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

Subversion Repositories bluespec-h264

[/] [bluespec-h264/] [trunk/] [release/] [mkInterpolator_3stage.bsv] - Blame information for rev 84

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

Line No. Rev Author Line
1 84 jamey.hick 
//**********************************************************************
2 
// interpolator implementation
3 
//----------------------------------------------------------------------
4 
//
5 
//
6 
 
7 
package mkInterpolator;
8 
 
9 
import H264Types::*;
10 
import IInterpolator::*;
11 
import FIFO::*;
12 
import Vector::*;
13 
 
14 
import Connectable::*;
15 
import GetPut::*;
16 
import ClientServer::*;
17 
 
18 
 
19 
//-----------------------------------------------------------
20 
// Local Datatypes
21 
//-----------------------------------------------------------
22 
 
23 
typedef union tagged
24 
{
25 
 struct { Bit#(2) xFracL; Bit#(2) yFracL; Bit#(2) offset; IPBlockType bt; } IPWLuma;
26 
 struct { Bit#(3) xFracC; Bit#(3) yFracC; Bit#(2) offset; IPBlockType bt; } IPWChroma;
27 
}
28 
InterpolatorWT deriving(Eq,Bits);
29 
 
30 
 
31 
//-----------------------------------------------------------
32 
// Helper functions
33 
 
34 
function Bit#(8) clip1y10to8( Bit#(10) innum );
35 
   if(innum[9] == 1)
36 
      return 0;
37 
   else if(innum[8] == 1)
38 
      return 255;
39 
   else
40 
      return truncate(innum);
41 
endfunction
42 
 
43 
function Bit#(15) interpolate8to15( Bit#(8) in0, Bit#(8) in1, Bit#(8) in2, Bit#(8) in3, Bit#(8) in4, Bit#(8) in5 );
44 
   return zeroExtend(in0) - 5*zeroExtend(in1) + 20*zeroExtend(in2) + 20*zeroExtend(in3) - 5*zeroExtend(in4) + zeroExtend(in5);
45 
endfunction
46 
 
47 
function Bit#(8) interpolate15to8( Bit#(15) in0, Bit#(15) in1, Bit#(15) in2, Bit#(15) in3, Bit#(15) in4, Bit#(15) in5 );
48 
   Bit#(20) temp = signExtend(in0) - 5*signExtend(in1) + 20*signExtend(in2) + 20*signExtend(in3) - 5*signExtend(in4) + signExtend(in5) + 512;
49 
   return clip1y10to8(truncate(temp>>10));
50 
endfunction
51 
 
52 
 
53 
 
54 
//-----------------------------------------------------------
55 
// Interpolation Module
56 
//-----------------------------------------------------------
57 
 
58 
 
59 
(* synthesize *)
60 
module mkInterpolator( Interpolator );
61 
 
62 
   FIFO#(InterpolatorIT) reqfifoLoad <- mkSizedFIFO(interpolator_reqfifoLoad_size);
63 
   FIFO#(InterpolatorWT) reqfifoWork <- mkSizedFIFO(interpolator_reqfifoWork_size);
64 
   FIFO#(Vector#(4,Bit#(8))) outfifo <- mkFIFO;
65 
   Reg#(Bool) endOfFrameFlag <- mkReg(False);
66 
   FIFO#(InterpolatorLoadReq)  memReqQ  <- mkFIFO;
67 
   FIFO#(InterpolatorLoadResp) memRespQ <- mkSizedFIFO(interpolator_memRespQ_size);
68 
 
69 
   Reg#(Bit#(PicWidthSz))  picWidth  <- mkReg(maxPicWidthInMB);
70 
   Reg#(Bit#(PicHeightSz)) picHeight <- mkReg(0);
71 
 
72 
   RFile1#(Bit#(5),Vector#(4,Bit#(15))) workFile  <- mkRFile1Full();
73 
   RFile1#(Bit#(4),Vector#(4,Bit#(8))) resultFile <- mkRFile1Full();
74 
 
75 
   Reg#(Bit#(1)) loadStage  <- mkReg(0);
76 
   Reg#(Bit#(2)) loadHorNum <- mkReg(0);
77 
   Reg#(Bit#(4)) loadVerNum <- mkReg(0);
78 
 
79 
   Reg#(Bit#(1)) workStage     <- mkReg(0);
80 
   Reg#(Bit#(2)) workMbPart    <- mkReg(0);//only for Chroma
81 
   Reg#(Bit#(2)) workSubMbPart <- mkReg(0);
82 
   Reg#(Bit#(2)) workHorNum    <- mkReg(0);
83 
   Reg#(Bit#(4)) workVerNum    <- mkReg(0);
84 
   Reg#(Vector#(20,Bit#(8))) workVector8 <- mkRegU;
85 
   Reg#(Vector#(20,Bit#(15))) workVector15 <- mkRegU;
86 
   Reg#(Vector#(4,Bit#(1))) resultReady <- mkRegU;
87 
   Reg#(Bool) workDone <- mkReg(False);
88 
 
89 
   Reg#(Bit#(2)) outBlockNum <- mkReg(0);
90 
   Reg#(Bit#(2)) outPixelNum <- mkReg(0);
91 
   Reg#(Bool) outDone <- mkReg(False);
92 
 
93 
 
94 
   rule sendEndOfFrameReq( endOfFrameFlag );
95 
      endOfFrameFlag <= False;
96 
      memReqQ.enq(IPLoadEndFrame);
97 
   endrule
98 
 
99 
 
100 
   rule loadLuma( reqfifoLoad.first() matches tagged IPLuma .reqdata &&& !endOfFrameFlag );
101 
      Bit#(2) xfracl = reqdata.mvhor[1:0];
102 
      Bit#(2) yfracl = reqdata.mvver[1:0];
103 
      Bool twoStage = (xfracl==1||xfracl==3) && (yfracl==1||yfracl==3);
104 
      Bool horInter = (twoStage ? loadStage==1 : xfracl!=0);
105 
      Bool verInter = (twoStage ? loadStage==0 : yfracl!=0);
106 
      Bit#(2) offset = reqdata.mvhor[3:2] + ((twoStage&&verInter&&xfracl==3) ? 1 : 0);
107 
      Bit#(1) horOut = 0;
108 
      Bit#(TAdd#(PicWidthSz,2)) horAddr;
109 
      Bit#(TAdd#(PicHeightSz,4)) verAddr;
110 
      Bit#(TAdd#(PicWidthSz,12)) horTemp = zeroExtend({reqdata.hor,2'b00}) + zeroExtend({loadHorNum,2'b00}) + (xfracl==3&&(yfracl==1||yfracl==3)&&loadStage==0 ? 1 : 0);
111 
      Bit#(TAdd#(PicHeightSz,10)) verTemp = zeroExtend(reqdata.ver) + zeroExtend(loadVerNum) + (yfracl==3&&(xfracl==1||xfracl==3)&&loadStage==1 ? 1 : 0);
112 
      Bit#(13) mvhortemp = signExtend(reqdata.mvhor[13:2])-(horInter?2:0);
113 
      Bit#(11) mvvertemp = signExtend(reqdata.mvver[11:2])-(verInter?2:0);
114 
      if(mvhortemp[12]==1 && zeroExtend(0-mvhortemp)>horTemp)
115 
         begin
116 
            horAddr = 0;
117 
            horOut = 1;
118 
         end
119 
      else
120 
         begin
121 
            horTemp = horTemp + signExtend(mvhortemp);
122 
            if(horTemp>=zeroExtend({picWidth,4'b0000}))
123 
               begin
124 
                  horAddr = {picWidth-1,2'b11};
125 
                  horOut = 1;
126 
               end
127 
            else
128 
               horAddr = truncate(horTemp>>2);
129 
         end
130 
      if(mvvertemp[10]==1 && zeroExtend(0-mvvertemp)>verTemp)
131 
         verAddr = 0;
132 
      else
133 
         begin
134 
            verTemp = verTemp + signExtend(mvvertemp);
135 
            if(verTemp>=zeroExtend({picHeight,4'b0000}))
136 
               verAddr = {picHeight-1,4'b1111};
137 
            else
138 
               verAddr = truncate(verTemp);
139 
         end
140 
      memReqQ.enq(IPLoadLuma {refIdx:reqdata.refIdx,horOutOfBounds:horOut,hor:horAddr,ver:verAddr});
141 
      Bool verFirst = (twoStage&&loadStage==0) || (yfracl==2&&(xfracl==1||xfracl==3));
142 
      Bit#(2) loadHorNumMax = (reqdata.bt==IP8x8||reqdata.bt==IP8x4 ? 1 : 0) + (horInter ? 2 : (offset==0 ? 0 : 1));
143 
      Bit#(4) loadVerNumMax = (reqdata.bt==IP8x8||reqdata.bt==IP4x8 ? 7 : 3) + (verInter ? 5 : 0);
144 
      if(verFirst)
145 
         begin
146 
            if(loadVerNum < loadVerNumMax)
147 
               loadVerNum <= loadVerNum+1;
148 
            else
149 
               begin
150 
                  loadVerNum <= 0;
151 
                  if(loadHorNum < loadHorNumMax)
152 
                     loadHorNum <= loadHorNum+1;
153 
                  else
154 
                     begin
155 
                        loadHorNum <= 0;
156 
                        if(twoStage)
157 
                           loadStage <= 1;
158 
                        else
159 
                           reqfifoLoad.deq();
160 
                     end
161 
               end
162 
         end
163 
      else
164 
         begin
165 
            if(loadHorNum < loadHorNumMax)
166 
               loadHorNum <= loadHorNum+1;
167 
            else
168 
               begin
169 
                  loadHorNum <= 0;
170 
                  if(loadVerNum < loadVerNumMax)
171 
                     loadVerNum <= loadVerNum+1;
172 
                  else
173 
                     begin
174 
                        loadVerNum <= 0;
175 
                        loadStage <= 0;
176 
                        reqfifoLoad.deq();
177 
                     end
178 
               end
179 
         end
180 
      if(reqdata.bt==IP16x16 || reqdata.bt==IP16x8 || reqdata.bt==IP8x16)
181 
         $display( "ERROR Interpolation: loadLuma block sizes > 8x8 not supported");
182 
      //$display( "Trace interpolator: loadLuma %h %h %h %h %h %h %h", xfracl, yfracl, loadHorNum, loadVerNum, reqdata.refIdx, horAddr, verAddr);
183 
   endrule
184 
 
185 
 
186 
   rule loadChroma( reqfifoLoad.first() matches tagged IPChroma .reqdata &&& !endOfFrameFlag );
187 
      Bit#(3) xfracc = reqdata.mvhor[2:0];
188 
      Bit#(3) yfracc = reqdata.mvver[2:0];
189 
      Bit#(2) offset = reqdata.mvhor[4:3]+{reqdata.hor[0],1'b0};
190 
      Bit#(1) horOut = 0;
191 
      Bit#(TAdd#(PicWidthSz,1)) horAddr;
192 
      Bit#(TAdd#(PicHeightSz,3)) verAddr;
193 
      Bit#(TAdd#(PicWidthSz,11)) horTemp = zeroExtend({reqdata.hor,1'b0}) + zeroExtend({loadHorNum,2'b00});
194 
      Bit#(TAdd#(PicHeightSz,9)) verTemp = zeroExtend(reqdata.ver) + zeroExtend(loadVerNum);
195 
      if(reqdata.mvhor[13]==1 && zeroExtend(0-reqdata.mvhor[13:3])>horTemp)
196 
         begin
197 
            horAddr = 0;
198 
            horOut = 1;
199 
         end
200 
      else
201 
         begin
202 
            horTemp = horTemp + signExtend(reqdata.mvhor[13:3]);
203 
            if(horTemp>=zeroExtend({picWidth,3'b000}))
204 
               begin
205 
                  horAddr = {picWidth-1,1'b1};
206 
                  horOut = 1;
207 
               end
208 
            else
209 
               horAddr = truncate(horTemp>>2);
210 
         end
211 
      if(reqdata.mvver[11]==1 && zeroExtend(0-reqdata.mvver[11:3])>verTemp)
212 
         verAddr = 0;
213 
      else
214 
         begin
215 
            verTemp = verTemp + signExtend(reqdata.mvver[11:3]);
216 
            if(verTemp>=zeroExtend({picHeight,3'b000}))
217 
               verAddr = {picHeight-1,3'b111};
218 
            else
219 
               verAddr = truncate(verTemp);
220 
         end
221 
      memReqQ.enq(IPLoadChroma {refIdx:reqdata.refIdx,uv:reqdata.uv,horOutOfBounds:horOut,hor:horAddr,ver:verAddr});
222 
      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)));
223 
      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);
224 
      if(loadHorNum < loadHorNumMax)
225 
         loadHorNum <= loadHorNum+1;
226 
      else
227 
         begin
228 
            loadHorNum <= 0;
229 
            if(loadVerNum < loadVerNumMax)
230 
               loadVerNum <= loadVerNum+1;
231 
            else
232 
               begin
233 
                  loadVerNum <= 0;
234 
                  reqfifoLoad.deq();
235 
               end
236 
         end
237 
      //$display( "Trace interpolator: loadChroma %h %h %h %h %h %h %h", xfracc, yfracc, loadHorNum, loadVerNum, reqdata.refIdx, horAddr, verAddr);
238 
   endrule
239 
 
240 
 
241 
   rule workLuma ( reqfifoWork.first() matches tagged IPWLuma .reqdata &&& !workDone );
242 
      let xfracl = reqdata.xFracL;
243 
      let yfracl = reqdata.yFracL;
244 
      let offset = reqdata.offset;
245 
      let blockT = reqdata.bt;
246 
      Vector#(20,Bit#(8)) workVector8Next = workVector8;
247 
      Vector#(20,Bit#(15)) workVector15Next = workVector15;
248 
      Vector#(4,Bit#(1)) resultReadyNext = resultReady;
249 
      if(workStage == 0)
250 
         begin
251 
            if(memRespQ.first() matches tagged IPLoadResp .tempreaddata)
252 
               begin
253 
                  memRespQ.deq();
254 
                  Vector#(4,Bit#(8)) readdata = replicate(0);
255 
                  readdata[0] = tempreaddata[7:0];
256 
                  readdata[1] = tempreaddata[15:8];
257 
                  readdata[2] = tempreaddata[23:16];
258 
                  readdata[3] = tempreaddata[31:24];
259 
                  //$display( "Trace interpolator: workLuma stage 0 readdata %h %h %h %h %h %h", workHorNum, workVerNum, readdata[3], readdata[2], readdata[1], readdata[0] );
260 
                  Vector#(4,Bit#(8)) tempResult8 = replicate(0);
261 
                  Vector#(4,Bit#(15)) tempResult15 = replicate(0);
262 
                  if(xfracl==0 || yfracl==0 || xfracl==2)
263 
                     begin
264 
                        if(xfracl==0)//reorder
265 
                           begin
266 
                              for(Integer ii=0; ii<4; ii=ii+1)
267 
                                 begin
268 
                                    Bit#(2) offsetplusii = offset+fromInteger(ii);
269 
                                    if(offset <= 3-fromInteger(ii) && offset!=0)
270 
                                       tempResult8[ii] = workVector8[offsetplusii];
271 
                                    else
272 
                                       tempResult8[ii] = readdata[offsetplusii];
273 
                                    workVector8Next[ii] = readdata[ii];
274 
                                 end
275 
                              for(Integer ii=0; ii<4; ii=ii+1)
276 
                                 tempResult15[ii] = zeroExtend({tempResult8[ii],5'b00000});
277 
                           end
278 
                        else//horizontal interpolation
279 
                           begin
280 
                              offset = offset-2;
281 
                              for(Integer ii=0; ii<8; ii=ii+1)
282 
                                 workVector8Next[ii] = workVector8[ii+4];
283 
                              for(Integer ii=0; ii<4; ii=ii+1)
284 
                                 begin
285 
                                    Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
286 
                                    workVector8Next[tempIndex] = readdata[ii];
287 
                                 end
288 
                              for(Integer ii=0; ii<4; ii=ii+1)
289 
                                 begin
290 
                                    tempResult15[ii] = interpolate8to15(workVector8Next[ii],workVector8Next[ii+1],workVector8Next[ii+2],workVector8Next[ii+3],workVector8Next[ii+4],workVector8Next[ii+5]);
291 
                                    tempResult8[ii] = clip1y10to8(truncate((tempResult15[ii]+16)>>5));
292 
                                    if(xfracl == 1)
293 
                                       tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,workVector8Next[ii+2]} + 1) >> 1);
294 
                                    else if(xfracl == 3)
295 
                                       tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,workVector8Next[ii+3]} + 1) >> 1);
296 
                                 end
297 
                           end
298 
                        Bit#(2) workHorNumOffset = (xfracl!=0 ? 2 : (reqdata.offset==0 ? 0 : 1));
299 
                        if(workHorNum >= workHorNumOffset)
300 
                           begin
301 
                              Bit#(1) horAddr = truncate(workHorNum-workHorNumOffset);
302 
                              if(yfracl == 0)//write to resultFile
303 
                                 begin
304 
                                    Bit#(3) verAddr = truncate(workVerNum);
305 
                                    horAddr = horAddr + ((blockT==IP4x8&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[0]==1) ? 1 : 0);
306 
                                    verAddr = verAddr + ((blockT==IP8x4&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[1]==1) ? 4 : 0);
307 
                                    resultFile.upd({verAddr,horAddr},tempResult8);
308 
                                    if(verAddr[1:0] == 3)
309 
                                       resultReadyNext[{verAddr[2],horAddr}] = 1;
310 
                                 end
311 
                              else//write to workFile
312 
                                 workFile.upd({workVerNum,horAddr},tempResult15);
313 
                           end
314 
                        Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + workHorNumOffset;
315 
                        Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + (yfracl!=0 ? 5 : 0);
316 
                        if(workHorNum < workHorNumMax)
317 
                           workHorNum <= workHorNum+1;
318 
                        else
319 
                           begin
320 
                              workHorNum <= 0;
321 
                              if(workVerNum < workVerNumMax)
322 
                                 workVerNum <= workVerNum+1;
323 
                              else
324 
                                 begin
325 
                                    workVerNum <= 0;
326 
                                    if(yfracl!=0)
327 
                                       workStage <= 1;
328 
                                    else
329 
                                       begin
330 
                                          if(((blockT==IP4x8 || blockT==IP8x4) && workSubMbPart==0) || (blockT==IP4x4 && workSubMbPart<3))
331 
                                             workSubMbPart <= workSubMbPart+1;
332 
                                          else
333 
                                             begin
334 
                                                workSubMbPart <= 0;
335 
                                                workDone <= True;
336 
                                             end
337 
                                          reqfifoWork.deq();
338 
                                       end
339 
                                 end
340 
                           end
341 
                     end
342 
                  else//vertical interpolation
343 
                     begin
344 
                        offset = offset + (xfracl==3&&(yfracl==1||yfracl==3) ? 1 : 0);
345 
                        for(Integer ii=0; ii<4; ii=ii+1)
346 
                           tempResult15[ii] = interpolate8to15(workVector8[ii],workVector8[ii+4],workVector8[ii+8],workVector8[ii+12],workVector8[ii+16],readdata[ii]);
347 
                        for(Integer ii=0; ii<16; ii=ii+1)
348 
                           workVector8Next[ii] = workVector8[ii+4];
349 
                        for(Integer ii=0; ii<4; ii=ii+1)
350 
                           workVector8Next[ii+16] = readdata[ii];
351 
                        Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + (yfracl==2 ? 2 : (offset==0 ? 0 : 1));
352 
                        Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + 5;
353 
                        Bit#(2) horAddr = workHorNum;
354 
                        Bit#(3) verAddr = truncate(workVerNum-5);
355 
                        if(workVerNum > 4)
356 
                           begin
357 
                              workFile.upd({verAddr,horAddr},tempResult15);
358 
                              //$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]);
359 
                           end
360 
                        if(workVerNum < workVerNumMax)
361 
                           workVerNum <= workVerNum+1;
362 
                        else
363 
                           begin
364 
                              workVerNum <= 0;
365 
                              if(workHorNum < workHorNumMax)
366 
                                 workHorNum <= workHorNum+1;
367 
                              else
368 
                                 begin
369 
                                    workHorNum <= 0;
370 
                                    workStage <= 1;
371 
                                 end
372 
                           end
373 
                     end
374 
               end
375 
         end
376 
      else
377 
         begin
378 
            Vector#(4,Bit#(8)) tempResult8 = replicate(0);
379 
            Vector#(4,Bit#(15)) readdata = replicate(0);
380 
            if(yfracl==0)
381 
               $display( "ERROR Interpolation: workLuma loadStage==1 and yfracl==0");
382 
            if(xfracl==0 || xfracl==2)//vertical interpolation
383 
               begin
384 
                  readdata = workFile.sub({workVerNum,workHorNum[0]});
385 
                  for(Integer ii=0; ii<4; ii=ii+1)
386 
                     begin
387 
                        tempResult8[ii] = interpolate15to8(workVector15[ii],workVector15[ii+4],workVector15[ii+8],workVector15[ii+12],workVector15[ii+16],readdata[ii]);
388 
                        if(yfracl == 1)
389 
                           tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((workVector15[ii+8]+16)>>5))} + 1) >> 1);
390 
                        else if(yfracl == 3)
391 
                           tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((workVector15[ii+12]+16)>>5))} + 1) >> 1);
392 
                     end
393 
                  for(Integer ii=0; ii<16; ii=ii+1)
394 
                     workVector15Next[ii] = workVector15[ii+4];
395 
                  for(Integer ii=0; ii<4; ii=ii+1)
396 
                     workVector15Next[ii+16] = readdata[ii];
397 
                  Bit#(2) workHorNumMax = 1;
398 
                  Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3) + 5;
399 
                  if(workVerNum > 4)
400 
                     begin
401 
                        Bit#(1) horAddr = truncate(workHorNum);
402 
                        Bit#(3) verAddr = truncate(workVerNum-5);
403 
                        horAddr = horAddr + ((blockT==IP4x8&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[0]==1) ? 1 : 0);
404 
                        verAddr = verAddr + ((blockT==IP8x4&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[1]==1) ? 4 : 0);
405 
                        resultFile.upd({verAddr,horAddr},tempResult8);
406 
                        if(verAddr[1:0] == 3)
407 
                           resultReadyNext[{verAddr[2],horAddr}] = 1;
408 
                     end
409 
                  if(workVerNum < workVerNumMax)
410 
                     workVerNum <= workVerNum+1;
411 
                  else
412 
                     begin
413 
                        workVerNum <= 0;
414 
                        if(workHorNum < workHorNumMax)
415 
                           workHorNum <= workHorNum+1;
416 
                        else
417 
                           begin
418 
                              workHorNum <= 0;
419 
                              workStage <= 0;
420 
                              if(((blockT==IP4x8 || blockT==IP8x4) && workSubMbPart==0) || (blockT==IP4x4 && workSubMbPart<3))
421 
                                 workSubMbPart <= workSubMbPart+1;
422 
                              else
423 
                                 begin
424 
                                    workSubMbPart <= 0;
425 
                                    workDone <= True;
426 
                                 end
427 
                              reqfifoWork.deq();
428 
                           end
429 
                     end
430 
               end
431 
            else//horizontal interpolation
432 
               begin
433 
                  offset = offset-2;
434 
                  if(yfracl == 2)
435 
                     begin
436 
                        readdata = workFile.sub({workVerNum[2:0],workHorNum});
437 
                        for(Integer ii=0; ii<8; ii=ii+1)
438 
                           workVector15Next[ii] = workVector15[ii+4];
439 
                        for(Integer ii=0; ii<4; ii=ii+1)
440 
                           begin
441 
                              Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
442 
                              workVector15Next[tempIndex] = readdata[ii];
443 
                           end
444 
                        for(Integer ii=0; ii<4; ii=ii+1)
445 
                           begin
446 
                              tempResult8[ii] = interpolate15to8(workVector15Next[ii],workVector15Next[ii+1],workVector15Next[ii+2],workVector15Next[ii+3],workVector15Next[ii+4],workVector15Next[ii+5]);
447 
                              if(xfracl == 1)
448 
                                 tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((workVector15Next[ii+2]+16)>>5))} + 1) >> 1);
449 
                              else if(xfracl == 3)
450 
                                 tempResult8[ii] = truncate(({1'b0,tempResult8[ii]} + {1'b0,clip1y10to8(truncate((workVector15Next[ii+3]+16)>>5))} + 1) >> 1);
451 
                           end
452 
                     end
453 
                  else
454 
                     begin
455 
                        if(memRespQ.first() matches tagged IPLoadResp .tempreaddata8)
456 
                           begin
457 
                              memRespQ.deq();
458 
                              Vector#(4,Bit#(8)) readdata8 = replicate(0);
459 
                              readdata8[0] = tempreaddata8[7:0];
460 
                              readdata8[1] = tempreaddata8[15:8];
461 
                              readdata8[2] = tempreaddata8[23:16];
462 
                              readdata8[3] = tempreaddata8[31:24];
463 
                              for(Integer ii=0; ii<8; ii=ii+1)
464 
                                 workVector8Next[ii] = workVector8[ii+4];
465 
                              for(Integer ii=0; ii<4; ii=ii+1)
466 
                                 begin
467 
                                    Bit#(4) tempIndex = fromInteger(ii) + 8 - zeroExtend(offset);
468 
                                    workVector8Next[tempIndex] = readdata8[ii];
469 
                                 end
470 
                              Vector#(4,Bit#(15)) tempResult15 = replicate(0);
471 
                              for(Integer ii=0; ii<4; ii=ii+1)
472 
                                 begin
473 
                                    tempResult15[ii] = interpolate8to15(workVector8Next[ii],workVector8Next[ii+1],workVector8Next[ii+2],workVector8Next[ii+3],workVector8Next[ii+4],workVector8Next[ii+5]);
474 
                                    tempResult8[ii] = clip1y10to8(truncate((tempResult15[ii]+16)>>5));
475 
                                 end
476 
                              Bit#(2) verOffset;
477 
                              Vector#(4,Bit#(15)) verResult15 = replicate(0);
478 
                              if(xfracl == 1)
479 
                                 verOffset = reqdata.offset;
480 
                              else
481 
                                 verOffset = reqdata.offset+1;
482 
                              readdata = workFile.sub({workVerNum[2:0],(workHorNum-2+(verOffset==0?0:1))});
483 
                              for(Integer ii=0; ii<4; ii=ii+1)
484 
                                 begin
485 
                                    Bit#(2) offsetplusii = verOffset+fromInteger(ii);
486 
                                    if(verOffset <= 3-fromInteger(ii) && verOffset!=0)
487 
                                       verResult15[ii] = workVector15[offsetplusii];
488 
                                    else
489 
                                       verResult15[ii] = readdata[offsetplusii];
490 
                                    workVector15Next[ii] = readdata[ii];
491 
                                 end
492 
                              for(Integer ii=0; ii<4; ii=ii+1)
493 
                                 begin
494 
                                    Bit#(9) tempVal = zeroExtend(clip1y10to8(truncate((verResult15[ii]+16)>>5)));
495 
                                    tempResult8[ii] = truncate((tempVal+zeroExtend(tempResult8[ii])+1)>>1);
496 
                                 end
497 
                           end
498 
                     end
499 
                  if(workHorNum >= 2)
500 
                     begin
501 
                        Bit#(1) horAddr = truncate(workHorNum-2);
502 
                        Bit#(3) verAddr = truncate(workVerNum);
503 
                        horAddr = horAddr + ((blockT==IP4x8&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[0]==1) ? 1 : 0);
504 
                        verAddr = verAddr + ((blockT==IP8x4&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[1]==1) ? 4 : 0);
505 
                        resultFile.upd({verAddr,horAddr},tempResult8);
506 
                        if(verAddr[1:0] == 3)
507 
                           resultReadyNext[{verAddr[2],horAddr}] = 1;
508 
                        //$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));
509 
                     end
510 
                  Bit#(2) workHorNumMax = (blockT==IP8x8||blockT==IP8x4 ? 1 : 0) + 2;
511 
                  Bit#(4) workVerNumMax = (blockT==IP8x8||blockT==IP4x8 ? 7 : 3);
512 
                  if(workHorNum < workHorNumMax)
513 
                     workHorNum <= workHorNum+1;
514 
                  else
515 
                     begin
516 
                        workHorNum <= 0;
517 
                        if(workVerNum < workVerNumMax)
518 
                           workVerNum <= workVerNum+1;
519 
                        else
520 
                           begin
521 
                              workVerNum <= 0;
522 
                              workStage <= 0;
523 
                              if(((blockT==IP4x8 || blockT==IP8x4) && workSubMbPart==0) || (blockT==IP4x4 && workSubMbPart<3))
524 
                                 workSubMbPart <= workSubMbPart+1;
525 
                              else
526 
                                 begin
527 
                                    workSubMbPart <= 0;
528 
                                    workDone <= True;
529 
                                 end
530 
                              reqfifoWork.deq();
531 
                           end
532 
                     end
533 
               end
534 
         end
535 
      workVector8 <= workVector8Next;
536 
      workVector15 <= workVector15Next;
537 
      resultReady <= resultReadyNext;
538 
      //$display( "Trace interpolator: workLuma %h %h %h %h %h %h", xfracl, yfracl, workHorNum, workVerNum, offset, workStage);
539 
   endrule
540 
 
541 
 
542 
   rule workChroma ( reqfifoWork.first() matches tagged IPWChroma .reqdata &&& !workDone );
543 
      Bit#(4) xfracc = zeroExtend(reqdata.xFracC);
544 
      Bit#(4) yfracc = zeroExtend(reqdata.yFracC);
545 
      let offset = reqdata.offset;
546 
      let blockT = reqdata.bt;
547 
      Vector#(20,Bit#(8)) workVector8Next = workVector8;
548 
      Vector#(4,Bit#(1)) resultReadyNext = resultReady;
549 
      if(memRespQ.first() matches tagged IPLoadResp .tempreaddata)
550 
         begin
551 
            memRespQ.deq();
552 
            Vector#(4,Bit#(8)) readdata = replicate(0);
553 
            readdata[0] = tempreaddata[7:0];
554 
            readdata[1] = tempreaddata[15:8];
555 
            readdata[2] = tempreaddata[23:16];
556 
            readdata[3] = tempreaddata[31:24];
557 
            Vector#(5,Bit#(8)) tempWork8 = replicate(0);
558 
            Vector#(5,Bit#(8)) tempPrev8 = replicate(0);
559 
            Vector#(4,Bit#(8)) tempResult8 = replicate(0);
560 
            Bool resultReadyFlag = False;
561 
            for(Integer ii=0; ii<4; ii=ii+1)
562 
               begin
563 
                  Bit#(2) offsetplusii = offset+fromInteger(ii);
564 
                  if(offset <= 3-fromInteger(ii) && !((blockT==IP4x8||blockT==IP4x4)&&(offset[1]==0||(xfracc==0&&offset!=3))) && !(xfracc==0&&offset==0))
565 
                     tempWork8[ii] = workVector8[offsetplusii];
566 
                  else
567 
                     tempWork8[ii] = readdata[offsetplusii];
568 
                  workVector8Next[ii] = readdata[ii];
569 
               end
570 
            tempWork8[4] = readdata[offset];
571 
            if((blockT==IP16x8 || blockT==IP16x16) && workHorNum==(xfracc==0&&offset==0 ? 1 : 2))
572 
               begin
573 
                  for(Integer ii=0; ii<5; ii=ii+1)
574 
                     begin
575 
                        tempPrev8[ii] = workVector8[ii+9];
576 
                        workVector8Next[ii+9] = tempWork8[ii];
577 
                     end
578 
               end
579 
            else
580 
               begin
581 
                  for(Integer ii=0; ii<5; ii=ii+1)
582 
                     tempPrev8[ii] = workVector8[ii+4];
583 
                  if(workHorNum==(xfracc==0&&offset==0 ? 0 : 1) || ((blockT==IP4x8||blockT==IP4x4)&&(offset[1]==0||(xfracc==0&&offset!=3))))
584 
                     begin
585 
                        for(Integer ii=0; ii<5; ii=ii+1)
586 
                           workVector8Next[ii+4] = tempWork8[ii];
587 
                     end
588 
               end
589 
            if(yfracc==0)
590 
               begin
591 
                  for(Integer ii=0; ii<5; ii=ii+1)
592 
                     tempPrev8[ii] = tempWork8[ii];
593 
               end
594 
            for(Integer ii=0; ii<4; ii=ii+1)
595 
               begin
596 
                  Bit#(14) tempVal = zeroExtend((8-xfracc))*zeroExtend((8-yfracc))*zeroExtend(tempPrev8[ii]);
597 
                  tempVal = tempVal + zeroExtend(xfracc)*zeroExtend((8-yfracc))*zeroExtend(tempPrev8[ii+1]);
598 
                  tempVal = tempVal + zeroExtend((8-xfracc))*zeroExtend(yfracc)*zeroExtend(tempWork8[ii]);
599 
                  tempVal = tempVal + zeroExtend(xfracc)*zeroExtend(yfracc)*zeroExtend(tempWork8[ii+1]);
600 
                  tempResult8[ii] = truncate((tempVal+32)>>6);
601 
               end
602 
            if(workVerNum > 0 || yfracc==0)
603 
               begin
604 
                  if(blockT==IP4x8 || blockT==IP4x4)
605 
                     begin
606 
                        Bit#(5) tempIndex = 10 + zeroExtend(workVerNum<<1);
607 
                        workVector8Next[tempIndex] = tempResult8[0];
608 
                        workVector8Next[tempIndex+1] = tempResult8[1];
609 
                        tempResult8[2] = tempResult8[0];
610 
                        tempResult8[3] = tempResult8[1];
611 
                        tempResult8[0] = workVector8[tempIndex];
612 
                        tempResult8[1] = workVector8[tempIndex+1];
613 
                        if((workHorNum>0 || offset[1]==0) && workSubMbPart[0]==1)
614 
                           resultReadyFlag = True;
615 
                     end
616 
                  else
617 
                     begin
618 
                        if(workHorNum>0 || (xfracc==0 && offset==0))
619 
                           resultReadyFlag = True;
620 
                     end
621 
               end
622 
            if(resultReadyFlag)
623 
               begin
624 
                  Bit#(1) horAddr = ((blockT==IP4x8 || blockT==IP4x4) ? 0 : truncate(((xfracc==0 && offset==0) ? workHorNum : workHorNum-1)));
625 
                  Bit#(3) verAddr = truncate((yfracc==0 ? workVerNum : workVerNum-1));
626 
                  horAddr = horAddr + ((blockT==IP16x8||blockT==IP16x16) ? 0 : workMbPart[0]);
627 
                  verAddr = verAddr + ((blockT==IP8x16||blockT==IP16x16) ? 0 : ((blockT==IP16x8) ? {workMbPart[0],2'b00} : {workMbPart[1],2'b00}));
628 
                  verAddr = verAddr + ((blockT==IP8x4&&workSubMbPart==1)||(blockT==IP4x4&&workSubMbPart[1]==1) ? 2 : 0);
629 
                  resultFile.upd({verAddr,horAddr},tempResult8);
630 
                  if(verAddr[1:0] == 3)
631 
                     resultReadyNext[{verAddr[2],horAddr}] = 1;
632 
               end
633 
            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)));
634 
            Bit#(4) workVerNumMax = (blockT==IP16x16||blockT==IP8x16 ? 7 : (blockT==IP16x8||blockT==IP8x8||blockT==IP4x8 ? 3 : 1)) + (yfracc==0 ? 0 : 1);
635 
            if(workHorNum < workHorNumMax)
636 
               workHorNum <= workHorNum+1;
637 
            else
638 
               begin
639 
                  workHorNum <= 0;
640 
                  if(workVerNum < workVerNumMax)
641 
                     workVerNum <= workVerNum+1;
642 
                  else
643 
                     begin
644 
                        workVerNum <= 0;
645 
                        if(((blockT==IP4x8 || blockT==IP8x4) && workSubMbPart==0) || (blockT==IP4x4 && workSubMbPart<3))
646 
                           workSubMbPart <= workSubMbPart+1;
647 
                        else
648 
                           begin
649 
                              workSubMbPart <= 0;
650 
                              if(((blockT==IP16x8 || blockT==IP8x16) && workMbPart==0) || (!(blockT==IP16x8 || blockT==IP8x16 || blockT==IP16x16) && workMbPart<3))
651 
                                 workMbPart <= workMbPart+1;
652 
                              else
653 
                                 begin
654 
                                    workMbPart <= 0;
655 
                                    workDone <= True;
656 
                                 end
657 
                           end
658 
                        reqfifoWork.deq();
659 
                     end
660 
               end
661 
         end
662 
      workVector8 <= workVector8Next;
663 
      resultReady <= resultReadyNext;
664 
      //$display( "Trace interpolator: workChroma %h %h %h %h %h", xfracc, yfracc, workHorNum, workVerNum, offset);
665 
   endrule
666 
 
667 
 
668 
   rule outputing( !outDone && resultReady[outBlockNum]==1 );
669 
      outfifo.enq(resultFile.sub({outBlockNum[1],outPixelNum,outBlockNum[0]}));
670 
      outPixelNum <= outPixelNum+1;
671 
      if(outPixelNum == 3)
672 
         begin
673 
            outBlockNum <= outBlockNum+1;
674 
            if(outBlockNum == 3)
675 
               outDone <= True;
676 
         end
677 
      //$display( "Trace interpolator: outputing %h %h %h %h %h %h", outBlockNum, outPixelNum, tempVector[3], tempVector[2], tempVector[1], tempVector[0]);
678 
   endrule
679 
 
680 
 
681 
   rule switching( outDone && workDone );
682 
      outDone <= False;
683 
      workDone <= False;
684 
      resultReady <= replicate(0);
685 
      //$display( "Trace interpolator: switching %h %h", outBlockNum, outPixelNum);
686 
   endrule
687 
 
688 
 
689 
   method Action   setPicWidth( Bit#(PicWidthSz) newPicWidth );
690 
      picWidth <= newPicWidth;
691 
   endmethod
692 
 
693 
   method Action   setPicHeight( Bit#(PicHeightSz) newPicHeight );
694 
      picHeight <= newPicHeight;
695 
   endmethod
696 
 
697 
   method Action request( InterpolatorIT inputdata );
698 
      reqfifoLoad.enq(inputdata);
699 
      if(inputdata matches tagged IPLuma .indata)
700 
         reqfifoWork.enq(IPWLuma {xFracL:indata.mvhor[1:0],yFracL:indata.mvver[1:0],offset:indata.mvhor[3:2],bt:indata.bt});
701 
      else if(inputdata matches tagged IPChroma .indata)
702 
         reqfifoWork.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});
703 
   endmethod
704 
 
705 
   method Vector#(4,Bit#(8)) first();
706 
      return outfifo.first();
707 
   endmethod
708 
 
709 
   method Action deq();
710 
      outfifo.deq();
711 
   endmethod
712 
 
713 
   method Action endOfFrame();
714 
      endOfFrameFlag <= True;
715 
   endmethod
716 
 
717 
   interface Client mem_client;
718 
      interface Get request  = fifoToGet(memReqQ);
719 
      interface Put response = fifoToPut(memRespQ);
720 
   endinterface
721 
 
722 
 
723 
endmodule
724 
 
725 
 
726 
endpackage

powered by: WebSVN 2.1.0

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