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