OpenCores
URL https://opencores.org/ocsvn/bluespec-h264/bluespec-h264/trunk

Subversion Repositories bluespec-h264

[/] [bluespec-h264/] [trunk/] [LumaChromaParallel/] [mkInterpolator.bsv] - Blame information for rev 91

Go to most recent revision | Details | Compare with Previous | View Log

Line No. Rev Author Line
1 90 jamey.hick 
// The MIT License
2 
 
3 
// Copyright (c) 2006-2007 Massachusetts Institute of Technology
4 
 
5 
// Permission is hereby granted, free of charge, to any person obtaining a copy
6 
// of this software and associated documentation files (the "Software"), to deal
7 
// in the Software without restriction, including without limitation the rights
8 
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9 
// copies of the Software, and to permit persons to whom the Software is
10 
// furnished to do so, subject to the following conditions:
11 
 
12 
// The above copyright notice and this permission notice shall be included in
13 
// all copies or substantial portions of the Software.
14 
 
15 
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
21 
// THE SOFTWARE.
22 
//**********************************************************************
23 
// interpolator implementation
24 
//----------------------------------------------------------------------
25 
//
26 
//
27 
 
28 
package mkInterpolator;
29 
 
30 
import H264Types::*;
31 
import IInterpolator::*;
32 
import FIFO::*;
33 91 jamey.hick 
import FIFOF::*;
34 90 jamey.hick 
import Vector::*;
35 
 
36 
import Connectable::*;
37 
import GetPut::*;
38 
import ClientServer::*;
39 
 
40 
 
41 
//-----------------------------------------------------------
42 
// Local Datatypes
43 
//-----------------------------------------------------------
44 
 
45 
typedef union tagged
46 
{
47 
 struct { Bit#(2) xFracL; Bit#(2) yFracL; Bit#(2) offset; IPBlockType bt; } IPWLuma;
48 
 struct { Bit#(3) xFracC; Bit#(3) yFracC; Bit#(2) offset; IPBlockType bt; } IPWChroma;
49 
}
50 
InterpolatorWT deriving(Eq,Bits);
51 
 
52 
 
53 
//-----------------------------------------------------------
54 
// Helper functions
55 
 
56 
function Bit#(8) clip1y10to8( Bit#(10) innum );
57 
   if(innum[9] == 1)
58 
      return 0;
59 
   else if(innum[8] == 1)
60 
      return 255;
61 
   else
62 
      return truncate(innum);
63 
endfunction
64 
 
65 
function Bit#(15) interpolate8to15( Bit#(8) in0, Bit#(8) in1, Bit#(8) in2, Bit#(8) in3, Bit#(8) in4, Bit#(8) in5 );
66 
   return zeroExtend(in0) - 5*zeroExtend(in1) + 20*zeroExtend(in2) + 20*zeroExtend(in3) - 5*zeroExtend(in4) + zeroExtend(in5);
67 
endfunction
68 
 
69 
function Bit#(8) interpolate15to8( Bit#(15) in0, Bit#(15) in1, Bit#(15) in2, Bit#(15) in3, Bit#(15) in4, Bit#(15) in5 );
70 
   Bit#(20) temp = signExtend(in0) - 5*signExtend(in1) + 20*signExtend(in2) + 20*signExtend(in3) - 5*signExtend(in4) + signExtend(in5) + 512;
71 
   return clip1y10to8(truncate(temp>>10));
72 
endfunction
73 
 
74 
 
75 
 
76 
//-----------------------------------------------------------
77 
// Interpolation Module
78 
//-----------------------------------------------------------
79 
 
80 
 
81 
(* synthesize *)
82 
module mkInterpolator( Interpolator );
83 
 
84 
   FIFO#(InterpolatorIT) reqfifoLoad <- mkSizedFIFO(interpolator_reqfifoLoad_size);
85 
   FIFO#(InterpolatorWT) reqfifoWork1 <- mkSizedFIFO(interpolator_reqfifoWork_size);
86 
   Reg#(Maybe#(InterpolatorWT)) reqregWork2 <- mkReg(Invalid);
87 
   FIFO#(Vector#(4,Bit#(8))) outfifo <- mkFIFO;
88 
   Reg#(Bool) endOfFrameFlag <- mkReg(False);
89 
   FIFO#(InterpolatorLoadReq)  memReqQ  <- mkFIFO;
90 
   FIFO#(InterpolatorLoadResp) memRespQ <- mkSizedFIFO(interpolator_memRespQ_size);
91 
 
92 
   Reg#(Bit#(PicWidthSz))  picWidth  <- mkReg(maxPicWidthInMB);
93 
   Reg#(Bit#(PicHeightSz)) picHeight <- mkReg(0);
94 
 
95 
   RFile1#(Bit#(6),Vector#(4,Bit#(15))) workFile  <- mkRFile1Full();
96 
   RFile1#(Bit#(6),Vector#(4,Bit#(8)))  storeFile <- mkRFile1Full();
97 
   Reg#(Bit#(1)) workFileFlag <- mkReg(0);
98 
   RFile1#(Bit#(4),Vector#(4,Bit#(8))) resultFile <- mkRFile1Full();
99 
 
100 
   Reg#(Bit#(1)) loadStage  <- mkReg(0);
101 
   Reg#(Bit#(2)) loadHorNum <- mkReg(0);
102 
   Reg#(Bit#(4)) loadVerNum <- mkReg(0);
103 
 
104 
   Reg#(Bit#(2)) work1MbPart    <- mkReg(0);//only for Chroma
105 
   Reg#(Bit#(2)) work1SubMbPart <- mkReg(0);//only for Chroma
106 
   Reg#(Bit#(1)) work1Stage     <- mkReg(0);
107 
   Reg#(Bit#(2)) work1HorNum    <- mkReg(0);
108 
   Reg#(Bit#(4)) work1VerNum    <- mkReg(0);
109 
   Reg#(Vector#(20,Bit#(8))) work1Vector8 <- mkRegU;
110 
   Reg#(Bool) work1Done <- mkReg(False);
111 
 
112 
   Reg#(Bit#(2)) work2SubMbPart <- mkReg(0);
113 
   Reg#(Bit#(2)) work2HorNum    <- mkReg(0);
114 
   Reg#(Bit#(4)) work2VerNum    <- mkReg(0);
115 
   Reg#(Vector#(20,Bit#(8))) work2Vector8 <- mkRegU;
116 
   Reg#(Vector#(20,Bit#(15))) work2Vector15 <- mkRegU;
117 
   Reg#(Vector#(16,Bit#(1))) resultReady <- mkReg(replicate(0));
118 
   Reg#(Bool) work2Done <- mkReg(False);
119 
   Reg#(Bool) work8x8Done <- mkReg(False);
120 
 
121 
   Reg#(Bit#(2)) outBlockNum <- mkReg(0);
122 
   Reg#(Bit#(2)) outPixelNum <- mkReg(0);
123 
   Reg#(Bool) outDone <- mkReg(False);
124 
 
125 91 jamey.hick 
   Reg#(Bit#(64)) total_cycles <- mkReg(0);
126 
 
127 
 
128 
   rule incr;
129 
     total_cycles <= total_cycles + 1;
130 
   endrule
131 
 
132 90 jamey.hick 
   rule sendEndOfFrameReq( endOfFrameFlag );
133 
      endOfFrameFlag <= False;
134 91 jamey.hick 
      $display("trace interpolator sending EndOfFrame");
135 90 jamey.hick 
      memReqQ.enq(IPLoadEndFrame);
136 
   endrule
137 
 
138 
 
139 
   rule loadLuma( reqfifoLoad.first() matches tagged IPLuma .reqdata &&& !endOfFrameFlag );
140 
      Bit#(2) xfracl = reqdata.mvhor[1:0];
141 
      Bit#(2) yfracl = reqdata.mvver[1:0];
142 
      Bit#(2) offset = reqdata.mvhor[3:2];
143 
      Bool twoStage = (xfracl==1||xfracl==3) && (yfracl==1||yfracl==3);
144 
      Bool horInter = (twoStage ? loadStage==1 : xfracl!=0);
145 
      Bool verInter = (twoStage ? loadStage==0 : yfracl!=0);
146 
      Bit#(2) offset2 = reqdata.mvhor[3:2] + ((twoStage&&verInter&&xfracl==3) ? 1 : 0);
147 
      Bit#(1) horOut = 0;
148 
      Bit#(TAdd#(PicWidthSz,2)) horAddr;
149 
      Bit#(TAdd#(PicHeightSz,4)) verAddr;
150 
      Bit#(TAdd#(PicWidthSz,12)) horTemp = zeroExtend({reqdata.hor,2'b00}) + zeroExtend({loadHorNum,2'b00}) + (xfracl==3&&(yfracl==1||yfracl==3)&&loadStage==0 ? 1 : 0);
151 
      Bit#(TAdd#(PicHeightSz,10)) verTemp = zeroExtend(reqdata.ver) + zeroExtend(loadVerNum) + (yfracl==3&&(xfracl==1||xfracl==3)&&loadStage==1 ? 1 : 0);
152 
      Bit#(13) mvhortemp = signExtend(reqdata.mvhor[13:2])-(horInter?2:0);
153 
      Bit#(11) mvvertemp = signExtend(reqdata.mvver[11:2])-(verInter?2:0);
154 
      if(mvhortemp[12]==1 && zeroExtend(0-mvhortemp)>horTemp)
155 
         begin
156 
            horAddr = 0;
157 
            horOut = 1;
158 
         end
159 
      else
160 
         begin
161 
            horTemp = horTemp + signExtend(mvhortemp);
162 
            if(horTemp>=zeroExtend({picWidth,4'b0000}))
163 
               begin
164 
                  horAddr = {picWidth-1,2'b11};
165 
                  horOut = 1;
166 
               end
167 
            else
168 
               horAddr = truncate(horTemp>>2);
169 
         end
170 
      if(mvvertemp[10]==1 && zeroExtend(0-mvvertemp)>verTemp)
171 
         verAddr = 0;
172 
      else
173 
         begin
174 
            verTemp = verTemp + signExtend(mvvertemp);
175 
            if(verTemp>=zeroExtend({picHeight,4'b0000}))
176 
               verAddr = {picHeight-1,4'b1111};
177 
            else
178 
               verAddr = truncate(verTemp);
179 
         end
180 91 jamey.hick 
      $display("PARDEBLOCK issuing Luma memory req");
181 90 jamey.hick 
      memReqQ.enq(IPLoadLuma {refIdx:reqdata.refIdx,horOutOfBounds:horOut,hor:horAddr,ver:verAddr});
182 
      Bool verFirst = twoStage || (yfracl==2&&(xfracl==1||xfracl==3));
183 
      Bit#(2) loadHorNumMax = (reqdata.bt==IP8x8||reqdata.bt==IP8x4 ? 1 : 0) + (horInter ? 2 : (offset2==0 ? 0 : 1));
184 
      Bit#(4) loadVerNumMax = (reqdata.bt==IP8x8||reqdata.bt==IP4x8 ? 7 : 3) + (verInter ? 5 : 0);
185 
      if(verFirst)
186 
         begin
187 
            if(loadVerNum < loadVerNumMax)
188 
               loadVerNum <= loadVerNum+1;
189 
            else
190 
               begin
191 
                  loadVerNum <= 0;
192 
                  if(loadHorNum < loadHorNumMax)
193 
                     begin
194 
                        if(loadStage == 1)
195 
                           begin
196 
                              offset = offset + (xfracl==3 ? 1 : 0);
197 
                              if(!(offset==1 || (xfracl==3 && offset==2)))
198 
                                 loadHorNum <= loadHorNumMax;
199 
                              else
200 
                                 begin
201 
                                    loadHorNum <= 0;
202 
                                    loadStage <= 0;
203 
                                    reqfifoLoad.deq();
204 
                                 end
205 
                           end
206 
                        else
207 
                           loadHorNum <= loadHorNum+1;
208 
                     end
209 
                  else
210 
                     begin
211 
                        if(twoStage && loadStage==0)
212 
                           begin
213 
                              offset = offset + (xfracl==3 ? 1 : 0);
214 
                              if((xfracl==3 ? offset<3 : offset<2))
215 
                                 loadHorNum <= 0;
216 
                              else
217 
                                 loadHorNum <= loadHorNumMax+1;
218 
                              loadStage <= 1;
219 
                           end
220 
                        else
221 
                           begin
222 
                              loadHorNum <= 0;
223 
                              loadStage <= 0;
224 
                              reqfifoLoad.deq();
225 
                           end
226 
                     end
227 
               end
228 
         end
229 
      else
230 
         begin
231 
            if(loadHorNum < loadHorNumMax)
232 
               loadHorNum <= loadHorNum+1;
233 
            else
234 
               begin
235 
                  loadHorNum <= 0;
236 
                  if(loadVerNum < loadVerNumMax)
237 
                     loadVerNum <= loadVerNum+1;
238 
                  else
239 
                     begin
240 
                        loadVerNum <= 0;
241 
                        reqfifoLoad.deq();
242 
                     end
243 
               end
244 
         end
245 
      if(reqdata.bt==IP16x16 || reqdata.bt==IP16x8 || reqdata.bt==IP8x16)
246 
         $display( "ERROR Interpolation: loadLuma block sizes > 8x8 not supported");
247 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator(%0d)): loadLuma %h %h %h %h %h %h %h", total_cycles,xfracl, yfracl, loadHorNum, loadVerNum, reqdata.refIdx, horAddr, verAddr);
248 90 jamey.hick 
   endrule
249 
 
250 
 
251 
   rule loadChroma( reqfifoLoad.first() matches tagged IPChroma .reqdata &&& !endOfFrameFlag );
252 
      Bit#(3) xfracc = reqdata.mvhor[2:0];
253 
      Bit#(3) yfracc = reqdata.mvver[2:0];
254 
      Bit#(2) offset = reqdata.mvhor[4:3]+{reqdata.hor[0],1'b0};
255 
      Bit#(1) horOut = 0;
256 
      Bit#(TAdd#(PicWidthSz,1)) horAddr;
257 
      Bit#(TAdd#(PicHeightSz,3)) verAddr;
258 
      Bit#(TAdd#(PicWidthSz,11)) horTemp = zeroExtend({reqdata.hor,1'b0}) + zeroExtend({loadHorNum,2'b00});
259 
      Bit#(TAdd#(PicHeightSz,9)) verTemp = zeroExtend(reqdata.ver) + zeroExtend(loadVerNum);
260 
      if(reqdata.mvhor[13]==1 && zeroExtend(0-reqdata.mvhor[13:3])>horTemp)
261 
         begin
262 
            horAddr = 0;
263 
            horOut = 1;
264 
         end
265 
      else
266 
         begin
267 
            horTemp = horTemp + signExtend(reqdata.mvhor[13:3]);
268 
            if(horTemp>=zeroExtend({picWidth,3'b000}))
269 
               begin
270 
                  horAddr = {picWidth-1,1'b1};
271 
                  horOut = 1;
272 
               end
273 
            else
274 
               horAddr = truncate(horTemp>>2);
275 
         end
276 
      if(reqdata.mvver[11]==1 && zeroExtend(0-reqdata.mvver[11:3])>verTemp)
277 
         verAddr = 0;
278 
      else
279 
         begin
280 
            verTemp = verTemp + signExtend(reqdata.mvver[11:3]);
281 
            if(verTemp>=zeroExtend({picHeight,3'b000}))
282 
               verAddr = {picHeight-1,3'b111};
283 
            else
284 
               verAddr = truncate(verTemp);
285 
         end
286 91 jamey.hick 
      $display("PARDEBLOCK issuing Chroma memory req");
287 90 jamey.hick 
      memReqQ.enq(IPLoadChroma {refIdx:reqdata.refIdx,uv:reqdata.uv,horOutOfBounds:horOut,hor:horAddr,ver:verAddr});
288 
      Bit#(2) loadHorNumMax = (reqdata.bt==IP4x8||reqdata.bt==IP4x4 ? (offset[1]==0||(xfracc==0&&offset!=3) ? 0 : 1) : ((reqdata.bt==IP16x16||reqdata.bt==IP16x8 ? 1 : 0) + (xfracc==0&&offset==0 ? 0 : 1)));
289 
      Bit#(4) loadVerNumMax = (reqdata.bt==IP16x16||reqdata.bt==IP8x16 ? 7 : (reqdata.bt==IP16x8||reqdata.bt==IP8x8||reqdata.bt==IP4x8 ? 3 : 1)) + (yfracc==0 ? 0 : 1);
290 
      if(loadHorNum < loadHorNumMax)
291 
         loadHorNum <= loadHorNum+1;
292 
      else
293 
         begin
294 
            loadHorNum <= 0;
295 
            if(loadVerNum < loadVerNumMax)
296 
               loadVerNum <= loadVerNum+1;
297 
            else
298 
               begin
299 
                  loadVerNum <= 0;
300 
                  reqfifoLoad.deq();
301 
               end
302 
         end
303 91 jamey.hick 
      $display( "PARDEB Trace interpolator(%0d): loadChroma %h %h %h %h %h %h %h", total_cycles, xfracc, yfracc, loadHorNum, loadVerNum, reqdata.refIdx, horAddr, verAddr);
304 90 jamey.hick 
   endrule
305 
 
306 
 
307 
   rule work1Luma ( reqfifoWork1.first() matches tagged IPWLuma .reqdata &&& !work1Done );
308 
      let xfracl = reqdata.xFracL;
309 
      let yfracl = reqdata.yFracL;
310 
      let offset = reqdata.offset;
311 
      let blockT = reqdata.bt;
312 
      Bool twoStage = (xfracl==1||xfracl==3) && (yfracl==1||yfracl==3);
313 
      Vector#(20,Bit#(8)) work1Vector8Next = work1Vector8;
314 
      if(memRespQ.first() matches tagged IPLoadResp .tempreaddata)
315 
         begin
316 
            memRespQ.deq();
317 
            Vector#(4,Bit#(8)) readdata = replicate(0);
318 
            readdata[0] = tempreaddata[7:0];
319 
            readdata[1] = tempreaddata[15:8];
320 
            readdata[2] = tempreaddata[23:16];
321 
            readdata[3] = tempreaddata[31:24];
322 
            //$display( "Trace interpolator: workLuma stage 0 readdata %h %h %h %h %h %h", workHorNum, workVerNum, readdata[3], readdata[2], readdata[1], readdata[0] );
323 
            Vector#(4,Bit#(8)) tempResult8 = replicate(0);
324 
            Vector#(4,Bit#(15)) tempResult15 = replicate(0);
325 
            if(xfracl==0 || yfracl==0 || xfracl==2)
326 
               begin
327 
                  if(xfracl==0)//reorder
328 
                     begin
329 
                        for(Integer ii=0; ii<4; ii=ii+1)
330 
                           begin
331 
                              Bit#(2) offsetplusii = offset+fromInteger(ii);
332 
                              if(offset <= 3-fromInteger(ii) && offset!=0)
333 
                                 tempResult8[ii] = work1Vector8[offsetplusii];
334 
                              else
335 
                                 tempResult8[ii] = readdata[offsetplusii];
336 
                              work1Vector8Next[ii] = readdata[ii];
337 
                           end
338 
                        for(Integer ii=0; ii<4; ii=ii+1)
339 
                           tempResult15[ii] = zeroExtend({tempResult8[ii],5'b00000});
340 
                     end
341 
                  else//horizontal interpolation
342 
                     begin
343 
                        offset = offset-2;
344 
                        for(Integer ii=0; ii<8; ii=ii+1)
345 
                           work1Vector8Next[ii] = work1Vector8[ii+4];
346 
                        for(Integer ii=0; ii<4; ii=ii+1)
347 
                           begin
348 
                              Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
349 
                              work1Vector8Next[tempIndex] = readdata[ii];
350 
                           end
351 
                        for(Integer ii=0; ii<4; ii=ii+1)
352 
                           begin
353 
                              tempResult15[ii] = interpolate8to15(work1Vector8Next[ii],work1Vector8Next[ii+1],work1Vector8Next[ii+2],work1Vector8Next[ii+3],work1Vector8Next[ii+4],work1Vector8Next[ii+5]);
354 
                              tempResult8[ii] = clip1y10to8(truncate((tempResult15[ii]+16)>>5));
355 
                              if(xfracl == 1)
356 
                                 tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,work1Vector8Next[ii+2]} + 1) >> 1);
357 
                              else if(xfracl == 3)
358 
                                 tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,work1Vector8Next[ii+3]} + 1) >> 1);
359 
                           end
360 
                     end
361 
                  Bit#(2) workHorNumOffset = (xfracl!=0 ? 2 : (reqdata.offset==0 ? 0 : 1));
362 
                  if(work1HorNum >= workHorNumOffset)
363 
                     begin
364 
                        Bit#(1) horAddr = truncate(work1HorNum-workHorNumOffset);
365 
                        if(yfracl == 0)
366 
                           begin
367 
                              for(Integer ii=0; ii<4; ii=ii+1)
368 
                                 tempResult15[ii] = zeroExtend({tempResult8[ii],5'b00000});
369 
                           end
370 
                        workFile.upd({workFileFlag,work1VerNum,horAddr},tempResult15);
371 
                     end
372 
                  Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + workHorNumOffset;
373 
                  Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + (yfracl!=0 ? 5 : 0);
374 
                  if(work1HorNum < workHorNumMax)
375 
                     work1HorNum <= work1HorNum+1;
376 
                  else
377 
                     begin
378 
                        work1HorNum <= 0;
379 
                        if(work1VerNum < workVerNumMax)
380 
                           work1VerNum <= work1VerNum+1;
381 
                        else
382 
                           begin
383 
                              work1VerNum <= 0;
384 
                              work1Done <= True;
385 
                           end
386 
                     end
387 
               end
388 
            else if(work1Stage == 0)//vertical interpolation
389 
               begin
390 
                  offset = offset + (xfracl==3&&(yfracl==1||yfracl==3) ? 1 : 0);
391 
                  for(Integer ii=0; ii<4; ii=ii+1)
392 
                     tempResult15[ii] = interpolate8to15(work1Vector8[ii],work1Vector8[ii+4],work1Vector8[ii+8],work1Vector8[ii+12],work1Vector8[ii+16],readdata[ii]);
393 
                  for(Integer ii=0; ii<16; ii=ii+1)
394 
                     work1Vector8Next[ii] = work1Vector8[ii+4];
395 
                  for(Integer ii=0; ii<4; ii=ii+1)
396 
                     work1Vector8Next[ii+16] = readdata[ii];
397 
                  Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + (yfracl==2 ? 2 : (offset==0 ? 0 : 1));
398 
                  Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + 5;
399 
                  Bit#(2) horAddr = work1HorNum;
400 
                  Bit#(3) verAddr = truncate(work1VerNum-5);
401 
                  if(work1VerNum > 4)
402 
                     begin
403 
                        workFile.upd({workFileFlag,verAddr,horAddr},tempResult15);
404 
                        //$display( "Trace interpolator: workLuma stage 0 result %h %h %h %h %h %h %h", workHorNum, workVerNum, {verAddr,horAddr}, tempResult15[3], tempResult15[2], tempResult15[1], tempResult15[0]);
405 
                     end
406 
                  if(twoStage)
407 
                     begin
408 
                        Bit#(2) storeHorAddr = work1HorNum;
409 
                        Bit#(4) storeVerAddr = work1VerNum;
410 
                        if((xfracl==3 ? offset<3 : offset<2))
411 
                           storeHorAddr = storeHorAddr+1;
412 
                        if(yfracl==3)
413 
                           storeVerAddr = storeVerAddr-3;
414 
                        else
415 
                           storeVerAddr = storeVerAddr-2;
416 
                        if(storeVerAddr < 8)
417 
                           storeFile.upd({workFileFlag,storeVerAddr[2:0],storeHorAddr},readdata);
418 
                     end
419 
                  if(work1VerNum < workVerNumMax)
420 
                     work1VerNum <= work1VerNum+1;
421 
                  else
422 
                     begin
423 
                        work1VerNum <= 0;
424 
                        if(work1HorNum < workHorNumMax)
425 
                           work1HorNum <= work1HorNum+1;
426 
                        else
427 
                           begin
428 
                              if(twoStage)
429 
                                 begin
430 
                                    work1Stage <= 1;
431 
                                    if((xfracl==3 ? offset<3 : offset<2))
432 
                                       work1HorNum <= 0;
433 
                                    else
434 
                                       work1HorNum <= workHorNumMax+1;
435 
                                 end
436 
                              else
437 
                                 begin
438 
                                    work1HorNum <= 0;
439 
                                    work1Done <= True;
440 
                                 end
441 
                           end
442 
                     end
443 
               end
444 
            else//second stage of twoStage
445 
               begin
446 
                  storeFile.upd({workFileFlag,work1VerNum[2:0],work1HorNum},readdata);
447 
                  Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + 2;
448 
                  Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3);
449 
                  if(work1VerNum < workVerNumMax)
450 
                     work1VerNum <= work1VerNum+1;
451 
                  else
452 
                     begin
453 
                        work1VerNum <= 0;
454 
                        offset = offset + (xfracl==3 ? 1 : 0);
455 
                        if(work1HorNum
456 
                           work1HorNum <= workHorNumMax;
457 
                        else
458 
                           begin
459 
                              work1HorNum <= 0;
460 
                              work1Stage <= 0;
461 
                              work1Done <= True;
462 
                           end
463 
                     end
464 
               end
465 
         end
466 
      work1Vector8 <= work1Vector8Next;
467 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator(%0d): work1Luma %h %h %h %h %h %h", total_cycles, xfracl, yfracl, work1HorNum, work1VerNum, offset, work1Stage);
468 90 jamey.hick 
   endrule
469 
 
470 
 
471 
   rule work2Luma ( reqregWork2 matches tagged Valid .vdata &&& vdata matches tagged IPWLuma .reqdata &&& !work2Done &&& !work8x8Done );
472 
      let xfracl = reqdata.xFracL;
473 
      let yfracl = reqdata.yFracL;
474 
      let offset = reqdata.offset;
475 
      let blockT = reqdata.bt;
476 
      Vector#(20,Bit#(8)) work2Vector8Next = work2Vector8;
477 
      Vector#(20,Bit#(15)) work2Vector15Next = work2Vector15;
478 
      Vector#(16,Bit#(1)) resultReadyNext = resultReady;
479 
      Vector#(4,Bit#(8)) tempResult8 = replicate(0);
480 
      Vector#(4,Bit#(15)) readdata = replicate(0);
481 
      if(yfracl==0)
482 
         begin
483 
            readdata = workFile.sub({(1-workFileFlag),1'b0,work2VerNum[1],work2HorNum,work2VerNum[0]});
484 
            for(Integer ii=0; ii<4; ii=ii+1)
485 
               tempResult8[ii] = (readdata[ii])[12:5];
486 
            resultFile.upd({work2VerNum[1],work2HorNum,work2VerNum[0]},tempResult8);
487 
            resultReadyNext[{work2VerNum[1],work2HorNum,work2VerNum[0]}] = 1;
488 
            work2HorNum <= work2HorNum+1;
489 
            if(work2HorNum == 3)
490 
               begin
491 
                  if(work2VerNum == 3)
492 
                     begin
493 
                        work2VerNum <= 0;
494 
                        work2Done <= True;
495 
                        if(((blockT==IP4x8 || blockT==IP8x4) && work2SubMbPart==0) || (blockT==IP4x4 && work2SubMbPart<3))
496 
                           work2SubMbPart <= work2SubMbPart+1;
497 
                        else
498 
                           begin
499 
                              work2SubMbPart <= 0;
500 
                              work8x8Done <= True;
501 
                           end
502 
                     end
503 
                  else
504 
                     work2VerNum <= work2VerNum+1;
505 
               end
506 
         end
507 
      else if(xfracl==0 || xfracl==2)//vertical interpolation
508 
         begin
509 
            readdata = workFile.sub({(1-workFileFlag),work2VerNum,work2HorNum[0]});
510 
            for(Integer ii=0; ii<4; ii=ii+1)
511 
               begin
512 
                  tempResult8[ii] = interpolate15to8(work2Vector15[ii],work2Vector15[ii+4],work2Vector15[ii+8],work2Vector15[ii+12],work2Vector15[ii+16],readdata[ii]);
513 
                  if(yfracl == 1)
514 
                     tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((work2Vector15[ii+8]+16)>>5))} + 1) >> 1);
515 
                  else if(yfracl == 3)
516 
                     tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((work2Vector15[ii+12]+16)>>5))} + 1) >> 1);
517 
               end
518 
            for(Integer ii=0; ii<16; ii=ii+1)
519 
               work2Vector15Next[ii] = work2Vector15[ii+4];
520 
            for(Integer ii=0; ii<4; ii=ii+1)
521 
               work2Vector15Next[ii+16] = readdata[ii];
522 
            Bit#(2) workHorNumMax = 1;
523 
            Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + 5;
524 
            if(work2VerNum > 4)
525 
               begin
526 
                  Bit#(1) horAddr = truncate(work2HorNum);
527 
                  Bit#(3) verAddr = truncate(work2VerNum-5);
528 
                  horAddr = horAddr + ((blockT==IP4x8&&work2SubMbPart==1)||(blockT==IP4x4&&work2SubMbPart[0]==1) ? 1 : 0);
529 
                  verAddr = verAddr + ((blockT==IP8x4&&work2SubMbPart==1)||(blockT==IP4x4&&work2SubMbPart[1]==1) ? 4 : 0);
530 
                  resultFile.upd({verAddr,horAddr},tempResult8);
531 
                  resultReadyNext[{verAddr,horAddr}] = 1;
532 
               end
533 
            if(work2VerNum < workVerNumMax)
534 
               work2VerNum <= work2VerNum+1;
535 
            else
536 
               begin
537 
                  work2VerNum <= 0;
538 
                  if(work2HorNum < workHorNumMax)
539 
                     work2HorNum <= work2HorNum+1;
540 
                  else
541 
                     begin
542 
                        work2HorNum <= 0;
543 
                        work2Done <= True;
544 
                        if(((blockT==IP4x8 || blockT==IP8x4) && work2SubMbPart==0) || (blockT==IP4x4 && work2SubMbPart<3))
545 
                           work2SubMbPart <= work2SubMbPart+1;
546 
                        else
547 
                           begin
548 
                              work2SubMbPart <= 0;
549 
                              work8x8Done <= True;
550 
                           end
551 
                     end
552 
               end
553 
         end
554 
      else//horizontal interpolation
555 
         begin
556 
            offset = offset-2;
557 
            if(yfracl == 2)
558 
               begin
559 
                  readdata = workFile.sub({(1-workFileFlag),work2VerNum[2:0],work2HorNum});
560 
                  for(Integer ii=0; ii<8; ii=ii+1)
561 
                     work2Vector15Next[ii] = work2Vector15[ii+4];
562 
                  for(Integer ii=0; ii<4; ii=ii+1)
563 
                     begin
564 
                        Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
565 
                        work2Vector15Next[tempIndex] = readdata[ii];
566 
                     end
567 
                  for(Integer ii=0; ii<4; ii=ii+1)
568 
                     begin
569 
                        tempResult8[ii] = interpolate15to8(work2Vector15Next[ii],work2Vector15Next[ii+1],work2Vector15Next[ii+2],work2Vector15Next[ii+3],work2Vector15Next[ii+4],work2Vector15Next[ii+5]);
570 
                        if(xfracl == 1)
571 
                           tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((work2Vector15Next[ii+2]+16)>>5))} + 1) >> 1);
572 
                        else if(xfracl == 3)
573 
                           tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((work2Vector15Next[ii+3]+16)>>5))} + 1) >> 1);
574 
                     end
575 
               end
576 
            else
577 
               begin
578 
                  Vector#(4,Bit#(8)) readdata8 = storeFile.sub({(1-workFileFlag),work2VerNum[2:0],work2HorNum});
579 
                  for(Integer ii=0; ii<8; ii=ii+1)
580 
                     work2Vector8Next[ii] = work2Vector8[ii+4];
581 
                  for(Integer ii=0; ii<4; ii=ii+1)
582 
                     begin
583 
                        Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
584 
                        work2Vector8Next[tempIndex] = readdata8[ii];
585 
                     end
586 
                  Vector#(4,Bit#(15)) tempResult15 = replicate(0);
587 
                  for(Integer ii=0; ii<4; ii=ii+1)
588 
                     begin
589 
                        tempResult15[ii] = interpolate8to15(work2Vector8Next[ii],work2Vector8Next[ii+1],work2Vector8Next[ii+2],work2Vector8Next[ii+3],work2Vector8Next[ii+4],work2Vector8Next[ii+5]);
590 
                        tempResult8[ii] = clip1y10to8(truncate((tempResult15[ii]+16)>>5));
591 
                     end
592 
                  Bit#(2) verOffset;
593 
                  Vector#(4,Bit#(15)) verResult15 = replicate(0);
594 
                  if(xfracl == 1)
595 
                     verOffset = reqdata.offset;
596 
                  else
597 
                     verOffset = reqdata.offset+1;
598 
                  readdata = workFile.sub({(1-workFileFlag),work2VerNum[2:0],(work2HorNum-2+(verOffset==0?0:1))});
599 
                  for(Integer ii=0; ii<4; ii=ii+1)
600 
                     begin
601 
                        Bit#(2) offsetplusii = verOffset+fromInteger(ii);
602 
                        if(verOffset <= 3-fromInteger(ii) && verOffset!=0)
603 
                           verResult15[ii] = work2Vector15[offsetplusii];
604 
                        else
605 
                           verResult15[ii] = readdata[offsetplusii];
606 
                        work2Vector15Next[ii] = readdata[ii];
607 
                     end
608 
                  for(Integer ii=0; ii<4; ii=ii+1)
609 
                     begin
610 
                        Bit#(9) tempVal = zeroExtend(clip1y10to8(truncate((verResult15[ii]+16)>>5)));
611 
                        tempResult8[ii] = truncate((tempVal+zeroExtend(tempResult8[ii])+1)>>1);
612 
                     end
613 
               end
614 
            if(work2HorNum >= 2)
615 
               begin
616 
                  Bit#(1) horAddr = truncate(work2HorNum-2);
617 
                  Bit#(3) verAddr = truncate(work2VerNum);
618 
                  horAddr = horAddr + ((blockT==IP4x8&&work2SubMbPart==1)||(blockT==IP4x4&&work2SubMbPart[0]==1) ? 1 : 0);
619 
                  verAddr = verAddr + ((blockT==IP8x4&&work2SubMbPart==1)||(blockT==IP4x4&&work2SubMbPart[1]==1) ? 4 : 0);
620 
                  resultFile.upd({verAddr,horAddr},tempResult8);
621 
                  resultReadyNext[{verAddr,horAddr}] = 1;
622 
                  //$display( "Trace interpolator: workLuma stage 1 result %h %h %h %h %h %h %h %h", workHorNum, workVerNum, {verAddr,horAddr}, tempResult8[3], tempResult8[2], tempResult8[1], tempResult8[0], pack(resultReadyNext));
623 
               end
624 
            Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + 2;
625 
            Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3);
626 
            if(work2HorNum < workHorNumMax)
627 
               work2HorNum <= work2HorNum+1;
628 
            else
629 
               begin
630 
                  work2HorNum <= 0;
631 
                  if(work2VerNum < workVerNumMax)
632 
                     work2VerNum <= work2VerNum+1;
633 
                  else
634 
                     begin
635 
                        work2VerNum <= 0;
636 
                        work2Done <= True;
637 
                        if(((blockT==IP4x8 || blockT==IP8x4) && work2SubMbPart==0) || (blockT==IP4x4 && work2SubMbPart<3))
638 
                           work2SubMbPart <= work2SubMbPart+1;
639 
                        else
640 
                           begin
641 
                              work2SubMbPart <= 0;
642 
                              work8x8Done <= True;
643 
                           end
644 
                     end
645 
               end
646 
         end
647 
      work2Vector8 <= work2Vector8Next;
648 
      work2Vector15 <= work2Vector15Next;
649 
      resultReady <= resultReadyNext;
650 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator%(0d): work2Luma %h %h %h %h %h",total_cycles, xfracl, yfracl, work2HorNum, work2VerNum, offset);
651 90 jamey.hick 
   endrule
652 
 
653 
 
654 
   rule work1Chroma ( reqfifoWork1.first() matches tagged IPWChroma .reqdata &&& !work1Done );
655 
      Bit#(4) xfracc = zeroExtend(reqdata.xFracC);
656 
      Bit#(4) yfracc = zeroExtend(reqdata.yFracC);
657 
      let offset = reqdata.offset;
658 
      let blockT = reqdata.bt;
659 
      Vector#(20,Bit#(8)) work1Vector8Next = work1Vector8;
660 
      if(memRespQ.first() matches tagged IPLoadResp .tempreaddata)
661 
         begin
662 
            memRespQ.deq();
663 
            Vector#(4,Bit#(8)) readdata = replicate(0);
664 
            readdata[0] = tempreaddata[7:0];
665 
            readdata[1] = tempreaddata[15:8];
666 
            readdata[2] = tempreaddata[23:16];
667 
            readdata[3] = tempreaddata[31:24];
668 
            Vector#(5,Bit#(8)) tempWork8 = replicate(0);
669 
            Vector#(5,Bit#(8)) tempPrev8 = replicate(0);
670 
            Vector#(4,Bit#(8)) tempResult8 = replicate(0);
671 
            Bool resultReadyFlag = False;
672 
            for(Integer ii=0; ii<4; ii=ii+1)
673 
               begin
674 
                  Bit#(2) offsetplusii = offset+fromInteger(ii);
675 
                  if(offset <= 3-fromInteger(ii) && !((blockT==IP4x8||blockT==IP4x4)&&(offset[1]==0||(xfracc==0&&offset!=3))) && !(xfracc==0&&offset==0))
676 
                     tempWork8[ii] = work1Vector8[offsetplusii];
677 
                  else
678 
                     tempWork8[ii] = readdata[offsetplusii];
679 
                  work1Vector8Next[ii] = readdata[ii];
680 
               end
681 
            tempWork8[4] = readdata[offset];
682 
            if((blockT==IP16x8 || blockT==IP16x16) && work1HorNum==(xfracc==0&&offset==0 ? 1 : 2))
683 
               begin
684 
                  for(Integer ii=0; ii<5; ii=ii+1)
685 
                     begin
686 
                        tempPrev8[ii] = work1Vector8[ii+9];
687 
                        work1Vector8Next[ii+9] = tempWork8[ii];
688 
                     end
689 
               end
690 
            else
691 
               begin
692 
                  for(Integer ii=0; ii<5; ii=ii+1)
693 
                     tempPrev8[ii] = work1Vector8[ii+4];
694 
                  if(work1HorNum==(xfracc==0&&offset==0 ? 0 : 1) || ((blockT==IP4x8||blockT==IP4x4)&&(offset[1]==0||(xfracc==0&&offset!=3))))
695 
                     begin
696 
                        for(Integer ii=0; ii<5; ii=ii+1)
697 
                           work1Vector8Next[ii+4] = tempWork8[ii];
698 
                     end
699 
               end
700 
            if(yfracc==0)
701 
               begin
702 
                  for(Integer ii=0; ii<5; ii=ii+1)
703 
                     tempPrev8[ii] = tempWork8[ii];
704 
               end
705 
            for(Integer ii=0; ii<4; ii=ii+1)
706 
               begin
707 
                  Bit#(14) tempVal = zeroExtend((8-xfracc))*zeroExtend((8-yfracc))*zeroExtend(tempPrev8[ii]);
708 
                  tempVal = tempVal + zeroExtend(xfracc)*zeroExtend((8-yfracc))*zeroExtend(tempPrev8[ii+1]);
709 
                  tempVal = tempVal + zeroExtend((8-xfracc))*zeroExtend(yfracc)*zeroExtend(tempWork8[ii]);
710 
                  tempVal = tempVal + zeroExtend(xfracc)*zeroExtend(yfracc)*zeroExtend(tempWork8[ii+1]);
711 
                  tempResult8[ii] = truncate((tempVal+32)>>6);
712 
               end
713 
            if(work1VerNum > 0 || yfracc==0)
714 
               begin
715 
                  if(blockT==IP4x8 || blockT==IP4x4)
716 
                     begin
717 
                        Bit#(5) tempIndex = 10 + zeroExtend(work1VerNum<<1);
718 
                        work1Vector8Next[tempIndex] = tempResult8[0];
719 
                        work1Vector8Next[tempIndex+1] = tempResult8[1];
720 
                        tempResult8[2] = tempResult8[0];
721 
                        tempResult8[3] = tempResult8[1];
722 
                        tempResult8[0] = work1Vector8[tempIndex];
723 
                        tempResult8[1] = work1Vector8[tempIndex+1];
724 
                        if((work1HorNum>0 || offset[1]==0) && work1SubMbPart[0]==1)
725 
                           resultReadyFlag = True;
726 
                     end
727 
                  else
728 
                     begin
729 
                        if(work1HorNum>0 || (xfracc==0 && offset==0))
730 
                           resultReadyFlag = True;
731 
                     end
732 
               end
733 
            if(resultReadyFlag)
734 
               begin
735 
                  Bit#(1) horAddr = ((blockT==IP4x8 || blockT==IP4x4) ? 0 : truncate(((xfracc==0 && offset==0) ? work1HorNum : work1HorNum-1)));
736 
                  Bit#(3) verAddr = truncate((yfracc==0 ? work1VerNum : work1VerNum-1));
737 
                  horAddr = horAddr + ((blockT==IP16x8||blockT==IP16x16) ? 0 : work1MbPart[0]);
738 
                  verAddr = verAddr + ((blockT==IP8x16||blockT==IP16x16) ? 0 : ((blockT==IP16x8) ? {work1MbPart[0],2'b00} : {work1MbPart[1],2'b00}));
739 
                  verAddr = verAddr + ((blockT==IP8x4&&work1SubMbPart==1)||(blockT==IP4x4&&work1SubMbPart[1]==1) ? 2 : 0);
740 
                  storeFile.upd({workFileFlag,1'b0,verAddr,horAddr},tempResult8);
741 
               end
742 
            Bit#(2) workHorNumMax = (blockT==IP4x8||blockT==IP4x4 ? (offset[1]==0||(xfracc==0&&offset!=3) ? 0 : 1) : ((blockT==IP16x16||blockT==IP16x8 ? 1 : 0) + (xfracc==0&&offset==0 ? 0 : 1)));
743 
            Bit#(4) workVerNumMax = (blockT==IP16x16||blockT==IP8x16 ? 7 : (blockT==IP16x8||blockT==IP8x8||blockT==IP4x8 ? 3 : 1)) + (yfracc==0 ? 0 : 1);
744 
            if(work1HorNum < workHorNumMax)
745 
               work1HorNum <= work1HorNum+1;
746 
            else
747 
               begin
748 
                  work1HorNum <= 0;
749 
                  if(work1VerNum < workVerNumMax)
750 
                     work1VerNum <= work1VerNum+1;
751 
                  else
752 
                     begin
753 
                        Bool allDone = False;
754 
                        work1VerNum <= 0;
755 
                        if(((blockT==IP4x8 || blockT==IP8x4) && work1SubMbPart==0) || (blockT==IP4x4 && work1SubMbPart<3))
756 
                           work1SubMbPart <= work1SubMbPart+1;
757 
                        else
758 
                           begin
759 
                              work1SubMbPart <= 0;
760 
                              if(((blockT==IP16x8 || blockT==IP8x16) && work1MbPart==0) || (!(blockT==IP16x8 || blockT==IP8x16 || blockT==IP16x16) && work1MbPart<3))
761 
                                 work1MbPart <= work1MbPart+1;
762 
                              else
763 
                                 begin
764 
                                    work1MbPart <= 0;
765 
                                    work1Done <= True;
766 
                                    allDone = True;
767 
                                 end
768 
                           end
769 
                        if(!allDone)
770 
                           reqfifoWork1.deq();
771 
                     end
772 
               end
773 
         end
774 
      work1Vector8 <= work1Vector8Next;
775 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator(%0d): work1Chroma %h %h %h %h %h", total_cycles, xfracc, yfracc, work1HorNum, work1VerNum, offset);
776 90 jamey.hick 
   endrule
777 
 
778 
 
779 
   rule work2Chroma ( reqregWork2 matches tagged Valid .vdata &&& vdata matches tagged IPWChroma .reqdata &&& !work2Done &&& !work8x8Done );
780 
      Vector#(16,Bit#(1)) resultReadyNext = resultReady;
781 
      resultFile.upd({work2VerNum[1],work2HorNum,work2VerNum[0]},storeFile.sub({(1-workFileFlag),1'b0,work2VerNum[1],work2HorNum,work2VerNum[0]}));
782 
      resultReadyNext[{work2VerNum[1],work2HorNum,work2VerNum[0]}] = 1;
783 
      work2HorNum <= work2HorNum+1;
784 
      if(work2HorNum == 3)
785 
         begin
786 
            if(work2VerNum == 3)
787 
               begin
788 
                  work2VerNum <= 0;
789 
                  work2Done <= True;
790 
                  work8x8Done <= True;
791 
               end
792 
            else
793 
               work2VerNum <= work2VerNum+1;
794 
         end
795 
      resultReady <= resultReadyNext;
796 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator(%0d): work2Chroma %h %h", total_cycles, work2HorNum, work2VerNum);
797 90 jamey.hick 
   endrule
798 
 
799 
 
800 
  rule outputing( !outDone && resultReady[{outBlockNum[1],outPixelNum,outBlockNum[0]}]==1 );
801 
      outfifo.enq(resultFile.sub({outBlockNum[1],outPixelNum,outBlockNum[0]}));
802 
      outPixelNum <= outPixelNum+1;
803 
      if(outPixelNum == 3)
804 
         begin
805 
            outBlockNum <= outBlockNum+1;
806 
            if(outBlockNum == 3)
807 
               outDone <= True;
808 
         end
809 91 jamey.hick 
      $display( "PARDEBLOCK Trace interpolator(%0d): outputing %h %h", total_cycles, outBlockNum, outPixelNum);
810 90 jamey.hick 
   endrule
811 
 
812 
 
813 
   rule switching( work1Done && (work2Done || reqregWork2==Invalid) && !work8x8Done);
814 
      work1Done <= False;
815 
      work2Done <= False;
816 
      reqregWork2 <= (tagged Valid reqfifoWork1.first());
817 
      workFileFlag <= 1-workFileFlag;
818 
      reqfifoWork1.deq();
819 
      $display( "Trace interpolator: switching %h %h", outBlockNum, outPixelNum);
820 
   endrule
821 
 
822 
 
823 
   rule switching8x8( work1Done && (work2Done || reqregWork2==Invalid) && work8x8Done && outDone);
824 
      outDone <= False;
825 
      work8x8Done <= False;
826 
      resultReady <= replicate(0);
827 
      work1Done <= False;
828 
      work2Done <= False;
829 
      reqregWork2 <= (tagged Valid reqfifoWork1.first());
830 
      workFileFlag <= 1-workFileFlag;
831 
      reqfifoWork1.deq();
832 
      $display( "Trace interpolator: switching8x8 %h %h", outBlockNum, outPixelNum);
833 
   endrule
834 
 
835 
 
836 
 
837 
   method Action   setPicWidth( Bit#(PicWidthSz) newPicWidth );
838 
      picWidth <= newPicWidth;
839 
   endmethod
840 
 
841 
   method Action   setPicHeight( Bit#(PicHeightSz) newPicHeight );
842 
      picHeight <= newPicHeight;
843 
   endmethod
844 
 
845 
   method Action request( InterpolatorIT inputdata );
846 
      reqfifoLoad.enq(inputdata);
847 
      if(inputdata matches tagged IPLuma .indata)
848 
         reqfifoWork1.enq(IPWLuma {xFracL:indata.mvhor[1:0],yFracL:indata.mvver[1:0],offset:indata.mvhor[3:2],bt:indata.bt});
849 
      else if(inputdata matches tagged IPChroma .indata)
850 
         reqfifoWork1.enq(IPWChroma {xFracC:indata.mvhor[2:0],yFracC:indata.mvver[2:0],offset:indata.mvhor[4:3]+{indata.hor[0],1'b0},bt:indata.bt});
851 
   endmethod
852 
 
853 
   method Vector#(4,Bit#(8)) first();
854 
      return outfifo.first();
855 
   endmethod
856 
 
857 
   method Action deq();
858 
      outfifo.deq();
859 
   endmethod
860 91 jamey.hick 
 
861 90 jamey.hick 
 
862 
   method Action endOfFrame();
863 
      endOfFrameFlag <= True;
864 
   endmethod
865 
 
866 91 jamey.hick 
   method InterpolatorLoadReq mem_request_first();
867 
     return memReqQ.first;
868 
   endmethod
869 90 jamey.hick 
 
870 91 jamey.hick 
   method Action mem_request_deq();
871 
     memReqQ.deq;
872 
   endmethod
873 
 
874 
   interface mem_client_resp = fifoToPut(memRespQ);
875 
 
876 90 jamey.hick 
endmodule
877 
 
878 
 
879 
endpackage

powered by: WebSVN 2.1.0

© copyright 1999-2024 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.