 63 additions
|
95 |
|
|
for each element of [Image], for a total of 4096 multiplications and 4032
|
96 |
|
|
additions.
|
97 |
|
|
|
98 |
|
|
The above equation is equivalent to a summation over v, followed by a
|
99 |
|
|
summation over u, as follows:
|
100 |
|
|
|
101 |
|
|
[Image]
|
102 |
|
|
|
103 |
|
|
This equation is the equivalent to applying a one-dimensional (1D) iDCT
|
104 |
|
|
eight times on each column of [Image] and then applying a 1D iDCT on the
|
105 |
|
|
rows of the result. Reversing this order by applying a 1D iDCT on the rows
|
106 |
|
|
first, and then on the columns, gives the same result.
|
107 |
|
|
|
108 |
|
|
2.1. Selecting a Fast iDCT Algorithm
|
109 |
|
|
|
110 |
|
|
Many algorithms have been proposed for efficient calculation of the 2D iDCT.
|
111 |
|
|
Some algorithms are based on efficient 1D iDCTs [4]; some rely on direct
|
112 |
|
|
analysis of the two-dimensional nature of iDCT [2]; and others combine a 2D
|
113 |
|
|
prescale with a very efficient 1D iDCT [1]. Some algorithms even take into
|
114 |
|
|
account the zero coefficients used in the MPEG bit streams to construct
|
115 |
|
|
statistically efficient iDCT algorithms [3]. Most fast 1D DCT and iDCT
|
116 |
|
|
algorithms are variants of Lee's Fast DCT algorithm [6], or are based on
|
117 |
|
|
variants of Winograd's FFT [5].
|
118 |
|
|
|
119 |
|
|
The following algorithms were evaluated for implementation using MMX
|
120 |
|
|
instructions:
|
121 |
|
|
|
122 |
|
|
* Statistic algorithm [3]
|
123 |
|
|
* Feig's two-dimensional algorithm [2]
|
124 |
|
|
* The LLM algorithm [4]
|
125 |
|
|
* The AAN algorithm [1]
|
126 |
|
|
|
127 |
|
|
In general, statistic algorithms inspect the input data and execute
|
128 |
|
|
conditional paths in the control flow. The overhead caused by the data
|
129 |
|
|
inspection and the jump instructions would be too expensive when compared to
|
130 |
|
|
the speed achievable by other algorithms implemented with MMX instructions.
|
131 |
|
|
|
132 |
|
|
Although Feig's 2D algorithm [2] reduces the multiplication count, the cost
|
133 |
|
|
of multiplication in MMX technology is small, so this reduction is not
|
134 |
|
|
critical. Also, the irregular memory-access pattern of this algorithm is not
|
135 |
|
|
conducive to efficient implementation using the MMX instruction set.
|
136 |
|
|
|
137 |
|
|
Both the LLM and AAN algorithms were implemented in MMX assembly code. The
|
138 |
|
|
LLM algorithm was implemented using accumulation in 32-bit elements, while
|
139 |
|
|
the AAN algorithm used accumulation in 16-bit elements. The resulting LLM
|
140 |
|
|
implementation was more accurate, conforming to the iDCT IEEE standard [7].
|
141 |
|
|
However, the AAN implementation was much faster. The AAN implementation is
|
142 |
|
|
presented in this document.
|
143 |
|
|
|
144 |
|
|
2.2. AAN Algorithm
|
145 |
|
|
|
146 |
|
|
The AAN algorithm is a one-dimensional, prescaled DCT/iDCT algorithm. First,
|
147 |
|
|
the eight input coefficients are prescaled by eight prescale values, which
|
148 |
|
|
requires eight multiplications. Then the algorithm is applied to the
|
149 |
|
|
prescaled coefficients, which requires five multiplications and 29 additions
|
150 |
|
|
for the transform. Although the 1D AAN algorithm is less efficient than
|
151 |
|
|
other 1D algorithms (for example, LLM), when applied to the two-dimensional,
|
152 |
|
|
8x8 iDCT, this algorithm takes only 64 multiplications for the prescale, and
|
153 |
|
|
80 multiplications and 464 additions for the transform.
|
154 |
|
|
|
155 |
|
|
3.0. AAN ALGORITHM IMPLEMENTED WITH MMX[tm] INSTRUCTIONS
|
156 |
|
|
|
157 |
|
|
The AAN implementation described in this document uses 16-bit data elements,
|
158 |
|
|
so that four variables can be processed in parallel using packed words. MMX
|
159 |
|
|
instructions that operate on packed words read or store four words
|
160 |
|
|
contiguously in memory. So, for an 8x8 matrix, half a row can be read or
|
161 |
|
|
stored at one time. If the 1D iDCT is applied to the columns of an 8x8
|
162 |
|
|
matrix, MMX instructions can operate on four columns at a time. Applying the
|
163 |
|
|
1D iDCT to the rows of the matrix is more involved and less efficient.
|
164 |
|
|
|
165 |
|
|
The AAN algorithm is performed in four steps:
|
166 |
|
|
|
167 |
|
|
1. Perform an iDCT on the columns of the matrix.
|
168 |
|
|
2. Transpose the matrix.
|
169 |
|
|
3. Perform a second iDCT.
|
170 |
|
|
4. Prescale the input coefficients of the iDCT on the columns of the
|
171 |
|
|
transposed matrix, which is equivalent to performing an iDCT on the
|
172 |
|
|
rows of the original matrix.
|
173 |
|
|
|
174 |
|
|
These steps result in a transposed matrix, which would have to be again
|
175 |
|
|
transposed to obtain the final result. To prevent this extra step, the input
|
176 |
|
|
matrix should be transposed initially. The cost of transposing the input is
|
177 |
|
|
negligible, since the matrix is constructed from the Zig-Zag scan [9].
|
178 |
|
|
Therefore, the actual implementation follows these steps:
|
179 |
|
|
|
180 |
|
|
1. Prescale the input coefficients of the transposed input matrix.
|
181 |
|
|
2. Perform an iDCT on the columns of the transposed input matrix, which is
|
182 |
|
|
equivalent to performing an iDCT on the rows of the final matrix.
|
183 |
|
|
3. Transpose the matrix.
|
184 |
|
|
4. Perform a second iDCT on the columns of the final matrix.
|
185 |
|
|
|
186 |
|
|
Another consideration is the limited length of the MMX registers. All the
|
187 |
|
|
iDCT algorithms mentioned in Section 3.1 are defined mathematically without
|
188 |
|
|
regard to the size of the accumulator or registers. To ensure adequate
|
189 |
|
|
precision for operations on 16-bit data elements, the algorithm was analyzed
|
190 |
|
|
carefully and appropriate precision was assigned during every intermediate
|
191 |
|
|
stage of the calculation.
|
192 |
|
|
|
193 |
|
|
Precision was controlled using packed shift instructions, which shift all
|
194 |
|
|
data elements in a register by the same amount. Shift right instructions
|
195 |
|
|
were used to prevent overflow of the most significant bit in each
|
196 |
|
|
intermediate step of the calculation.
|
197 |
|
|
|
198 |
|
|
3.1. iDCT Routine Interface
|
199 |
|
|
|
200 |
|
|
The iDCT routine is called from within an assembly module. The routine gets
|
201 |
|
|
a pointer to an 8x8, 16-bit element matrix; the pointer is located in the
|
202 |
|
|
ESI register. The matrix should be aligned on an 8-byte boundary. Each data
|
203 |
|
|
element is actually a 12-bit quantity that is left-adjusted, meaning that
|
204 |
|
|
the four least significant bits equal zero. The input matrix should be
|
205 |
|
|
transposed; otherwise, the output matrix must be transposed. The value 0.5
|
206 |
|
|
must be added to the DC value, input matrix element [0][0].
|
207 |
|
|
|
208 |
|
|
The routine returns the same memory array. Therefore, if the original input
|
209 |
|
|
operands are needed (for example, in test mode), they must be copied before
|
210 |
|
|
the call to the iDCT routine.
|
211 |
|
|
|
212 |
|
|
3.2. Optimization Considerations
|
213 |
|
|
|
214 |
|
|
One standard Pentium® processor optimization technique is code rescheduling
|
215 |
|
|
to exploit parallelism in an algorithm. Parallelism in the 2D iDCT was
|
216 |
|
|
approached from two directions, as illustrated in Figure 1:
|
217 |
|
|
|
218 |
|
|
* Within a single 8x8 iDCT
|
219 |
|
|
* Between four 8x8 iDCTs.
|
220 |
|
|
|
221 |
|
|
In the first approach, data is accessed by rows within the matrix. In the
|
222 |
|
|
second approach, data from the four matrixes is interleaved to enable
|
223 |
|
|
efficient use of the MMX instructions.
|
224 |
|
|
|
225 |
|
|
Figure 1. Single iDCT vs. Four iDCTs
|
226 |
|
|
|
227 |
|
|
[Image]
|
228 |
|
|
|
229 |
|
|
The advantage of the single-iDCT approach is that the interface to an MPEG
|
230 |
|
|
player is simpler. Its disadvantages are:
|
231 |
|
|
|
232 |
|
|
* The matrix must be transposed in order to operate on several rows in
|
233 |
|
|
parallel.
|
234 |
|
|
* To prevent overflow, packed shift instructions must be used. Since in a
|
235 |
|
|
given register, the accuracy of the four data elements varies, the
|
236 |
|
|
shift count is determined by the worst case among the four elements.
|
237 |
|
|
This results in an extra loss of accuracy.
|
238 |
|
|
|
239 |
|
|
The advantages of the four-iDCTs approach are:
|
240 |
|
|
|
241 |
|
|
* Matrix transposition is not required.
|
242 |
|
|
* To prevent overflow, packed shift instructions must still be used.
|
243 |
|
|
However, since all the data elements in a register have the same
|
244 |
|
|
accuracy, there is no extra loss of accuracy to accommodate the worst
|
245 |
|
|
case among four elements.
|
246 |
|
|
|
247 |
|
|
The disadvantage of the single-iDCT approach is that, to take advantage of
|
248 |
|
|
the MMX instructions, the input data from the four matrixes must first be
|
249 |
|
|
interleaved (see Figure 1). Then, after the transform, the resulting data
|
250 |
|
|
must be restored to four 8x8 matrixes.
|
251 |
|
|
|
252 |
|
|
Because of its simplicity, the single-iDCT approach was chosen. Instruction
|
253 |
|
|
scheduling was done manually to ensure optimal performance.
|
254 |
|
|
|
255 |
|
|
Register use was carefully considered as well. In most cases, intermediate
|
256 |
|
|
results were kept in registers; temporary storage to memory was needed in
|
257 |
|
|
only a few cases. For example, consider the implementation of the matrix
|
258 |
|
|
transpose. The basic operation is the transpose of 4x4 elements [8], as
|
259 |
|
|
illustrated in Figure 2.
|
260 |
|
|
|
261 |
|
|
Figure 2. Matrix Transpose Operation
|
262 |
|
|
|
263 |
|
|
[Image]
|
264 |
|
|
|
265 |
|
|
The 8x8 iDCT requires four of these operations. The sequence of these
|
266 |
|
|
operations was carefully chosen to save memory stores and loads. First, M4
|
267 |
|
|
was transposed, followed by M3. These two results were immediately used to
|
268 |
|
|
perform the iDCT on the last four columns. Similarly, M2 was transposed,
|
269 |
|
|
followed by M1. These results were used for the iDCT on the first four
|
270 |
|
|
columns.
|
271 |
|
|
|
272 |
|
|
The detailed steps of the matrix transpose algorithm are:
|
273 |
|
|
|
274 |
|
|
1. Prescale: 16 packed multiplies
|
275 |
|
|
2. Column 0: even part
|
276 |
|
|
3. Column 0: odd part
|
277 |
|
|
4. Column 0: output butterfly
|
278 |
|
|
5. Column 1: even part
|
279 |
|
|
6. Column 1: odd part
|
280 |
|
|
7. Column 1: output butterfly
|
281 |
|
|
8. Transpose: M4 part
|
282 |
|
|
9. Transpose: M3 part
|
283 |
|
|
10. Column 1: even part (after transpose)
|
284 |
|
|
11. Column 1: odd part (after transpose)
|
285 |
|
|
12. Column 1: output butterfly (after transform)
|
286 |
|
|
13. Transpose: M2 parts
|
287 |
|
|
14. Transpose: M1 parts
|
288 |
|
|
15. Column 0: odd part (after transpose)
|
289 |
|
|
16. Column 0: odd part (after transpose)
|
290 |
|
|
17. Column 0: output butterfly (after transform)
|
291 |
|
|
18. Cleanup
|
292 |
|
|
|
293 |
|
|
where:
|
294 |
|
|
|
295 |
|
|
* Column 0 represents the first four columns.
|
296 |
|
|
* Even part calculates the part of the iDCT that uses even-indexed
|
297 |
|
|
elements.
|
298 |
|
|
* Odd part calculates the part of the iDCT that uses odd-indexed
|
299 |
|
|
elements.
|
300 |
|
|
* Output butterfly generates the 1D iDCT using the results of the even
|
301 |
|
|
and odd parts.
|
302 |
|
|
|
303 |
|
|
During the rescheduling process, instructions from one block were moved to
|
304 |
|
|
the previous block whenever an empty slot could be filled. This reordering
|
305 |
|
|
is marked by comments in the code listed in Section 6.0.
|
306 |
|
|
|
307 |
|
|
4.0. PERFORMANCE GAINS
|
308 |
|
|
|
309 |
|
|
The cycle count for this implementation of the AAN algorithm, using MMX
|
310 |
|
|
instructions, is 240 clocks. A direct comparison of this implementation with
|
311 |
|
|
a scalar implementation of the AAN algorithm would be misleading, since the
|
312 |
|
|
AAN algorithm is not the fastest scalar implementation of an iDCT. However,
|
313 |
|
|
the implementation presented here is estimated to be 3 to 3.5 times faster
|
314 |
|
|
than the general performance of scalar iDCT algorithms.
|
315 |
|
|
|
316 |
|
|
5.0. REFERENCES
|
317 |
|
|
|
318 |
|
|
1. Arai, Y., T. Agui, and M. Nakajima, (1988). A Fast DCT-SQ Scheme for
|
319 |
|
|
Images; Trans IEICE, 71, pp. 1095-1097.
|
320 |
|
|
2. Feig E., and S. Winograd, (1992). Fast Algorithms for Discrete Cosine
|
321 |
|
|
Transform, IEEE Trans. Signal Proc., 40, pp. 2174-2193.
|
322 |
|
|
3. Hung, A.C., and Thy Meng, (1994). A Fast Statistical Inverse Discrete
|
323 |
|
|
cosine Transform for Image Compression, SPIE/IS&Teletronic Imaging
|
324 |
|
|
,2187 , pp. 196-205.
|
325 |
|
|
4. Loeffler, C., A. Ligtenberg, and C. S. Moschytz, (1989). Practical Fast
|
326 |
|
|
1D DCT Algorithm with Eleven Multiplications, Proc. ICASSP 1989, pp.
|
327 |
|
|
988-991.
|
328 |
|
|
5. Winograd S. (1976). On Computing the Discrete Fourier Transform, IBM
|
329 |
|
|
Res. Rep, RC-6291.
|
330 |
|
|
6. Lee, B. A New Algorithm to Compute the Discrete Cosine Transform, IEEE
|
331 |
|
|
Trans. Signal Proc., Dec/84, pp. 1243-1245.
|
332 |
|
|
7. IEEE standard specification for the implementation of 8x8 iDCT IEEE std
|
333 |
|
|
1180-1990
|
334 |
|
|
8. MPEG standard, Coding of Moving Pictures, ISO/IEC DIS 11172.
|
335 |
|
|
|
336 |
|
|
6.0. TWO-DIMENSIONAL IDCT CODE LISTING
|
337 |
|
|
|
338 |
|
|
; esi - input and output data pointer
|
339 |
|
|
; the input data is tranposed and each 16 bit element in the 8x8 matrix
|
340 |
|
|
;is left aligned:
|
341 |
|
|
; for example in 11...1110000 format
|
342 |
|
|
; If the iDCT is of I macroblock then 0.5 needs to be added to the
|
343 |
|
|
;DC Component
|
344 |
|
|
; (element[0][0] of the matrix)
|
345 |
|
|
|
346 |
|
|
.nolist
|
347 |
|
|
include iammx.inc ; IAMMX Emulator Macros
|
348 |
|
|
MMWORD TEXTEQU
|
349 |
|
|
.list
|
350 |
|
|
|
351 |
|
|
.586
|
352 |
|
|
.model flat
|
353 |
|
|
|
354 |
|
|
_DATA SEGMENT PARA PUBLIC USE32 'DATA'
|
355 |
|
|
x0005000200010001 DQ 0005000200010001h
|
356 |
|
|
x0040000000000000 DQ 40000000000000h
|
357 |
|
|
x5a825a825a825a82 DW 5a82h, 5a82h, 5a82h, 5a82h ; 23170
|
358 |
|
|
x539f539f539f539f DW 539fh, 539fh, 539fh, 539fh ; 21407
|
359 |
|
|
x4546454645464546 DW 4546h, 4546h, 4546h, 4546h ; 17734
|
360 |
|
|
x61f861f861f861f8 DW 61f8h, 61f8h, 61f8h, 61f8h ; 25080
|
361 |
|
|
scratch1 DQ 0
|
362 |
|
|
scratch3 DQ 0
|
363 |
|
|
scratch5 DQ 0
|
364 |
|
|
scratch7 DQ 0
|
365 |
|
|
; for debug only
|
366 |
|
|
x0 DQ 0
|
367 |
|
|
|
368 |
|
|
preSC DW 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520
|
369 |
|
|
DW 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270
|
370 |
|
|
DW 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906
|
371 |
|
|
DW 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315
|
372 |
|
|
DW 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520
|
373 |
|
|
DW 12873, 17855, 16819, 15137, 25746, 20228, 13933, 7103
|
374 |
|
|
DW 17734, 24598, 23170, 20853, 17734, 13933, 9597, 4892
|
375 |
|
|
DW 18081, 25080, 23624, 21261, 18081, 14206, 9785, 4988
|
376 |
|
|
|
377 |
|
|
_DATA ENDS
|
378 |
|
|
|
379 |
|
|
_TEXT SEGMENT PARA PUBLIC USE32 'CODE'
|
380 |
|
|
|
381 |
|
|
COMMENT ^
|
382 |
|
|
void idct8x8aan (
|
383 |
|
|
int16 *src_result);
|
384 |
|
|
^
|
385 |
|
|
public _idct8x8aan
|
386 |
|
|
_idct8x8aan proc near
|
387 |
|
|
|
388 |
|
|
push ebp
|
389 |
|
|
lea ecx, [preSC]
|
390 |
|
|
|
391 |
|
|
mov ebp, esp
|
392 |
|
|
push esi
|
393 |
|
|
|
394 |
|
|
mov esi, DWORD PTR [ebp+8] ; source
|
395 |
|
|
;slot
|
396 |
|
|
|
397 |
|
|
; column 0: even part
|
398 |
|
|
; use V4, V12, V0, V8 to produce V22..V25
|
399 |
|
|
movq mm0, mmword ptr [ecx+8*12] ; maybe the first mul can be done together
|
400 |
|
|
; with the dequantization in iHuff module ?
|
401 |
|
|
;slot
|
402 |
|
|
|
403 |
|
|
pmulhw mm0, mmword ptr [esi+8*12] ; V12
|
404 |
|
|
;slot
|
405 |
|
|
|
406 |
|
|
movq mm1, mmword ptr [ecx+8*4]
|
407 |
|
|
;slot
|
408 |
|
|
|
409 |
|
|
pmulhw mm1, mmword ptr [esi+8*4] ; V4
|
410 |
|
|
;slot
|
411 |
|
|
|
412 |
|
|
movq mm3, mmword ptr [ecx+8*0]
|
413 |
|
|
psraw mm0, 1 ; t64=t66
|
414 |
|
|
|
415 |
|
|
pmulhw mm3, mmword ptr [esi+8*0] ; V0
|
416 |
|
|
;slot
|
417 |
|
|
|
418 |
|
|
movq mm5, mmword ptr [ecx+8*8] ; duplicate V4
|
419 |
|
|
movq mm2, mm1 ; added 11/1/96
|
420 |
|
|
|
421 |
|
|
pmulhw mm5, mmword ptr [esi+8*8] ; V8
|
422 |
|
|
psubsw mm1, mm0 ; V16
|
423 |
|
|
|
424 |
|
|
pmulhw mm1, mmword ptr x5a825a825a825a82 ; 23170 ->V18
|
425 |
|
|
paddsw mm2, mm0 ; V17
|
426 |
|
|
|
427 |
|
|
movq mm0, mm2 ; duplicate V17
|
428 |
|
|
psraw mm2, 1 ; t75=t82
|
429 |
|
|
|
430 |
|
|
psraw mm0, 2 ; t72
|
431 |
|
|
movq mm4, mm3 ; duplicate V0
|
432 |
|
|
|
433 |
|
|
paddsw mm3, mm5 ; V19
|
434 |
|
|
psubsw mm4, mm5 ; V20 ;mm5 free
|
435 |
|
|
|
436 |
|
|
;moved from the block below
|
437 |
|
|
movq mm7, mmword ptr [ecx+8*10]
|
438 |
|
|
psraw mm3, 1 ; t74=t81
|
439 |
|
|
|
440 |
|
|
movq mm6, mm3 ; duplicate t74=t81
|
441 |
|
|
psraw mm4, 2 ; t77=t79
|
442 |
|
|
|
443 |
|
|
psubsw mm1, mm0 ; V21 ; mm0 free
|
444 |
|
|
paddsw mm3, mm2 ; V22
|
445 |
|
|
|
446 |
|
|
movq mm5, mm1 ; duplicate V21
|
447 |
|
|
paddsw mm1, mm4 ; V23
|
448 |
|
|
|
449 |
|
|
movq mmword ptr [esi+8*4], mm3 ; V22
|
450 |
|
|
psubsw mm4, mm5 ; V24; mm5 free
|
451 |
|
|
|
452 |
|
|
movq mmword ptr [esi+8*12], mm1 ; V23
|
453 |
|
|
psubsw mm6, mm2 ; V25; mm2 free
|
454 |
|
|
|
455 |
|
|
movq mmword ptr [esi+8*0], mm4 ; V24
|
456 |
|
|
;slot
|
457 |
|
|
|
458 |
|
|
; keep mm6 alive all along the next block
|
459 |
|
|
;movq mmword ptr [esi+8*8], mm6 ; V25
|
460 |
|
|
|
461 |
|
|
; column 0: odd part
|
462 |
|
|
; use V2, V6, V10, V14 to produce V31, V39, V40, V41
|
463 |
|
|
|
464 |
|
|
;moved above
|
465 |
|
|
;movq mm7, mmword ptr [ecx+8*10]
|
466 |
|
|
|
467 |
|
|
pmulhw mm7, mmword ptr [esi+8*10] ; V10
|
468 |
|
|
;slot
|
469 |
|
|
|
470 |
|
|
movq mm0, mmword ptr [ecx+8*6]
|
471 |
|
|
;slot
|
472 |
|
|
|
473 |
|
|
pmulhw mm0, mmword ptr [esi+8*6] ; V6
|
474 |
|
|
;slot
|
475 |
|
|
|
476 |
|
|
movq mm5, mmword ptr [ecx+8*2]
|
477 |
|
|
movq mm3, mm7 ; duplicate V10
|
478 |
|
|
|
479 |
|
|
pmulhw mm5, mmword ptr [esi+8*2] ; V2
|
480 |
|
|
;slot
|
481 |
|
|
|
482 |
|
|
movq mm4, mmword ptr [ecx+8*14]
|
483 |
|
|
psubsw mm7, mm0 ; V26
|
484 |
|
|
|
485 |
|
|
pmulhw mm4, mmword ptr [esi+8*14] ; V14
|
486 |
|
|
paddsw mm3, mm0 ; V29 ; free mm0
|
487 |
|
|
|
488 |
|
|
movq mm1, mm7 ; duplicate V26
|
489 |
|
|
psraw mm3, 1 ; t91=t94
|
490 |
|
|
|
491 |
|
|
pmulhw mm7, mmword ptr x539f539f539f539f ; V33
|
492 |
|
|
psraw mm1, 1 ; t96
|
493 |
|
|
|
494 |
|
|
movq mm0, mm5 ; duplicate V2
|
495 |
|
|
psraw mm4, 2 ; t85=t87
|
496 |
|
|
|
497 |
|
|
paddsw mm5, mm4 ; V27
|
498 |
|
|
psubsw mm0, mm4 ; V28 ; free mm4
|
499 |
|
|
|
500 |
|
|
movq mm2, mm0 ; duplicate V28
|
501 |
|
|
psraw mm5, 1 ; t90=t93
|
502 |
|
|
|
503 |
|
|
pmulhw mm0, mmword ptr x4546454645464546 ; V35
|
504 |
|
|
psraw mm2, 1 ; t97
|
505 |
|
|
|
506 |
|
|
movq mm4, mm5 ; duplicate t90=t93
|
507 |
|
|
psubsw mm1, mm2 ; V32 ; free mm2
|
508 |
|
|
|
509 |
|
|
pmulhw mm1, mmword ptr x61f861f861f861f8 ; V36
|
510 |
|
|
psllw mm7, 1 ; t107
|
511 |
|
|
|
512 |
|
|
paddsw mm5, mm3 ; V31
|
513 |
|
|
psubsw mm4, mm3 ; V30 ; free mm3
|
514 |
|
|
|
515 |
|
|
pmulhw mm4, mmword ptr x5a825a825a825a82 ; V34
|
516 |
|
|
nop ;slot
|
517 |
|
|
|
518 |
|
|
psubsw mm0, mm1 ; V38
|
519 |
|
|
psubsw mm1, mm7 ; V37 ; free mm7
|
520 |
|
|
|
521 |
|
|
psllw mm1, 1 ; t114
|
522 |
|
|
;move from the next block
|
523 |
|
|
movq mm3, mm6 ; duplicate V25
|
524 |
|
|
|
525 |
|
|
;move from the next block
|
526 |
|
|
movq mm7, mmword ptr [esi+8*4] ; V22
|
527 |
|
|
psllw mm0, 1 ; t110
|
528 |
|
|
|
529 |
|
|
psubsw mm0, mm5 ; V39 (mm5 still needed for next block)
|
530 |
|
|
psllw mm4, 2 ; t112
|
531 |
|
|
|
532 |
|
|
;move from the next block
|
533 |
|
|
movq mm2, mmword ptr [esi+8*12] ; V23
|
534 |
|
|
psubsw mm4, mm0 ; V40
|
535 |
|
|
|
536 |
|
|
paddsw mm1, mm4 ; V41; free mm0
|
537 |
|
|
;move from the next block
|
538 |
|
|
psllw mm2, 1 ; t117=t125
|
539 |
|
|
|
540 |
|
|
; column 0: output butterfly
|
541 |
|
|
;move above
|
542 |
|
|
;movq mm3, mm6 ; duplicate V25
|
543 |
|
|
;movq mm7, mmword ptr [esi+8*4] ; V22
|
544 |
|
|
;movq mm2, mmword ptr [esi+8*12] ; V23
|
545 |
|
|
;psllw mm2, 1 ; t117=t125
|
546 |
|
|
|
547 |
|
|
psubsw mm6, mm1 ; tm6
|
548 |
|
|
paddsw mm3, mm1 ; tm8; free mm1
|
549 |
|
|
|
550 |
|
|
movq mm1, mm7 ; duplicate V22
|
551 |
|
|
paddsw mm7, mm5 ; tm0
|
552 |
|
|
|
553 |
|
|
movq mmword ptr [esi+8*8], mm3 ; tm8; free mm3
|
554 |
|
|
psubsw mm1, mm5 ; tm14; free mm5
|
555 |
|
|
|
556 |
|
|
movq mmword ptr [esi+8*6], mm6 ; tm6; free mm6
|
557 |
|
|
movq mm3, mm2 ; duplicate t117=t125
|
558 |
|
|
|
559 |
|
|
movq mm6, mmword ptr [esi+8*0] ; V24
|
560 |
|
|
paddsw mm2, mm0 ; tm2
|
561 |
|
|
|
562 |
|
|
movq mmword ptr [esi+8*0], mm7 ; tm0; free mm7
|
563 |
|
|
psubsw mm3, mm0 ; tm12; free mm0
|
564 |
|
|
|
565 |
|
|
movq mmword ptr [esi+8*14], mm1 ; tm14; free mm1
|
566 |
|
|
psllw mm6, 1 ; t119=t123
|
567 |
|
|
|
568 |
|
|
movq mmword ptr [esi+8*2], mm2 ; tm2; free mm2
|
569 |
|
|
movq mm0, mm6 ; duplicate t119=t123
|
570 |
|
|
|
571 |
|
|
movq mmword ptr [esi+8*12], mm3 ; tm12; free mm3
|
572 |
|
|
paddsw mm6, mm4 ; tm4
|
573 |
|
|
|
574 |
|
|
;moved from next block
|
575 |
|
|
movq mm1, mmword ptr [ecx+8*5]
|
576 |
|
|
psubsw mm0, mm4 ; tm10; free mm4
|
577 |
|
|
|
578 |
|
|
;moved from next block
|
579 |
|
|
pmulhw mm1, mmword ptr [esi+8*5] ; V5
|
580 |
|
|
;slot
|
581 |
|
|
|
582 |
|
|
movq mmword ptr [esi+8*4], mm6 ; tm4; free mm6
|
583 |
|
|
;slot
|
584 |
|
|
|
585 |
|
|
movq mmword ptr [esi+8*10], mm0 ; tm10; free mm0
|
586 |
|
|
;slot
|
587 |
|
|
|
588 |
|
|
; column 1: even part
|
589 |
|
|
; use V5, V13, V1, V9 to produce V56..V59
|
590 |
|
|
;moved to prev block
|
591 |
|
|
;movq mm1, mmword ptr [ecx+8*5]
|
592 |
|
|
;pmulhw mm1, mmword ptr [esi+8*5] ; V5
|
593 |
|
|
|
594 |
|
|
movq mm7, mmword ptr [ecx+8*13]
|
595 |
|
|
psllw mm1, 1 ; t128=t130
|
596 |
|
|
|
597 |
|
|
pmulhw mm7, mmword ptr [esi+8*13] ; V13
|
598 |
|
|
movq mm2, mm1 ; duplicate t128=t130
|
599 |
|
|
|
600 |
|
|
movq mm3, mmword ptr [ecx+8*1]
|
601 |
|
|
;slot
|
602 |
|
|
|
603 |
|
|
pmulhw mm3, mmword ptr [esi+8*1] ; V1
|
604 |
|
|
;slot
|
605 |
|
|
|
606 |
|
|
movq mm5, mmword ptr [ecx+8*9]
|
607 |
|
|
psubsw mm1, mm7 ; V50
|
608 |
|
|
|
609 |
|
|
pmulhw mm5, mmword ptr [esi+8*9] ; V9
|
610 |
|
|
paddsw mm2, mm7 ; V51
|
611 |
|
|
|
612 |
|
|
pmulhw mm1, mmword ptr x5a825a825a825a82 ; 23170 ->V52
|
613 |
|
|
movq mm6, mm2 ; duplicate V51
|
614 |
|
|
|
615 |
|
|
psraw mm2, 1 ; t138=t144
|
616 |
|
|
movq mm4, mm3 ; duplicate V1
|
617 |
|
|
|
618 |
|
|
psraw mm6, 2 ; t136
|
619 |
|
|
paddsw mm3, mm5 ; V53
|
620 |
|
|
|
621 |
|
|
psubsw mm4, mm5 ; V54 ;mm5 free
|
622 |
|
|
movq mm7, mm3 ; duplicate V53
|
623 |
|
|
|
624 |
|
|
;moved from next block
|
625 |
|
|
movq mm0, mmword ptr [ecx+8*11]
|
626 |
|
|
psraw mm4, 1 ; t140=t142
|
627 |
|
|
|
628 |
|
|
psubsw mm1, mm6 ; V55 ; mm6 free
|
629 |
|
|
paddsw mm3, mm2 ; V56
|
630 |
|
|
|
631 |
|
|
movq mm5, mm4 ; duplicate t140=t142
|
632 |
|
|
paddsw mm4, mm1 ; V57
|
633 |
|
|
|
634 |
|
|
movq mmword ptr [esi+8*5], mm3 ; V56
|
635 |
|
|
psubsw mm5, mm1 ; V58; mm1 free
|
636 |
|
|
|
637 |
|
|
movq mmword ptr [esi+8*13], mm4 ; V57
|
638 |
|
|
psubsw mm7, mm2 ; V59; mm2 free
|
639 |
|
|
|
640 |
|
|
movq mmword ptr [esi+8*9], mm5 ; V58
|
641 |
|
|
;slot
|
642 |
|
|
|
643 |
|
|
; keep mm7 alive all along the next block
|
644 |
|
|
;movq mmword ptr [esi+8*1], mm7 ; V59
|
645 |
|
|
|
646 |
|
|
;moved above
|
647 |
|
|
;movq mm0, mmword ptr [ecx+8*11]
|
648 |
|
|
|
649 |
|
|
pmulhw mm0, mmword ptr [esi+8*11] ; V11
|
650 |
|
|
;slot
|
651 |
|
|
|
652 |
|
|
movq mm6, mmword ptr [ecx+8*7]
|
653 |
|
|
;slot
|
654 |
|
|
|
655 |
|
|
pmulhw mm6, mmword ptr [esi+8*7] ; V7
|
656 |
|
|
;slot
|
657 |
|
|
|
658 |
|
|
movq mm4, mmword ptr [ecx+8*15]
|
659 |
|
|
movq mm3, mm0 ; duplicate V11
|
660 |
|
|
|
661 |
|
|
pmulhw mm4, mmword ptr [esi+8*15] ; V15
|
662 |
|
|
;slot
|
663 |
|
|
|
664 |
|
|
movq mm5, mmword ptr [ecx+8*3]
|
665 |
|
|
psllw mm6,1 ; t146=t152
|
666 |
|
|
|
667 |
|
|
pmulhw mm5, mmword ptr [esi+8*3] ; V3
|
668 |
|
|
paddsw mm0, mm6 ; V63
|
669 |
|
|
|
670 |
|
|
; note that V15 computation has a correction step:
|
671 |
|
|
; this is a 'magic' constant that rebiases the results to be closer to the expected result
|
672 |
|
|
; this magic constant can be refined to reduce the error even more
|
673 |
|
|
; by doing the correction step in a later stage when the number is actually multiplied by 16
|
674 |
|
|
|
675 |
|
|
paddw mm4, mmword ptr x0005000200010001
|
676 |
|
|
psubsw mm3, mm6 ; V60 ; free mm6
|
677 |
|
|
|
678 |
|
|
psraw mm0, 1 ; t154=t156
|
679 |
|
|
movq mm1, mm3 ; duplicate V60
|
680 |
|
|
|
681 |
|
|
pmulhw mm1, mmword ptr x539f539f539f539f ; V67
|
682 |
|
|
movq mm6, mm5 ; duplicate V3
|
683 |
|
|
|
684 |
|
|
psraw mm4, 2 ; t148=t150
|
685 |
|
|
;slot
|
686 |
|
|
|
687 |
|
|
paddsw mm5, mm4 ; V61
|
688 |
|
|
psubsw mm6, mm4 ; V62 ; free mm4
|
689 |
|
|
|
690 |
|
|
movq mm4, mm5 ; duplicate V61
|
691 |
|
|
psllw mm1, 1 ; t169
|
692 |
|
|
|
693 |
|
|
paddsw mm5, mm0 ; V65 -> result
|
694 |
|
|
psubsw mm4, mm0 ; V64 ; free mm0
|
695 |
|
|
|
696 |
|
|
pmulhw mm4, mmword ptr x5a825a825a825a82 ; V68
|
697 |
|
|
psraw mm3, 1 ; t158
|
698 |
|
|
|
699 |
|
|
psubsw mm3, mm6 ; V66
|
700 |
|
|
movq mm2, mm5 ; duplicate V65
|
701 |
|
|
|
702 |
|
|
pmulhw mm3, mmword ptr x61f861f861f861f8 ; V70
|
703 |
|
|
psllw mm6, 1 ; t165
|
704 |
|
|
|
705 |
|
|
pmulhw mm6, mmword ptr x4546454645464546 ; V69
|
706 |
|
|
psraw mm2, 1 ; t172
|
707 |
|
|
|
708 |
|
|
;moved from next block
|
709 |
|
|
movq mm0, mmword ptr [esi+8*5] ; V56
|
710 |
|
|
psllw mm4, 1 ; t174
|
711 |
|
|
|
712 |
|
|
;moved from next block
|
713 |
|
|
psraw mm0, 1 ; t177=t188
|
714 |
|
|
nop ; slot
|
715 |
|
|
|
716 |
|
|
psubsw mm6, mm3 ; V72
|
717 |
|
|
psubsw mm3, mm1 ; V71 ; free mm1
|
718 |
|
|
|
719 |
|
|
psubsw mm6, mm2 ; V73 ; free mm2
|
720 |
|
|
;moved from next block
|
721 |
|
|
psraw mm5, 1 ; t178=t189
|
722 |
|
|
|
723 |
|
|
psubsw mm4, mm6 ; V74
|
724 |
|
|
;moved from next block
|
725 |
|
|
movq mm1, mm0 ; duplicate t177=t188
|
726 |
|
|
|
727 |
|
|
paddsw mm3, mm4 ; V75
|
728 |
|
|
;moved from next block
|
729 |
|
|
paddsw mm0, mm5 ; tm1
|
730 |
|
|
|
731 |
|
|
;location
|
732 |
|
|
; 5 - V56
|
733 |
|
|
; 13 - V57
|
734 |
|
|
; 9 - V58
|
735 |
|
|
; X - V59, mm7
|
736 |
|
|
; X - V65, mm5
|
737 |
|
|
; X - V73, mm6
|
738 |
|
|
; X - V74, mm4
|
739 |
|
|
; X - V75, mm3
|
740 |
|
|
; free mm0, mm1 & mm2
|
741 |
|
|
;move above
|
742 |
|
|
;movq mm0, mmword ptr [esi+8*5] ; V56
|
743 |
|
|
;psllw mm0, 1 ; t177=t188 ! new !!
|
744 |
|
|
;psllw mm5, 1 ; t178=t189 ! new !!
|
745 |
|
|
;movq mm1, mm0 ; duplicate t177=t188
|
746 |
|
|
;paddsw mm0, mm5 ; tm1
|
747 |
|
|
|
748 |
|
|
movq mm2, mmword ptr [esi+8*13] ; V57
|
749 |
|
|
psubsw mm1, mm5 ; tm15; free mm5
|
750 |
|
|
|
751 |
|
|
movq mmword ptr [esi+8*1], mm0 ; tm1; free mm0
|
752 |
|
|
psraw mm7, 1 ; t182=t184 ! new !!
|
753 |
|
|
|
754 |
|
|
;save the store as used directly in the transpose
|
755 |
|
|
;movq mmword ptr [esi+8*15], mm1 ; tm15; free mm1
|
756 |
|
|
movq mm5, mm7 ; duplicate t182=t184
|
757 |
|
|
psubsw mm7, mm3 ; tm7
|
758 |
|
|
|
759 |
|
|
paddsw mm5, mm3 ; tm9; free mm3
|
760 |
|
|
;slot
|
761 |
|
|
|
762 |
|
|
movq mm0, mmword ptr [esi+8*9] ; V58
|
763 |
|
|
movq mm3, mm2 ; duplicate V57
|
764 |
|
|
|
765 |
|
|
movq mmword ptr [esi+8*7], mm7 ; tm7; free mm7
|
766 |
|
|
psubsw mm3, mm6 ; tm13
|
767 |
|
|
|
768 |
|
|
paddsw mm2, mm6 ; tm3 ; free mm6
|
769 |
|
|
; moved up from the transpose
|
770 |
|
|
movq mm7, mm3
|
771 |
|
|
|
772 |
|
|
; moved up from the transpose
|
773 |
|
|
punpcklwd mm3, mm1
|
774 |
|
|
movq mm6, mm0 ; duplicate V58
|
775 |
|
|
|
776 |
|
|
movq mmword ptr [esi+8*3], mm2 ; tm3; free mm2
|
777 |
|
|
paddsw mm0, mm4 ; tm5
|
778 |
|
|
|
779 |
|
|
psubsw mm6, mm4 ; tm11; free mm4
|
780 |
|
|
; moved up from the transpose
|
781 |
|
|
punpckhwd mm7, mm1
|
782 |
|
|
|
783 |
|
|
movq mmword ptr [esi+8*5], mm0 ; tm5; free mm0
|
784 |
|
|
; moved up from the transpose
|
785 |
|
|
movq mm2, mm5
|
786 |
|
|
|
787 |
|
|
; transpose - M4 part
|
788 |
|
|
; --------- ---------
|
789 |
|
|
; | M1 | M2 | | M1'| M3'|
|
790 |
|
|
; --------- --> ---------
|
791 |
|
|
; | M3 | M4 | | M2'| M4'|
|
792 |
|
|
; --------- ---------
|
793 |
|
|
; Two alternatives: use full mmword approach so the following code can be
|
794 |
|
|
; scheduled before the transpose is done without stores, or use the faster
|
795 |
|
|
; half mmword stores (when possible)
|
796 |
|
|
|
797 |
|
|
movdf dword ptr [esi+8*9+4], mm3 ; MS part of tmt9
|
798 |
|
|
punpcklwd mm5, mm6
|
799 |
|
|
|
800 |
|
|
movdf dword ptr [esi+8*13+4], mm7 ; MS part of tmt13
|
801 |
|
|
punpckhwd mm2, mm6
|
802 |
|
|
|
803 |
|
|
movdf dword ptr [esi+8*9], mm5 ; LS part of tmt9
|
804 |
|
|
punpckhdq mm5, mm3 ; free mm3
|
805 |
|
|
|
806 |
|
|
movdf dword ptr [esi+8*13], mm2 ; LS part of tmt13
|
807 |
|
|
punpckhdq mm2, mm7 ; free mm7
|
808 |
|
|
|
809 |
|
|
; moved up from the M3 transpose
|
810 |
|
|
movq mm0, mmword ptr [esi+8*8]
|
811 |
|
|
;slot
|
812 |
|
|
|
813 |
|
|
; moved up from the M3 transpose
|
814 |
|
|
movq mm1, mmword ptr [esi+8*10]
|
815 |
|
|
; moved up from the M3 transpose
|
816 |
|
|
movq mm3, mm0
|
817 |
|
|
|
818 |
|
|
; shuffle the rest of the data, and write it with 2 mmword writes
|
819 |
|
|
movq mmword ptr [esi+8*11], mm5 ; tmt11
|
820 |
|
|
; moved up from the M3 transpose
|
821 |
|
|
punpcklwd mm0, mm1
|
822 |
|
|
|
823 |
|
|
movq mmword ptr [esi+8*15], mm2 ; tmt15
|
824 |
|
|
; moved up from the M3 transpose
|
825 |
|
|
punpckhwd mm3, mm1
|
826 |
|
|
|
827 |
|
|
; transpose - M3 part
|
828 |
|
|
|
829 |
|
|
; moved up to previous code section
|
830 |
|
|
;movq mm0, mmword ptr [esi+8*8]
|
831 |
|
|
;movq mm1, mmword ptr [esi+8*10]
|
832 |
|
|
;movq mm3, mm0
|
833 |
|
|
;punpcklwd mm0, mm1
|
834 |
|
|
;punpckhwd mm3, mm1
|
835 |
|
|
|
836 |
|
|
movq mm6, mmword ptr [esi+8*12]
|
837 |
|
|
;slot
|
838 |
|
|
|
839 |
|
|
movq mm4, mmword ptr [esi+8*14]
|
840 |
|
|
movq mm2, mm6
|
841 |
|
|
|
842 |
|
|
; shuffle the data and write out the lower parts of the transposed in 4 dwords
|
843 |
|
|
punpcklwd mm6, mm4
|
844 |
|
|
movq mm1, mm0
|
845 |
|
|
|
846 |
|
|
punpckhdq mm1, mm6
|
847 |
|
|
movq mm7, mm3
|
848 |
|
|
|
849 |
|
|
punpckhwd mm2, mm4 ; free mm4
|
850 |
|
|
;slot
|
851 |
|
|
|
852 |
|
|
punpckldq mm0, mm6 ; free mm6
|
853 |
|
|
;slot
|
854 |
|
|
|
855 |
|
|
;moved from next block
|
856 |
|
|
movq mm4, mmword ptr [esi+8*13] ; tmt13
|
857 |
|
|
punpckldq mm3, mm2
|
858 |
|
|
|
859 |
|
|
punpckhdq mm7, mm2 ; free mm2
|
860 |
|
|
;moved from next block
|
861 |
|
|
movq mm5, mm3 ; duplicate tmt5
|
862 |
|
|
|
863 |
|
|
; column 1: even part (after transpose)
|
864 |
|
|
|
865 |
|
|
;moved above
|
866 |
|
|
;movq mm5, mm3 ; duplicate tmt5
|
867 |
|
|
;movq mm4, mmword ptr [esi+8*13] ; tmt13
|
868 |
|
|
|
869 |
|
|
psubsw mm3, mm4 ; V134
|
870 |
|
|
;slot
|
871 |
|
|
|
872 |
|
|
pmulhw mm3, mmword ptr x5a825a825a825a82 ; 23170 ->V136
|
873 |
|
|
;slot
|
874 |
|
|
|
875 |
|
|
movq mm6, mmword ptr [esi+8*9] ; tmt9
|
876 |
|
|
paddsw mm5, mm4 ; V135 ; mm4 free
|
877 |
|
|
|
878 |
|
|
movq mm4, mm0 ; duplicate tmt1
|
879 |
|
|
paddsw mm0, mm6 ; V137
|
880 |
|
|
|
881 |
|
|
psubsw mm4, mm6 ; V138 ; mm6 free
|
882 |
|
|
psllw mm3, 2 ; t290
|
883 |
|
|
psubsw mm3, mm5 ; V139
|
884 |
|
|
movq mm6, mm0 ; duplicate V137
|
885 |
|
|
|
886 |
|
|
paddsw mm0, mm5 ; V140
|
887 |
|
|
movq mm2, mm4 ; duplicate V138
|
888 |
|
|
|
889 |
|
|
paddsw mm2, mm3 ; V141
|
890 |
|
|
psubsw mm4, mm3 ; V142 ; mm3 free
|
891 |
|
|
|
892 |
|
|
movq mmword ptr [esi+8*9], mm0 ; V140
|
893 |
|
|
psubsw mm6, mm5 ; V143 ; mm5 free
|
894 |
|
|
|
895 |
|
|
;moved from next block
|
896 |
|
|
movq mm0, mmword ptr[esi+8*11] ; tmt11
|
897 |
|
|
;slot
|
898 |
|
|
|
899 |
|
|
movq mmword ptr [esi+8*13], mm2 ; V141
|
900 |
|
|
;moved from next block
|
901 |
|
|
movq mm2, mm0 ; duplicate tmt11
|
902 |
|
|
|
903 |
|
|
; column 1: odd part (after transpose)
|
904 |
|
|
|
905 |
|
|
;moved up to the prev block
|
906 |
|
|
;movq mm0, mmword ptr[esi+8*11] ; tmt11
|
907 |
|
|
;movq mm2, mm0 ; duplicate tmt11
|
908 |
|
|
|
909 |
|
|
movq mm5, mmword ptr[esi+8*15] ; tmt15
|
910 |
|
|
psubsw mm0, mm7 ; V144
|
911 |
|
|
|
912 |
|
|
movq mm3, mm0 ; duplicate V144
|
913 |
|
|
paddsw mm2, mm7 ; V147 ; free mm7
|
914 |
|
|
|
915 |
|
|
pmulhw mm0, mmword ptr x539f539f539f539f ; 21407-> V151
|
916 |
|
|
movq mm7, mm1 ; duplicate tmt3
|
917 |
|
|
|
918 |
|
|
paddsw mm7, mm5 ; V145
|
919 |
|
|
psubsw mm1, mm5 ; V146 ; free mm5
|
920 |
|
|
|
921 |
|
|
psubsw mm3, mm1 ; V150
|
922 |
|
|
movq mm5, mm7 ; duplicate V145
|
923 |
|
|
|
924 |
|
|
pmulhw mm1, mmword ptr x4546454645464546 ; 17734-> V153
|
925 |
|
|
psubsw mm5, mm2 ; V148
|
926 |
|
|
|
927 |
|
|
pmulhw mm3, mmword ptr x61f861f861f861f8 ; 25080-> V154
|
928 |
|
|
psllw mm0, 2 ; t311
|
929 |
|
|
|
930 |
|
|
pmulhw mm5, mmword ptr x5a825a825a825a82 ; 23170-> V152
|
931 |
|
|
paddsw mm7, mm2 ; V149 ; free mm2
|
932 |
|
|
|
933 |
|
|
psllw mm1, 1 ; t313
|
934 |
|
|
nop ; slot
|
935 |
|
|
|
936 |
|
|
;without the nop above - freeze here for one clock
|
937 |
|
|
;the nop cleans the mess a little bit
|
938 |
|
|
movq mm2, mm3 ; duplicate V154
|
939 |
|
|
psubsw mm3, mm0 ; V155 ; free mm0
|
940 |
|
|
|
941 |
|
|
psubsw mm1, mm2 ; V156 ; free mm2
|
942 |
|
|
;moved from the next block
|
943 |
|
|
movq mm2, mm6 ; duplicate V143
|
944 |
|
|
|
945 |
|
|
;moved from the next block
|
946 |
|
|
movq mm0, mmword ptr[esi+8*13] ; V141
|
947 |
|
|
psllw mm1, 1 ; t315
|
948 |
|
|
|
949 |
|
|
psubsw mm1, mm7 ; V157 (keep V149)
|
950 |
|
|
psllw mm5, 2 ; t317
|
951 |
|
|
|
952 |
|
|
psubsw mm5, mm1 ; V158
|
953 |
|
|
psllw mm3, 1 ; t319
|
954 |
|
|
|
955 |
|
|
paddsw mm3, mm5 ; V159
|
956 |
|
|
;slot
|
957 |
|
|
|
958 |
|
|
; column 1: output butterfly (after transform)
|
959 |
|
|
;moved to the prev block
|
960 |
|
|
;movq mm2, mm6 ; duplicate V143
|
961 |
|
|
;movq mm0, mmword ptr[esi+8*13] ; V141
|
962 |
|
|
|
963 |
|
|
psubsw mm2, mm3 ; V163
|
964 |
|
|
paddsw mm6, mm3 ; V164 ; free mm3
|
965 |
|
|
|
966 |
|
|
movq mm3, mm4 ; duplicate V142
|
967 |
|
|
psubsw mm4, mm5 ; V165 ; free mm5
|
968 |
|
|
|
969 |
|
|
movq mmword ptr scratch7, mm2 ; out7
|
970 |
|
|
psraw mm6, 4
|
971 |
|
|
|
972 |
|
|
psraw mm4, 4
|
973 |
|
|
paddsw mm3, mm5 ; V162
|
974 |
|
|
|
975 |
|
|
movq mm2, mmword ptr[esi+8*9] ; V140
|
976 |
|
|
movq mm5, mm0 ; duplicate V141
|
977 |
|
|
|
978 |
|
|
;in order not to perculate this line up, we read [esi+8*9] very near to this location
|
979 |
|
|
movq mmword ptr [esi+8*9], mm6 ; out9
|
980 |
|
|
paddsw mm0, mm1 ; V161
|
981 |
|
|
|
982 |
|
|
movq mmword ptr scratch5, mm3 ; out5
|
983 |
|
|
psubsw mm5, mm1 ; V166 ; free mm1
|
984 |
|
|
|
985 |
|
|
movq mmword ptr[esi+8*11], mm4 ; out11
|
986 |
|
|
psraw mm5, 4
|
987 |
|
|
|
988 |
|
|
movq mmword ptr scratch3, mm0 ; out3
|
989 |
|
|
movq mm4, mm2 ; duplicate V140
|
990 |
|
|
|
991 |
|
|
movq mmword ptr[esi+8*13], mm5 ; out13
|
992 |
|
|
paddsw mm2, mm7 ; V160
|
993 |
|
|
|
994 |
|
|
;moved from the next block
|
995 |
|
|
movq mm0, mmword ptr [esi+8*1]
|
996 |
|
|
psubsw mm4, mm7 ; V167 ; free mm7
|
997 |
|
|
|
998 |
|
|
;moved from the next block
|
999 |
|
|
movq mm7, mmword ptr [esi+8*3]
|
1000 |
|
|
psraw mm4, 4
|
1001 |
|
|
|
1002 |
|
|
movq mmword ptr scratch1, mm2 ; out1
|
1003 |
|
|
;moved from the next block
|
1004 |
|
|
movq mm1, mm0
|
1005 |
|
|
|
1006 |
|
|
movq mmword ptr[esi+8*15], mm4 ; out15
|
1007 |
|
|
;moved from the next block
|
1008 |
|
|
punpcklwd mm0, mm7
|
1009 |
|
|
|
1010 |
|
|
; transpose - M2 parts
|
1011 |
|
|
;moved up to the prev block
|
1012 |
|
|
;movq mm0, mmword ptr [esi+8*1]
|
1013 |
|
|
;movq mm7, mmword ptr [esi+8*3]
|
1014 |
|
|
;movq mm1, mm0
|
1015 |
|
|
;punpcklwd mm0, mm7
|
1016 |
|
|
|
1017 |
|
|
movq mm5, mmword ptr [esi+8*5]
|
1018 |
|
|
punpckhwd mm1, mm7
|
1019 |
|
|
|
1020 |
|
|
movq mm4, mmword ptr [esi+8*7]
|
1021 |
|
|
movq mm3, mm5
|
1022 |
|
|
|
1023 |
|
|
; shuffle the data and write out the lower parts of the trasposed in 4 dwords
|
1024 |
|
|
movdf dword ptr [esi+8*8], mm0 ; LS part of tmt8
|
1025 |
|
|
punpcklwd mm5, mm4
|
1026 |
|
|
|
1027 |
|
|
movdf dword ptr [esi+8*12], mm1 ; LS part of tmt12
|
1028 |
|
|
punpckhwd mm3, mm4
|
1029 |
|
|
|
1030 |
|
|
movdf dword ptr [esi+8*8+4], mm5 ; MS part of tmt8
|
1031 |
|
|
punpckhdq mm0, mm5 ; tmt10
|
1032 |
|
|
|
1033 |
|
|
movdf dword ptr [esi+8*12+4], mm3 ; MS part of tmt12
|
1034 |
|
|
punpckhdq mm1, mm3 ; tmt14
|
1035 |
|
|
|
1036 |
|
|
; transpose - M1 parts
|
1037 |
|
|
movq mm7, mmword ptr [esi]
|
1038 |
|
|
;slot
|
1039 |
|
|
|
1040 |
|
|
movq mm2, mmword ptr [esi+8*2]
|
1041 |
|
|
movq mm6, mm7
|
1042 |
|
|
|
1043 |
|
|
movq mm5, mmword ptr [esi+8*4]
|
1044 |
|
|
punpcklwd mm7, mm2
|
1045 |
|
|
|
1046 |
|
|
movq mm4, mmword ptr [esi+8*6]
|
1047 |
|
|
punpckhwd mm6, mm2 ; free mm2
|
1048 |
|
|
|
1049 |
|
|
movq mm3, mm5
|
1050 |
|
|
punpcklwd mm5, mm4
|
1051 |
|
|
|
1052 |
|
|
punpckhwd mm3, mm4 ; free mm4
|
1053 |
|
|
movq mm2, mm7
|
1054 |
|
|
|
1055 |
|
|
movq mm4, mm6
|
1056 |
|
|
punpckldq mm7, mm5 ; tmt0
|
1057 |
|
|
|
1058 |
|
|
punpckhdq mm2, mm5 ; tmt2 ; free mm5
|
1059 |
|
|
;slot
|
1060 |
|
|
|
1061 |
|
|
; shuffle the rest of the data, and write it with 2 mmword writes
|
1062 |
|
|
punpckldq mm6, mm3 ; tmt4
|
1063 |
|
|
;move from next block
|
1064 |
|
|
movq mm5, mm2 ; duplicate tmt2
|
1065 |
|
|
|
1066 |
|
|
punpckhdq mm4, mm3 ; tmt6 ; free mm3
|
1067 |
|
|
;move from next block
|
1068 |
|
|
movq mm3, mm0 ; duplicate tmt10
|
1069 |
|
|
|
1070 |
|
|
; column 0: odd part (after transpose)
|
1071 |
|
|
;moved up to prev block
|
1072 |
|
|
;movq mm3, mm0 ; duplicate tmt10
|
1073 |
|
|
;movq mm5, mm2 ; duplicate tmt2
|
1074 |
|
|
|
1075 |
|
|
psubsw mm0, mm4 ; V110
|
1076 |
|
|
paddsw mm3, mm4 ; V113 ; free mm4
|
1077 |
|
|
|
1078 |
|
|
movq mm4, mm0 ; duplicate V110
|
1079 |
|
|
paddsw mm2, mm1 ; V111
|
1080 |
|
|
|
1081 |
|
|
pmulhw mm0, mmword ptr x539f539f539f539f ; 21407-> V117
|
1082 |
|
|
psubsw mm5, mm1 ; V112 ; free mm1
|
1083 |
|
|
|
1084 |
|
|
psubsw mm4, mm5 ; V116
|
1085 |
|
|
movq mm1, mm2 ; duplicate V111
|
1086 |
|
|
|
1087 |
|
|
pmulhw mm5, mmword ptr x4546454645464546 ; 17734-> V119
|
1088 |
|
|
psubsw mm2, mm3 ; V114
|
1089 |
|
|
|
1090 |
|
|
pmulhw mm4, mmword ptr x61f861f861f861f8 ; 25080-> V120
|
1091 |
|
|
paddsw mm1, mm3 ; V115 ; free mm3
|
1092 |
|
|
|
1093 |
|
|
pmulhw mm2, mmword ptr x5a825a825a825a82 ; 23170-> V118
|
1094 |
|
|
psllw mm0, 2 ; t266
|
1095 |
|
|
|
1096 |
|
|
movq mmword ptr[esi+8*0], mm1 ; save V115
|
1097 |
|
|
psllw mm5, 1 ; t268
|
1098 |
|
|
|
1099 |
|
|
psubsw mm5, mm4 ; V122
|
1100 |
|
|
psubsw mm4, mm0 ; V121 ; free mm0
|
1101 |
|
|
|
1102 |
|
|
psllw mm5, 1 ; t270
|
1103 |
|
|
;slot
|
1104 |
|
|
|
1105 |
|
|
psubsw mm5, mm1 ; V123 ; free mm1
|
1106 |
|
|
psllw mm2, 2 ; t272
|
1107 |
|
|
|
1108 |
|
|
psubsw mm2, mm5 ; V124 (keep V123)
|
1109 |
|
|
psllw mm4, 1 ; t274
|
1110 |
|
|
|
1111 |
|
|
movq mmword ptr[esi+8*2], mm5 ; save V123 ; free mm5
|
1112 |
|
|
paddsw mm4, mm2 ; V125 (keep V124)
|
1113 |
|
|
|
1114 |
|
|
; column 0: even part (after transpose)
|
1115 |
|
|
movq mm0, mmword ptr[esi+8*12] ; tmt12
|
1116 |
|
|
movq mm3, mm6 ; duplicate tmt4
|
1117 |
|
|
|
1118 |
|
|
psubsw mm6, mm0 ; V100
|
1119 |
|
|
paddsw mm3, mm0 ; V101 ; free mm0
|
1120 |
|
|
|
1121 |
|
|
pmulhw mm6, mmword ptr x5a825a825a825a82 ; 23170 ->V102
|
1122 |
|
|
movq mm5, mm7 ; duplicate tmt0
|
1123 |
|
|
|
1124 |
|
|
movq mm1, mmword ptr[esi+8*8] ; tmt8
|
1125 |
|
|
;slot
|
1126 |
|
|
|
1127 |
|
|
paddsw mm7, mm1 ; V103
|
1128 |
|
|
psubsw mm5, mm1 ; V104 ; free mm1
|
1129 |
|
|
|
1130 |
|
|
movq mm0, mm7 ; duplicate V103
|
1131 |
|
|
psllw mm6, 2 ; t245
|
1132 |
|
|
|
1133 |
|
|
paddsw mm7, mm3 ; V106
|
1134 |
|
|
movq mm1, mm5 ; duplicate V104
|
1135 |
|
|
|
1136 |
|
|
psubsw mm6, mm3 ; V105
|
1137 |
|
|
psubsw mm0, mm3 ; V109; free mm3
|
1138 |
|
|
|
1139 |
|
|
paddsw mm5, mm6 ; V107
|
1140 |
|
|
psubsw mm1, mm6 ; V108 ; free mm6
|
1141 |
|
|
|
1142 |
|
|
; column 0: output butterfly (after transform)
|
1143 |
|
|
movq mm3, mm1 ; duplicate V108
|
1144 |
|
|
paddsw mm1, mm2 ; out4
|
1145 |
|
|
|
1146 |
|
|
psraw mm1, 4
|
1147 |
|
|
psubsw mm3, mm2 ; out10 ; free mm2
|
1148 |
|
|
|
1149 |
|
|
psraw mm3, 4
|
1150 |
|
|
movq mm6, mm0 ; duplicate V109
|
1151 |
|
|
|
1152 |
|
|
movq mmword ptr[esi+8*4], mm1 ; out4 ; free mm1
|
1153 |
|
|
psubsw mm0, mm4 ; out6
|
1154 |
|
|
|
1155 |
|
|
movq mmword ptr[esi+8*10], mm3 ; out10 ; free mm3
|
1156 |
|
|
psraw mm0, 4
|
1157 |
|
|
|
1158 |
|
|
paddsw mm6, mm4 ; out8 ; free mm4
|
1159 |
|
|
movq mm1, mm7 ; duplicate V106
|
1160 |
|
|
|
1161 |
|
|
movq mmword ptr[esi+8*6], mm0 ; out6 ; free mm0
|
1162 |
|
|
psraw mm6, 4
|
1163 |
|
|
|
1164 |
|
|
movq mm4, mmword ptr[esi+8*0] ; V115
|
1165 |
|
|
;slot
|
1166 |
|
|
|
1167 |
|
|
movq mmword ptr[esi+8*8], mm6 ; out8 ; free mm6
|
1168 |
|
|
movq mm2, mm5 ; duplicate V107
|
1169 |
|
|
|
1170 |
|
|
movq mm3, mmword ptr[esi+8*2] ; V123
|
1171 |
|
|
paddsw mm7, mm4 ; out0
|
1172 |
|
|
|
1173 |
|
|
;moved up from next block
|
1174 |
|
|
movq mm0, mmword ptr scratch3
|
1175 |
|
|
psraw mm7, 4
|
1176 |
|
|
|
1177 |
|
|
;moved up from next block
|
1178 |
|
|
movq mm6, mmword ptr scratch5
|
1179 |
|
|
psubsw mm1, mm4 ; out14 ; free mm4
|
1180 |
|
|
|
1181 |
|
|
paddsw mm5, mm3 ; out2
|
1182 |
|
|
psraw mm1, 4
|
1183 |
|
|
|
1184 |
|
|
movq mmword ptr[esi], mm7 ; out0 ; free mm7
|
1185 |
|
|
psraw mm5, 4
|
1186 |
|
|
|
1187 |
|
|
movq mmword ptr[esi+8*14], mm1 ; out14 ; free mm1
|
1188 |
|
|
psubsw mm2, mm3 ; out12 ; free mm3
|
1189 |
|
|
|
1190 |
|
|
movq mmword ptr[esi+8*2], mm5 ; out2 ; free mm5
|
1191 |
|
|
psraw mm2, 4
|
1192 |
|
|
|
1193 |
|
|
;moved up to the prev block
|
1194 |
|
|
movq mm4, mmword ptr scratch7
|
1195 |
|
|
;moved up to the prev block
|
1196 |
|
|
psraw mm0, 4
|
1197 |
|
|
|
1198 |
|
|
movq mmword ptr[esi+8*12], mm2 ; out12 ; free mm2
|
1199 |
|
|
;moved up to the prev block
|
1200 |
|
|
psraw mm6, 4
|
1201 |
|
|
|
1202 |
|
|
;move back the data to its correct place
|
1203 |
|
|
;moved up to the prev block
|
1204 |
|
|
;movq mm0, mmword ptr scratch3
|
1205 |
|
|
;movq mm6, mmword ptr scratch5
|
1206 |
|
|
;movq mm4, mmword ptr scratch7
|
1207 |
|
|
;psraw mm0, 4
|
1208 |
|
|
;psraw mm6, 4
|
1209 |
|
|
|
1210 |
|
|
movq mm1, mmword ptr scratch1
|
1211 |
|
|
psraw mm4, 4
|
1212 |
|
|
|
1213 |
|
|
movq mmword ptr [esi+8*3], mm0 ; out3
|
1214 |
|
|
psraw mm1, 4
|
1215 |
|
|
|
1216 |
|
|
movq mmword ptr [esi+8*5], mm6 ; out5
|
1217 |
|
|
;slot
|
1218 |
|
|
|
1219 |
|
|
movq mmword ptr [esi+8*7], mm4 ; out7
|
1220 |
|
|
;slot
|
1221 |
|
|
|
1222 |
|
|
movq mmword ptr [esi+8*1], mm1 ; out1
|
1223 |
|
|
;slot
|
1224 |
|
|
|
1225 |
|
|
emms
|
1226 |
|
|
|
1227 |
|
|
pop esi
|
1228 |
|
|
pop ebp
|
1229 |
|
|
|
1230 |
|
|
ret 0
|
1231 |
|
|
|
1232 |
|
|
_idct8x8aan ENDP
|
1233 |
|
|
_TEXT ENDS
|
1234 |
|
|
|
1235 |
|
|
END
|
1236 |
|
|
|
1237 |
|
|
* Legal Information © 1998 Intel Corporation
|