| 1 |
83 |
jamey.hick |
|
| 2 |
|
|
// The MIT License
|
| 3 |
|
|
|
| 4 |
|
|
// Copyright (c) 2006-2007 Massachusetts Institute of Technology
|
| 5 |
|
|
|
| 6 |
|
|
// Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 7 |
|
|
// of this software and associated documentation files (the "Software"), to deal
|
| 8 |
|
|
// in the Software without restriction, including without limitation the rights
|
| 9 |
|
|
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 10 |
|
|
// copies of the Software, and to permit persons to whom the Software is
|
| 11 |
|
|
// furnished to do so, subject to the following conditions:
|
| 12 |
|
|
|
| 13 |
|
|
// The above copyright notice and this permission notice shall be included in
|
| 14 |
|
|
// all copies or substantial portions of the Software.
|
| 15 |
|
|
|
| 16 |
|
|
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 17 |
|
|
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 18 |
|
|
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 19 |
|
|
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 20 |
|
|
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 21 |
|
|
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
| 22 |
|
|
// THE SOFTWARE.
|
| 23 |
|
|
|
| 24 |
2 |
jamey.hick |
//**********************************************************************
|
| 25 |
|
|
// Prediction
|
| 26 |
|
|
//----------------------------------------------------------------------
|
| 27 |
|
|
//
|
| 28 |
|
|
//
|
| 29 |
|
|
|
| 30 |
|
|
package mkPrediction;
|
| 31 |
|
|
|
| 32 |
|
|
import H264Types::*;
|
| 33 |
|
|
|
| 34 |
|
|
import IPrediction::*;
|
| 35 |
|
|
import IInterpolator::*;
|
| 36 |
|
|
import mkInterpolator::*;
|
| 37 |
|
|
import FIFO::*;
|
| 38 |
|
|
import FIFOF::*;
|
| 39 |
|
|
import Vector::*;
|
| 40 |
19 |
jamey.hick |
|
| 41 |
2 |
jamey.hick |
import Connectable::*;
|
| 42 |
|
|
import GetPut::*;
|
| 43 |
|
|
import ClientServer::*;
|
| 44 |
|
|
|
| 45 |
|
|
|
| 46 |
|
|
//-----------------------------------------------------------
|
| 47 |
|
|
// Local Datatypes
|
| 48 |
|
|
//-----------------------------------------------------------
|
| 49 |
|
|
|
| 50 |
|
|
typedef union tagged
|
| 51 |
|
|
{
|
| 52 |
|
|
void Intra; //Intra non-4x4
|
| 53 |
|
|
void Intra4x4;
|
| 54 |
|
|
void Inter;
|
| 55 |
|
|
}
|
| 56 |
|
|
OutState deriving(Eq,Bits);
|
| 57 |
|
|
|
| 58 |
|
|
typedef union tagged
|
| 59 |
|
|
{
|
| 60 |
|
|
void Start; //not working on anything in particular
|
| 61 |
|
|
void Intra16x16;
|
| 62 |
|
|
void Intra4x4;
|
| 63 |
|
|
void IntraPCM;
|
| 64 |
|
|
}
|
| 65 |
|
|
IntraState deriving(Eq,Bits);
|
| 66 |
|
|
|
| 67 |
|
|
typedef union tagged
|
| 68 |
|
|
{
|
| 69 |
|
|
void Start; //not working on anything in particular
|
| 70 |
|
|
void InterP16x16;
|
| 71 |
|
|
void InterP16x8;
|
| 72 |
|
|
void InterP8x16;
|
| 73 |
|
|
void InterP8x8;
|
| 74 |
|
|
void InterP8x8ref0;
|
| 75 |
|
|
void InterPskip;
|
| 76 |
|
|
}
|
| 77 |
|
|
InterState deriving(Eq,Bits);
|
| 78 |
|
|
|
| 79 |
|
|
typedef union tagged
|
| 80 |
|
|
{
|
| 81 |
|
|
Bit#(1) NotInter;//0 for not available, 1 for intra-coded
|
| 82 |
|
|
struct {Bit#(4) refIdx; Bit#(14) mvhor; Bit#(12) mvver; Bit#(1) nonZeroTransCoeff;} BlockMv;
|
| 83 |
|
|
}
|
| 84 |
|
|
InterBlockMv deriving(Eq,Bits);
|
| 85 |
|
|
|
| 86 |
|
|
typedef union tagged
|
| 87 |
|
|
{
|
| 88 |
|
|
void SkipMB;
|
| 89 |
|
|
void NonSkipMB;
|
| 90 |
|
|
void Intra4x4;
|
| 91 |
|
|
void Intra4x4PlusChroma;
|
| 92 |
|
|
}
|
| 93 |
|
|
NextOutput deriving(Eq,Bits);
|
| 94 |
|
|
|
| 95 |
|
|
|
| 96 |
|
|
|
| 97 |
|
|
//-----------------------------------------------------------
|
| 98 |
|
|
// Helper functions
|
| 99 |
|
|
|
| 100 |
|
|
function Bit#(8) intra4x4SelectTop( Bit#(72) valVector, Bit#(4) idx );
|
| 101 |
|
|
case(idx)
|
| 102 |
|
|
0: return valVector[15:8];
|
| 103 |
|
|
1: return valVector[23:16];
|
| 104 |
|
|
2: return valVector[31:24];
|
| 105 |
|
|
3: return valVector[39:32];
|
| 106 |
|
|
4: return valVector[47:40];
|
| 107 |
|
|
5: return valVector[55:48];
|
| 108 |
|
|
6: return valVector[63:56];
|
| 109 |
|
|
7: return valVector[71:64];
|
| 110 |
|
|
default: return valVector[7:0];
|
| 111 |
|
|
endcase
|
| 112 |
|
|
endfunction
|
| 113 |
|
|
|
| 114 |
|
|
function Bit#(8) intra4x4SelectLeft( Bit#(40) valVector, Bit#(3) idx );
|
| 115 |
|
|
case(idx)
|
| 116 |
|
|
0: return valVector[15:8];
|
| 117 |
|
|
1: return valVector[23:16];
|
| 118 |
|
|
2: return valVector[31:24];
|
| 119 |
|
|
3: return valVector[39:32];
|
| 120 |
|
|
default: return valVector[7:0];
|
| 121 |
|
|
endcase
|
| 122 |
|
|
endfunction
|
| 123 |
|
|
|
| 124 |
|
|
function Bit#(8) select32to8( Bit#(32) valVector, Bit#(2) idx );
|
| 125 |
|
|
case(idx)
|
| 126 |
|
|
0: return valVector[7:0];
|
| 127 |
|
|
1: return valVector[15:8];
|
| 128 |
|
|
2: return valVector[23:16];
|
| 129 |
|
|
3: return valVector[31:24];
|
| 130 |
|
|
endcase
|
| 131 |
|
|
endfunction
|
| 132 |
|
|
|
| 133 |
|
|
function Bit#(8) select16to8( Bit#(16) valVector, Bit#(1) idx );
|
| 134 |
|
|
case(idx)
|
| 135 |
|
|
0: return valVector[7:0];
|
| 136 |
|
|
1: return valVector[15:8];
|
| 137 |
|
|
endcase
|
| 138 |
|
|
endfunction
|
| 139 |
|
|
|
| 140 |
|
|
function Bool absDiffGEFour14( Bit#(14) val1, Bit#(14) val2 );
|
| 141 |
|
|
Int#(15) int1 = unpack(signExtend(val1));
|
| 142 |
|
|
Int#(15) int2 = unpack(signExtend(val2));
|
| 143 |
|
|
if(int1>=int2)
|
| 144 |
|
|
return (int1 >= (int2+4));
|
| 145 |
|
|
else
|
| 146 |
|
|
return (int2 >= (int1+4));
|
| 147 |
|
|
endfunction
|
| 148 |
|
|
|
| 149 |
|
|
function Bool absDiffGEFour12( Bit#(12) val1, Bit#(12) val2 );
|
| 150 |
|
|
Int#(13) int1 = unpack(signExtend(val1));
|
| 151 |
|
|
Int#(13) int2 = unpack(signExtend(val2));
|
| 152 |
|
|
if(int1>=int2)
|
| 153 |
|
|
return (int1 >= (int2+4));
|
| 154 |
|
|
else
|
| 155 |
|
|
return (int2 >= (int1+4));
|
| 156 |
|
|
endfunction
|
| 157 |
|
|
|
| 158 |
|
|
|
| 159 |
|
|
//-----------------------------------------------------------
|
| 160 |
|
|
// Prediction Module
|
| 161 |
|
|
//-----------------------------------------------------------
|
| 162 |
|
|
|
| 163 |
|
|
|
| 164 |
|
|
(* synthesize *)
|
| 165 |
|
|
module mkPrediction( IPrediction );
|
| 166 |
|
|
|
| 167 |
|
|
//Common state
|
| 168 |
|
|
FIFO#(EntropyDecOT) infifo <- mkSizedFIFO(prediction_infifo_size);
|
| 169 |
|
|
FIFO#(InverseTransOT) infifo_ITB <- mkSizedFIFO(prediction_infifo_ITB_size);
|
| 170 |
|
|
FIFO#(EntropyDecOT) outfifo <- mkFIFO;
|
| 171 |
|
|
Reg#(Bool) passFlag <- mkReg(True);
|
| 172 |
|
|
Reg#(Bit#(4)) blockNum <- mkReg(0);
|
| 173 |
|
|
Reg#(Bit#(4)) pixelNum <- mkReg(0);
|
| 174 |
|
|
|
| 175 |
|
|
Reg#(Bit#(PicWidthSz)) picWidth <- mkReg(maxPicWidthInMB);
|
| 176 |
|
|
Reg#(Bit#(PicHeightSz)) picHeight <- mkReg(0);
|
| 177 |
|
|
Reg#(Bit#(PicAreaSz)) firstMb <- mkReg(0);
|
| 178 |
|
|
Reg#(Bit#(PicAreaSz)) currMb <- mkReg(0);
|
| 179 |
|
|
Reg#(Bit#(PicAreaSz)) currMbHor <- mkReg(0);//horizontal position of currMb
|
| 180 |
|
|
Reg#(Bit#(PicHeightSz)) currMbVer <- mkReg(0);//vertical position of currMb
|
| 181 |
|
|
|
| 182 |
|
|
FIFOF#(OutState) outstatefifo <- mkFIFOF;
|
| 183 |
|
|
FIFOF#(NextOutput) nextoutputfifo <- mkFIFOF;
|
| 184 |
|
|
Reg#(Bit#(4)) outBlockNum <- mkReg(0);
|
| 185 |
|
|
Reg#(Bit#(4)) outPixelNum <- mkReg(0);
|
| 186 |
|
|
FIFO#(Vector#(4,Bit#(8))) predictedfifo <- mkSizedFIFO(prediction_predictedfifo_size);
|
| 187 |
|
|
Reg#(Bit#(1)) outChromaFlag <- mkReg(0);
|
| 188 |
|
|
Reg#(Bool) outFirstQPFlag <- mkReg(False);
|
| 189 |
|
|
|
| 190 |
|
|
DoNotFire donotfire <- mkDoNotFire();
|
| 191 |
|
|
|
| 192 |
|
|
//Reg#(Vector#(16,Bit#(8))) workVector <- mkRegU();
|
| 193 |
|
|
|
| 194 |
|
|
//Inter state
|
| 195 |
|
|
Interpolator interpolator <- mkInterpolator();
|
| 196 |
|
|
Reg#(InterState) interstate <- mkReg(Start);
|
| 197 |
|
|
Reg#(Bit#(PicAreaSz)) interPskipCount <- mkReg(0);
|
| 198 |
|
|
Reg#(Vector#(5,InterBlockMv)) interTopVal <- mkRegU();
|
| 199 |
|
|
Reg#(Vector#(4,InterBlockMv)) interLeftVal <- mkRegU();
|
| 200 |
|
|
Reg#(Vector#(4,InterBlockMv)) interTopLeftVal <- mkRegU();
|
| 201 |
|
|
FIFO#(MemReq#(TAdd#(PicWidthSz,2),32)) interMemReqQ <- mkFIFO;
|
| 202 |
|
|
Reg#(MemReq#(TAdd#(PicWidthSz,2),32)) interMemReqQdelay <- mkRegU();
|
| 203 |
|
|
FIFO#(MemResp#(32)) interMemRespQ <- mkFIFO;
|
| 204 |
|
|
Reg#(Bit#(3)) interReqCount <- mkReg(0);
|
| 205 |
|
|
Reg#(Bit#(3)) interRespCount <- mkReg(0);
|
| 206 |
|
|
|
| 207 |
33 |
jamey.hick |
Reg#(Bit#(2)) interStepCount <- mkReg(0);
|
| 208 |
2 |
jamey.hick |
Reg#(Bit#(2)) interMbPartNum <- mkReg(0);
|
| 209 |
|
|
Reg#(Bit#(2)) interSubMbPartNum <- mkReg(0);
|
| 210 |
|
|
Reg#(Bit#(2)) interPassingCount <- mkReg(0);
|
| 211 |
|
|
Reg#(Vector#(4,Bit#(4))) interRefIdxVector <- mkRegU();
|
| 212 |
|
|
Reg#(Vector#(4,Bit#(2))) interSubMbTypeVector <- mkRegU();
|
| 213 |
19 |
jamey.hick |
RFile1#(Bit#(4),Tuple2#(Bit#(14),Bit#(12))) interMvFile <- mkRFile1Full();
|
| 214 |
2 |
jamey.hick |
Reg#(Bit#(15)) interMvDiffTemp <- mkReg(0);
|
| 215 |
|
|
FIFO#(Tuple2#(Bit#(15),Bit#(13))) interMvDiff <- mkFIFO;
|
| 216 |
|
|
Reg#(Bit#(5)) interNewestMv <- mkReg(0);
|
| 217 |
33 |
jamey.hick |
|
| 218 |
|
|
// Registers for pipelining the interStage rule
|
| 219 |
|
|
|
| 220 |
|
|
Reg#(Bit#(3)) partWidthR <- mkRegU();
|
| 221 |
|
|
Reg#(Bit#(3)) partHeightR <- mkRegU();
|
| 222 |
|
|
Reg#(Bit#(3)) numPartR <- mkRegU();
|
| 223 |
|
|
Reg#(Bit#(3)) numSubPartR <- mkRegU();
|
| 224 |
|
|
Reg#(Bit#(2)) subMbTypeR <- mkRegU();
|
| 225 |
|
|
Reg#(Bool) calcmvR <- mkRegU();
|
| 226 |
|
|
Reg#(Bool) leftmvR <- mkRegU();
|
| 227 |
|
|
Reg#(Bit#(4)) refIndexR <- mkRegU();
|
| 228 |
|
|
Reg#(Vector#(3,InterBlockMv)) blockABCR <- mkRegU();
|
| 229 |
|
|
Reg#(Bit#(14)) mvhorfinalR <- mkRegU();
|
| 230 |
|
|
Reg#(Bit#(12)) mvverfinalR <- mkRegU();
|
| 231 |
|
|
Reg#(Bit#(5)) interNewestMvNextR <- mkRegU();
|
| 232 |
|
|
|
| 233 |
2 |
jamey.hick |
|
| 234 |
|
|
Reg#(Bit#(2)) interIPStepCount <- mkReg(0);
|
| 235 |
|
|
Reg#(Bit#(2)) interIPMbPartNum <- mkReg(0);
|
| 236 |
|
|
Reg#(Bit#(2)) interIPSubMbPartNum <- mkReg(0);
|
| 237 |
|
|
|
| 238 |
|
|
Reg#(Bit#(PicWidthSz)) interCurrMbDiff <- mkReg(0);
|
| 239 |
|
|
|
| 240 |
|
|
Reg#(Vector#(4,Bool)) interTopNonZeroTransCoeff <- mkRegU();
|
| 241 |
|
|
Reg#(Vector#(4,Bool)) interLeftNonZeroTransCoeff <- mkRegU();
|
| 242 |
|
|
FIFO#(Tuple2#(Bit#(2),Bit#(2))) interBSfifo <- mkSizedFIFO(32);
|
| 243 |
|
|
Reg#(Bool) interBSoutput <- mkReg(True);
|
| 244 |
|
|
FIFO#(InterBlockMv) interOutBlockMvfifo <- mkSizedFIFO(8);
|
| 245 |
|
|
|
| 246 |
|
|
|
| 247 |
|
|
//Intra state
|
| 248 |
|
|
Reg#(IntraState) intrastate <- mkReg(Start);
|
| 249 |
|
|
Reg#(Bit#(1)) intraChromaFlag <- mkReg(0);
|
| 250 |
|
|
FIFO#(MemReq#(TAdd#(PicWidthSz,2),68)) intraMemReqQ <- mkFIFO;
|
| 251 |
|
|
Reg#(MemReq#(TAdd#(PicWidthSz,2),68)) intraMemReqQdelay <- mkRegU;
|
| 252 |
|
|
FIFO#(MemResp#(68)) intraMemRespQ <- mkFIFO;
|
| 253 |
|
|
Reg#(Vector#(4,Bit#(4))) intra4x4typeLeft <- mkRegU();//15=unavailable, 14=inter-MB, 13=intra-non-4x4
|
| 254 |
|
|
Reg#(Vector#(4,Bit#(4))) intra4x4typeTop <- mkRegU();//15=unavailable, 14=inter-MB, 13=intra-non-4x4
|
| 255 |
|
|
Reg#(Bit#(1)) ppsconstrained_intra_pred_flag <- mkReg(0);
|
| 256 |
|
|
Reg#(Vector#(4,Bit#(40))) intraLeftVal <- mkRegU();
|
| 257 |
|
|
Reg#(Vector#(9,Bit#(8))) intraLeftValChroma0 <- mkRegU();
|
| 258 |
|
|
Reg#(Vector#(9,Bit#(8))) intraLeftValChroma1 <- mkRegU();
|
| 259 |
|
|
Reg#(Vector#(5,Bit#(32))) intraTopVal <- mkRegU();
|
| 260 |
|
|
Reg#(Vector#(4,Bit#(16))) intraTopValChroma0 <- mkRegU();
|
| 261 |
|
|
Reg#(Vector#(4,Bit#(16))) intraTopValChroma1 <- mkRegU();
|
| 262 |
|
|
Reg#(Bit#(32)) intraLeftValNext <- mkReg(0);
|
| 263 |
|
|
Reg#(Bit#(2)) intra16x16_pred_mode <- mkReg(0);
|
| 264 |
|
|
FIFO#(Bit#(4)) rem_intra4x4_pred_mode <- mkSizedFIFO(16);
|
| 265 |
|
|
FIFO#(Bit#(2)) intra_chroma_pred_mode <- mkFIFO;
|
| 266 |
|
|
Reg#(Bit#(4)) cur_intra4x4_pred_mode <- mkReg(0);
|
| 267 |
|
|
Reg#(Bit#(1)) intraChromaTopAvailable <- mkReg(0);
|
| 268 |
|
|
Reg#(Bit#(1)) intraChromaLeftAvailable <- mkReg(0);
|
| 269 |
|
|
|
| 270 |
|
|
Reg#(Bit#(3)) intraReqCount <- mkReg(0);
|
| 271 |
|
|
Reg#(Bit#(3)) intraRespCount <- mkReg(0);
|
| 272 |
|
|
Reg#(Bit#(4)) intraStepCount <- mkReg(0);
|
| 273 |
|
|
Reg#(Bit#(13)) intraSumA <- mkReg(0);
|
| 274 |
|
|
Reg#(Bit#(15)) intraSumB <- mkReg(0);
|
| 275 |
|
|
Reg#(Bit#(15)) intraSumC <- mkReg(0);
|
| 276 |
|
|
|
| 277 |
33 |
jamey.hick |
Reg#(Vector#(4,Bit#(8))) intraPredVector <- mkRegU();
|
| 278 |
2 |
jamey.hick |
|
| 279 |
|
|
//-----------------------------------------------------------
|
| 280 |
|
|
// Rules
|
| 281 |
|
|
|
| 282 |
|
|
//////////////////////////////////////////////////////////////////////////////
|
| 283 |
|
|
// rule stateMonitor ( True );
|
| 284 |
|
|
// if(predictedfifo.notEmpty())
|
| 285 |
|
|
// $display( "TRACE Prediction: stateMonitor predictedfifo.first() %0d", predictedfifo.first());////////////////////
|
| 286 |
|
|
// if(infifo.first() matches tagged ITBresidual .xdata)
|
| 287 |
|
|
// $display( "TRACE Prediction: stateMonitor infifo.first() %0d", xdata);////////////////////
|
| 288 |
|
|
// if(infifo.first() matches tagged ITBresidual .xdata)
|
| 289 |
|
|
// $display( "TRACE Prediction: stateMonitor outBlockNum outPixelNum outChromaFlag %0d %0d", outBlockNum, outPixelNum, outChromaFlag);////////////////////
|
| 290 |
|
|
// endrule
|
| 291 |
|
|
//////////////////////////////////////////////////////////////////////////////
|
| 292 |
|
|
|
| 293 |
|
|
rule checkFIFO ( True );
|
| 294 |
33 |
jamey.hick |
$display( "Trace Prediction: checkFIFO %h", infifo.first() );
|
| 295 |
2 |
jamey.hick |
endrule
|
| 296 |
33 |
jamey.hick |
rule checkFIFO_ITB ( True );
|
| 297 |
|
|
$display( "Trace Prediction: checkFIFO_ITB %h", infifo_ITB.first() );
|
| 298 |
|
|
endrule
|
| 299 |
|
|
rule checkFIFO_predicted ( True );
|
| 300 |
|
|
$display( "Trace Prediction: checkFIFO_predicted %h", predictedfifo.first() );
|
| 301 |
|
|
endrule
|
| 302 |
|
|
|
| 303 |
2 |
jamey.hick |
|
| 304 |
|
|
rule passing ( passFlag && !outstatefifo.notEmpty() && currMbHor
|
| 305 |
|
|
$display( "Trace Prediction: passing infifo packed %h", pack(infifo.first()));
|
| 306 |
|
|
case (infifo.first()) matches
|
| 307 |
|
|
tagged NewUnit . xdata :
|
| 308 |
|
|
begin
|
| 309 |
|
|
infifo.deq();
|
| 310 |
|
|
outfifo.enq(infifo.first());
|
| 311 |
|
|
$display("ccl4newunit");
|
| 312 |
|
|
$display("ccl4rbspbyte %h", xdata);
|
| 313 |
|
|
end
|
| 314 |
|
|
tagged SPSpic_width_in_mbs .xdata :
|
| 315 |
|
|
begin
|
| 316 |
|
|
infifo.deq();
|
| 317 |
|
|
outfifo.enq(infifo.first());
|
| 318 |
|
|
picWidth <= xdata;
|
| 319 |
|
|
interpolator.setPicWidth(xdata);
|
| 320 |
|
|
end
|
| 321 |
|
|
tagged SPSpic_height_in_map_units .xdata :
|
| 322 |
|
|
begin
|
| 323 |
|
|
infifo.deq();
|
| 324 |
|
|
outfifo.enq(infifo.first());
|
| 325 |
|
|
picHeight <= xdata;
|
| 326 |
|
|
interpolator.setPicHeight(xdata);
|
| 327 |
|
|
end
|
| 328 |
|
|
tagged PPSconstrained_intra_pred_flag .xdata :
|
| 329 |
|
|
begin
|
| 330 |
|
|
infifo.deq();
|
| 331 |
|
|
////outfifo.enq(infifo.first());
|
| 332 |
|
|
ppsconstrained_intra_pred_flag <= xdata;
|
| 333 |
|
|
end
|
| 334 |
|
|
tagged SHfirst_mb_in_slice .xdata :
|
| 335 |
|
|
begin
|
| 336 |
|
|
infifo.deq();
|
| 337 |
|
|
outfifo.enq(infifo.first());
|
| 338 |
|
|
firstMb <= xdata;
|
| 339 |
|
|
currMb <= xdata;
|
| 340 |
|
|
currMbHor <= xdata;
|
| 341 |
|
|
currMbVer <= 0;
|
| 342 |
|
|
intra4x4typeLeft <= replicate(15);
|
| 343 |
62 |
jamey.hick |
interTopLeftVal <= replicate(tagged NotInter 0);
|
| 344 |
2 |
jamey.hick |
if(xdata==0)
|
| 345 |
62 |
jamey.hick |
interLeftVal <= replicate(tagged NotInter 0);
|
| 346 |
2 |
jamey.hick |
outFirstQPFlag <= True;
|
| 347 |
|
|
end
|
| 348 |
|
|
tagged SDmb_skip_run .xdata : passFlag <= False;
|
| 349 |
|
|
tagged SDMmbtype .xdata : passFlag <= False;
|
| 350 |
|
|
tagged EndOfFile :
|
| 351 |
|
|
begin
|
| 352 |
|
|
infifo.deq();
|
| 353 |
|
|
outfifo.enq(infifo.first());
|
| 354 |
|
|
$display( "INFO Prediction: EndOfFile reached" );
|
| 355 |
|
|
//$finish(0);////////////////////////////////
|
| 356 |
|
|
end
|
| 357 |
|
|
default:
|
| 358 |
|
|
begin
|
| 359 |
|
|
infifo.deq();
|
| 360 |
|
|
outfifo.enq(infifo.first());
|
| 361 |
|
|
end
|
| 362 |
|
|
endcase
|
| 363 |
|
|
endrule
|
| 364 |
|
|
|
| 365 |
|
|
|
| 366 |
|
|
rule inputing ( !passFlag );
|
| 367 |
|
|
$display( "Trace Prediction: inputing infifo packed %h", pack(infifo.first()));
|
| 368 |
|
|
case (infifo.first()) matches
|
| 369 |
|
|
tagged SDmb_skip_run .xdata :
|
| 370 |
|
|
begin
|
| 371 |
|
|
if(interstate==Start && intrastate==Start)
|
| 372 |
|
|
begin
|
| 373 |
|
|
if(interPskipCount < xdata)
|
| 374 |
|
|
begin
|
| 375 |
|
|
if(!outstatefifo.notEmpty() || interCurrMbDiff
|
| 376 |
|
|
begin
|
| 377 |
|
|
$display( "Trace Prediction: passing SDmb_skip_run %0d", xdata);
|
| 378 |
|
|
outstatefifo.enq(Inter);
|
| 379 |
|
|
interstate <= InterPskip;
|
| 380 |
|
|
interReqCount <= 1;
|
| 381 |
|
|
interRespCount <= 1;
|
| 382 |
|
|
intra4x4typeLeft <= replicate(14);
|
| 383 |
|
|
intra4x4typeTop <= replicate(14);
|
| 384 |
62 |
jamey.hick |
interTopLeftVal <= update(interTopLeftVal , 0, (tagged NotInter 0));
|
| 385 |
|
|
interTopVal <= replicate(tagged NotInter 0);
|
| 386 |
2 |
jamey.hick |
interPskipCount <= interPskipCount+1;
|
| 387 |
|
|
interNewestMv <= 0;
|
| 388 |
|
|
interRefIdxVector <= replicate(0);
|
| 389 |
|
|
interCurrMbDiff <= interCurrMbDiff+1;
|
| 390 |
|
|
nextoutputfifo.enq(SkipMB);
|
| 391 |
|
|
end
|
| 392 |
|
|
else
|
| 393 |
|
|
donotfire.doNotFire();
|
| 394 |
|
|
end
|
| 395 |
|
|
else
|
| 396 |
|
|
begin
|
| 397 |
|
|
$display( "Trace Prediction: passing no SDmb_skip_run");
|
| 398 |
|
|
interPskipCount <= 0;
|
| 399 |
|
|
infifo.deq();
|
| 400 |
|
|
end
|
| 401 |
|
|
end
|
| 402 |
|
|
else
|
| 403 |
|
|
donotfire.doNotFire();
|
| 404 |
|
|
end
|
| 405 |
|
|
tagged SDMmbtype .xdata :
|
| 406 |
|
|
begin
|
| 407 |
|
|
if(interstate==Start && intrastate==Start)//not necessary (just need to keep inter from feeding predictedfifo or change intra state until intrastate==Start)
|
| 408 |
|
|
begin
|
| 409 |
|
|
infifo.deq();
|
| 410 |
|
|
$display( "INFO Prediction: SDMmbtype %0d", xdata);
|
| 411 |
|
|
if(mbPartPredMode(xdata,0)==Intra_16x16)
|
| 412 |
|
|
begin
|
| 413 |
|
|
if(!outstatefifo.notEmpty())
|
| 414 |
|
|
begin
|
| 415 |
|
|
outstatefifo.enq(Intra);
|
| 416 |
|
|
intrastate <= Intra16x16;
|
| 417 |
|
|
if(xdata matches tagged I_16x16 {intra16x16PredMode:.tempv1, codedBlockPatternChroma:.tempv2, codedBlockPatternLuma:.tempv3})
|
| 418 |
|
|
intra16x16_pred_mode <= tempv1;
|
| 419 |
|
|
else
|
| 420 |
|
|
$display( "ERROR Prediction: MacroblockLayer 5 sdmmbtype not I_16x16" );
|
| 421 |
|
|
intraReqCount <= 1;
|
| 422 |
|
|
intraRespCount <= 1;
|
| 423 |
62 |
jamey.hick |
interTopLeftVal <= replicate(tagged NotInter 1);
|
| 424 |
|
|
interLeftVal <= replicate(tagged NotInter 1);
|
| 425 |
|
|
interTopVal <= replicate(tagged NotInter 1);
|
| 426 |
2 |
jamey.hick |
end
|
| 427 |
|
|
else
|
| 428 |
|
|
donotfire.doNotFire();
|
| 429 |
|
|
end
|
| 430 |
|
|
else if(xdata==I_NxN)
|
| 431 |
|
|
begin
|
| 432 |
|
|
if(!outstatefifo.notEmpty())
|
| 433 |
|
|
begin
|
| 434 |
|
|
outstatefifo.enq(Intra4x4);
|
| 435 |
|
|
intrastate <= Intra4x4;
|
| 436 |
|
|
intraReqCount <= 1;
|
| 437 |
|
|
intraRespCount <= 1;
|
| 438 |
62 |
jamey.hick |
interTopLeftVal <= replicate(tagged NotInter 1);
|
| 439 |
|
|
interLeftVal <= replicate(tagged NotInter 1);
|
| 440 |
|
|
interTopVal <= replicate(tagged NotInter 1);
|
| 441 |
2 |
jamey.hick |
end
|
| 442 |
|
|
else
|
| 443 |
|
|
donotfire.doNotFire();
|
| 444 |
|
|
end
|
| 445 |
|
|
else if(xdata==I_PCM)
|
| 446 |
|
|
begin
|
| 447 |
|
|
$display( "ERROR Prediction: I_PCM not implemented yet");
|
| 448 |
|
|
$finish;////////////////////////////////////////////////////////////////////////////////////////
|
| 449 |
|
|
intra4x4typeLeft <= replicate(13);
|
| 450 |
|
|
intra4x4typeTop <= replicate(13);
|
| 451 |
62 |
jamey.hick |
interTopLeftVal <= replicate(tagged NotInter 1);
|
| 452 |
|
|
interLeftVal <= replicate(tagged NotInter 1);
|
| 453 |
|
|
interTopVal <= replicate(tagged NotInter 1);
|
| 454 |
2 |
jamey.hick |
end
|
| 455 |
|
|
else
|
| 456 |
|
|
begin
|
| 457 |
|
|
if(!outstatefifo.notEmpty() || interCurrMbDiff
|
| 458 |
|
|
begin
|
| 459 |
|
|
outstatefifo.enq(Inter);
|
| 460 |
|
|
case(xdata)
|
| 461 |
|
|
P_L0_16x16: interstate <= InterP16x16;
|
| 462 |
|
|
P_L0_L0_16x8: interstate <= InterP16x8;
|
| 463 |
|
|
P_L0_L0_8x16: interstate <= InterP8x16;
|
| 464 |
|
|
P_8x8: interstate <= InterP8x8;
|
| 465 |
|
|
P_8x8ref0: interstate <= InterP8x8ref0;
|
| 466 |
|
|
default: $display( "ERROR Prediction: passing SDMmbtype inter prediction unknown mbtype");
|
| 467 |
|
|
endcase
|
| 468 |
|
|
interReqCount <= 1;
|
| 469 |
|
|
interRespCount <= 1;
|
| 470 |
|
|
intra4x4typeLeft <= replicate(14);/////////////////////////////////////////////////////////////////////////////
|
| 471 |
|
|
intra4x4typeTop <= replicate(14);
|
| 472 |
62 |
jamey.hick |
interTopLeftVal <= update(interTopLeftVal , 0, (tagged NotInter 0));
|
| 473 |
|
|
interTopVal <= replicate(tagged NotInter 0);
|
| 474 |
2 |
jamey.hick |
interNewestMv <= 0;
|
| 475 |
|
|
interRefIdxVector <= replicate(0);
|
| 476 |
|
|
nextoutputfifo.enq(NonSkipMB);
|
| 477 |
|
|
end
|
| 478 |
|
|
else
|
| 479 |
|
|
donotfire.doNotFire();
|
| 480 |
|
|
end
|
| 481 |
|
|
interCurrMbDiff <= interCurrMbDiff+1;
|
| 482 |
|
|
end
|
| 483 |
|
|
else
|
| 484 |
|
|
donotfire.doNotFire();
|
| 485 |
|
|
end
|
| 486 |
|
|
tagged SDMMrem_intra4x4_pred_mode .xdata :
|
| 487 |
|
|
begin
|
| 488 |
|
|
infifo.deq();
|
| 489 |
|
|
////outfifo.enq(infifo.first());
|
| 490 |
|
|
rem_intra4x4_pred_mode.enq(xdata);
|
| 491 |
|
|
end
|
| 492 |
|
|
tagged SDMMintra_chroma_pred_mode .xdata :
|
| 493 |
|
|
begin
|
| 494 |
|
|
infifo.deq();
|
| 495 |
|
|
////outfifo.enq(infifo.first());
|
| 496 |
|
|
intra_chroma_pred_mode.enq(xdata);
|
| 497 |
|
|
end
|
| 498 |
|
|
tagged SDMMref_idx_l0 .xdata :
|
| 499 |
|
|
begin
|
| 500 |
|
|
infifo.deq();
|
| 501 |
|
|
////outfifo.enq(infifo.first());
|
| 502 |
|
|
interRefIdxVector <= update(interRefIdxVector,interPassingCount,xdata[3:0]);
|
| 503 |
|
|
if(interstate==InterP16x16 || interPassingCount==1)
|
| 504 |
|
|
interPassingCount <= 0;
|
| 505 |
|
|
else
|
| 506 |
|
|
interPassingCount <= interPassingCount+1;
|
| 507 |
|
|
end
|
| 508 |
|
|
tagged SDMMmvd_l0 .xdata :
|
| 509 |
|
|
begin
|
| 510 |
|
|
infifo.deq();
|
| 511 |
|
|
////outfifo.enq(infifo.first());
|
| 512 |
|
|
if(interPassingCount==1)
|
| 513 |
|
|
begin
|
| 514 |
|
|
Bit#(13) interMvDiffTemp2 = truncate(xdata);
|
| 515 |
|
|
interMvDiff.enq(tuple2(interMvDiffTemp,interMvDiffTemp2));
|
| 516 |
|
|
interPassingCount <= 0;
|
| 517 |
|
|
end
|
| 518 |
|
|
else
|
| 519 |
|
|
begin
|
| 520 |
|
|
interMvDiffTemp <= truncate(xdata);
|
| 521 |
|
|
interPassingCount <= interPassingCount+1;
|
| 522 |
|
|
end
|
| 523 |
|
|
end
|
| 524 |
|
|
tagged SDMSsub_mb_type .xdata :
|
| 525 |
|
|
begin
|
| 526 |
|
|
infifo.deq();
|
| 527 |
|
|
////outfifo.enq(infifo.first());
|
| 528 |
|
|
interSubMbTypeVector <= update(interSubMbTypeVector,interPassingCount,xdata);
|
| 529 |
|
|
interPassingCount <= interPassingCount+1;
|
| 530 |
|
|
end
|
| 531 |
|
|
tagged SDMSref_idx_l0 .xdata :
|
| 532 |
|
|
begin
|
| 533 |
|
|
infifo.deq();
|
| 534 |
|
|
////outfifo.enq(infifo.first());
|
| 535 |
|
|
interRefIdxVector <= update(interRefIdxVector,interPassingCount,xdata[3:0]);
|
| 536 |
|
|
interPassingCount <= interPassingCount+1;
|
| 537 |
|
|
end
|
| 538 |
|
|
tagged SDMSmvd_l0 .xdata :
|
| 539 |
|
|
begin
|
| 540 |
|
|
infifo.deq();
|
| 541 |
|
|
////outfifo.enq(infifo.first());
|
| 542 |
|
|
if(interPassingCount==1)
|
| 543 |
|
|
begin
|
| 544 |
|
|
Bit#(13) interMvDiffTemp2 = truncate(xdata);
|
| 545 |
|
|
interMvDiff.enq(tuple2(interMvDiffTemp,interMvDiffTemp2));
|
| 546 |
|
|
interPassingCount <= 0;
|
| 547 |
|
|
end
|
| 548 |
|
|
else
|
| 549 |
|
|
begin
|
| 550 |
|
|
interMvDiffTemp <= truncate(xdata);
|
| 551 |
|
|
interPassingCount <= interPassingCount+1;
|
| 552 |
|
|
end
|
| 553 |
|
|
end
|
| 554 |
|
|
default: passFlag <= True;
|
| 555 |
|
|
endcase
|
| 556 |
|
|
endrule
|
| 557 |
|
|
|
| 558 |
|
|
|
| 559 |
|
|
rule outputing ( currMbHor
|
| 560 |
|
|
Bit#(1) outputFlag = 0;
|
| 561 |
|
|
Vector#(4,Bit#(8)) outputVector = replicate(0);
|
| 562 |
|
|
Bit#(2) blockHor = {outBlockNum[2],outBlockNum[0]};
|
| 563 |
|
|
Bit#(2) blockVer = {outBlockNum[3],outBlockNum[1]};
|
| 564 |
|
|
Bit#(2) pixelVer = {outPixelNum[3],outPixelNum[2]};
|
| 565 |
|
|
Bit#(4) totalVer = {blockVer,pixelVer};
|
| 566 |
|
|
//$display( "Trace Prediction: outputing" );
|
| 567 |
|
|
if(outFirstQPFlag)
|
| 568 |
|
|
begin
|
| 569 |
|
|
if(infifo_ITB.first() matches tagged IBTmb_qp .xdata)
|
| 570 |
|
|
begin
|
| 571 |
|
|
infifo_ITB.deq();
|
| 572 |
|
|
outfifo.enq(IBTmb_qp {qpy:xdata.qpy,qpc:xdata.qpc});
|
| 573 |
|
|
outFirstQPFlag <= False;
|
| 574 |
|
|
$display( "Trace Prediction: outputing outFirstQP %h %h %h", outBlockNum, outPixelNum, xdata);
|
| 575 |
|
|
end
|
| 576 |
|
|
else
|
| 577 |
|
|
$display( "ERROR Prediction: outputing unexpected infifo_ITB.first()");
|
| 578 |
|
|
end
|
| 579 |
|
|
else if(nextoutputfifo.first() == SkipMB)
|
| 580 |
|
|
begin
|
| 581 |
|
|
if(interBSoutput && outChromaFlag==0 && outPixelNum==0)
|
| 582 |
|
|
begin
|
| 583 |
|
|
interBSoutput <= False;
|
| 584 |
|
|
interBSfifo.deq();
|
| 585 |
|
|
Bit#(2) tempHorBS = tpl_1(interBSfifo.first());
|
| 586 |
|
|
Bit#(2) tempVerBS = tpl_2(interBSfifo.first());
|
| 587 |
|
|
Bit#(3) horBS = (tempHorBS==3 ? 4 : (interLeftNonZeroTransCoeff[blockVer] ? 2 : zeroExtend(tempHorBS)));
|
| 588 |
|
|
Bit#(3) verBS = (tempVerBS==3 ? 4 : (interTopNonZeroTransCoeff[blockHor]&&blockVer!=0 ? 2 : zeroExtend(tempVerBS)));
|
| 589 |
|
|
outfifo.enq(PBbS {bShor:horBS,bSver:verBS});
|
| 590 |
|
|
interLeftNonZeroTransCoeff <= update(interLeftNonZeroTransCoeff, blockVer, False);
|
| 591 |
|
|
interTopNonZeroTransCoeff <= update(interTopNonZeroTransCoeff, blockHor, False);
|
| 592 |
|
|
$display( "Trace Prediction: outputing SkipMB bS %h %h %h %h", outBlockNum, outPixelNum, currMbHor, currMbVer);
|
| 593 |
|
|
end
|
| 594 |
|
|
else
|
| 595 |
|
|
begin
|
| 596 |
|
|
interBSoutput <= True;
|
| 597 |
|
|
outputVector = predictedfifo.first();
|
| 598 |
62 |
jamey.hick |
outfifo.enq(tagged PBoutput outputVector);
|
| 599 |
2 |
jamey.hick |
outputFlag = 1;
|
| 600 |
|
|
predictedfifo.deq();
|
| 601 |
|
|
$display( "Trace Prediction: outputing SkipMB out %h %h %h", outBlockNum, outPixelNum, outputVector);
|
| 602 |
|
|
end
|
| 603 |
|
|
end
|
| 604 |
|
|
else
|
| 605 |
|
|
begin
|
| 606 |
|
|
case ( infifo_ITB.first() ) matches
|
| 607 |
|
|
tagged IBTmb_qp .xdata :
|
| 608 |
|
|
begin
|
| 609 |
|
|
infifo_ITB.deq();
|
| 610 |
62 |
jamey.hick |
outfifo.enq( tagged IBTmb_qp {qpy:xdata.qpy,qpc:xdata.qpc});
|
| 611 |
2 |
jamey.hick |
outFirstQPFlag <= False;
|
| 612 |
|
|
$display( "Trace Prediction: outputing ITBmb_qp %h %h %h", outBlockNum, outPixelNum, xdata);
|
| 613 |
|
|
end
|
| 614 |
|
|
tagged ITBresidual .xdata :
|
| 615 |
|
|
begin
|
| 616 |
|
|
if(interBSoutput && outChromaFlag==0 && outPixelNum==0)
|
| 617 |
|
|
begin
|
| 618 |
|
|
interBSoutput <= False;
|
| 619 |
|
|
if(outstatefifo.first() != Inter)
|
| 620 |
62 |
jamey.hick |
outfifo.enq( tagged PBbS {bShor:(blockHor==0 ? 4 : 3),bSver:(blockVer==0 ? 4 : 3)});
|
| 621 |
2 |
jamey.hick |
else
|
| 622 |
|
|
begin
|
| 623 |
|
|
interBSfifo.deq();
|
| 624 |
|
|
Bit#(2) tempHorBS = tpl_1(interBSfifo.first());
|
| 625 |
|
|
Bit#(2) tempVerBS = tpl_2(interBSfifo.first());
|
| 626 |
|
|
Bit#(3) horBS = (tempHorBS==3 ? 4 : 2);
|
| 627 |
|
|
Bit#(3) verBS = (tempVerBS==3 ? 4 : 2);
|
| 628 |
62 |
jamey.hick |
outfifo.enq( tagged PBbS {bShor:horBS,bSver:verBS});
|
| 629 |
2 |
jamey.hick |
end
|
| 630 |
|
|
interLeftNonZeroTransCoeff <= update(interLeftNonZeroTransCoeff, blockVer, True);
|
| 631 |
|
|
interTopNonZeroTransCoeff <= update(interTopNonZeroTransCoeff, blockHor, True);
|
| 632 |
|
|
$display( "Trace Prediction: outputing ITBresidual bS %h %h %h %h %h", outChromaFlag, outBlockNum, outPixelNum, currMbHor, currMbVer);
|
| 633 |
|
|
end
|
| 634 |
|
|
else
|
| 635 |
|
|
begin
|
| 636 |
|
|
interBSoutput <= True;
|
| 637 |
|
|
Bit#(11) tempOutputValue = 0;
|
| 638 |
|
|
for(Integer ii=0; ii<4; ii=ii+1)
|
| 639 |
|
|
begin
|
| 640 |
|
|
tempOutputValue = signExtend(xdata[ii]) + zeroExtend((predictedfifo.first())[ii]);
|
| 641 |
|
|
if(tempOutputValue[10]==1)
|
| 642 |
|
|
outputVector[ii] = 0;
|
| 643 |
|
|
else if(tempOutputValue[9:0] > 255)
|
| 644 |
|
|
outputVector[ii] = 255;
|
| 645 |
|
|
else
|
| 646 |
|
|
outputVector[ii] = tempOutputValue[7:0];
|
| 647 |
|
|
end
|
| 648 |
62 |
jamey.hick |
outfifo.enq( tagged PBoutput outputVector);
|
| 649 |
2 |
jamey.hick |
infifo_ITB.deq();
|
| 650 |
|
|
predictedfifo.deq();
|
| 651 |
|
|
outputFlag = 1;
|
| 652 |
|
|
$display( "Trace Prediction: outputing ITBresidual out %h %h %h %h %h %h", outChromaFlag, outBlockNum, outPixelNum, predictedfifo.first(), xdata, outputVector);
|
| 653 |
|
|
end
|
| 654 |
|
|
end
|
| 655 |
|
|
tagged ITBcoeffLevelZeros :
|
| 656 |
|
|
begin
|
| 657 |
|
|
if(interBSoutput && outChromaFlag==0 && outPixelNum==0)
|
| 658 |
|
|
begin
|
| 659 |
|
|
interBSoutput <= False;
|
| 660 |
|
|
if(outstatefifo.first() != Inter)
|
| 661 |
62 |
jamey.hick |
outfifo.enq( tagged PBbS {bShor:(blockHor==0 ? 4 : 3),bSver:(blockVer==0 ? 4 : 3)});
|
| 662 |
2 |
jamey.hick |
else
|
| 663 |
|
|
begin
|
| 664 |
|
|
interBSfifo.deq();
|
| 665 |
|
|
Bit#(2) tempHorBS = tpl_1(interBSfifo.first());
|
| 666 |
|
|
Bit#(2) tempVerBS = tpl_2(interBSfifo.first());
|
| 667 |
|
|
Bit#(3) horBS = (tempHorBS==3 ? 4 : (interLeftNonZeroTransCoeff[blockVer] ? 2 : zeroExtend(tempHorBS)));
|
| 668 |
|
|
Bit#(3) verBS = (tempVerBS==3 ? 4 : (interTopNonZeroTransCoeff[blockHor]&&blockVer!=0 ? 2 : zeroExtend(tempVerBS)));
|
| 669 |
62 |
jamey.hick |
outfifo.enq( tagged PBbS {bShor:horBS,bSver:verBS});
|
| 670 |
2 |
jamey.hick |
end
|
| 671 |
|
|
interLeftNonZeroTransCoeff <= update(interLeftNonZeroTransCoeff, blockVer, False);
|
| 672 |
|
|
interTopNonZeroTransCoeff <= update(interTopNonZeroTransCoeff, blockHor, False);
|
| 673 |
|
|
$display( "Trace Prediction: outputing ITBcoeffLevelZeros bS %h %h %h %h %h", outChromaFlag, outBlockNum, outPixelNum, currMbHor, currMbVer);
|
| 674 |
|
|
end
|
| 675 |
|
|
else
|
| 676 |
|
|
begin
|
| 677 |
|
|
interBSoutput <= True;
|
| 678 |
|
|
if(outPixelNum == 12)
|
| 679 |
|
|
infifo_ITB.deq();
|
| 680 |
|
|
outputVector = predictedfifo.first();
|
| 681 |
62 |
jamey.hick |
outfifo.enq( tagged PBoutput outputVector);
|
| 682 |
2 |
jamey.hick |
outputFlag = 1;
|
| 683 |
|
|
predictedfifo.deq();
|
| 684 |
|
|
$display( "Trace Prediction: outputing ITBcoeffLevelZeros out %h %h %h %h %h", outChromaFlag, outBlockNum, outPixelNum, predictedfifo.first(), outputVector);
|
| 685 |
|
|
end
|
| 686 |
|
|
end
|
| 687 |
|
|
default: $display( "ERROR Prediction: outputing unknown infifo_ITB input" );
|
| 688 |
|
|
endcase
|
| 689 |
|
|
end
|
| 690 |
|
|
|
| 691 |
|
|
if(outputFlag == 1)
|
| 692 |
|
|
begin
|
| 693 |
|
|
$display("ccl4PBoutput %0d", outputVector[0]);
|
| 694 |
|
|
$display("ccl4PBoutput %0d", outputVector[1]);
|
| 695 |
|
|
$display("ccl4PBoutput %0d", outputVector[2]);
|
| 696 |
|
|
$display("ccl4PBoutput %0d", outputVector[3]);
|
| 697 |
|
|
|
| 698 |
|
|
if(outBlockNum==0 && pixelVer==0 && outChromaFlag==0 && currMb!=firstMb && picWidth>1)
|
| 699 |
|
|
begin
|
| 700 |
|
|
intraMemReqQ.enq(intraMemReqQdelay);
|
| 701 |
|
|
interMemReqQ.enq(interMemReqQdelay);
|
| 702 |
|
|
//$display( "TRACE Prediction: passing storing addr data");//////////////////
|
| 703 |
|
|
end
|
| 704 |
|
|
|
| 705 |
|
|
if(blockHor==3 || (blockHor[0]==1 && outChromaFlag==1) || (outstatefifo.first()==Intra4x4 && outChromaFlag==0))
|
| 706 |
|
|
begin
|
| 707 |
|
|
if(outChromaFlag==0)
|
| 708 |
|
|
begin
|
| 709 |
|
|
Bit#(32) intraLeftValNextTemp = intraLeftValNext;
|
| 710 |
|
|
if(totalVer==0 || (outstatefifo.first()==Intra4x4 && pixelVer==0))
|
| 711 |
|
|
begin
|
| 712 |
|
|
Bit#(32) tempValSet = select(intraTopVal,zeroExtend(blockHor));
|
| 713 |
|
|
intraLeftValNextTemp = zeroExtend(tempValSet[31:24]);
|
| 714 |
|
|
end
|
| 715 |
|
|
case(pixelVer)
|
| 716 |
|
|
0:intraLeftValNext <= {intraLeftValNextTemp[31:16],outputVector[3],intraLeftValNextTemp[7:0]};
|
| 717 |
|
|
1:intraLeftValNext <= {intraLeftValNextTemp[31:24],outputVector[3],intraLeftValNextTemp[15:0]};
|
| 718 |
|
|
2:intraLeftValNext <= {outputVector[3],intraLeftValNextTemp[23:0]};
|
| 719 |
|
|
3:
|
| 720 |
|
|
begin
|
| 721 |
|
|
intraLeftVal <= update(intraLeftVal,blockVer,{outputVector[3],intraLeftValNextTemp});
|
| 722 |
|
|
intraLeftValNext <= zeroExtend(outputVector[3]);
|
| 723 |
|
|
if(outstatefifo.first()==Intra4x4)
|
| 724 |
|
|
intra4x4typeLeft <= update(intra4x4typeLeft,blockVer,cur_intra4x4_pred_mode);
|
| 725 |
|
|
else if(outstatefifo.first()==Intra)
|
| 726 |
|
|
intra4x4typeLeft <= update(intra4x4typeLeft,blockVer,13);
|
| 727 |
|
|
else
|
| 728 |
|
|
intra4x4typeLeft <= update(intra4x4typeLeft,blockVer,14);
|
| 729 |
|
|
end
|
| 730 |
|
|
endcase
|
| 731 |
|
|
end
|
| 732 |
|
|
else
|
| 733 |
|
|
begin
|
| 734 |
|
|
if(outBlockNum[2]==0)
|
| 735 |
|
|
intraLeftValChroma0 <= update(intraLeftValChroma0,totalVer+1,outputVector[3]);
|
| 736 |
|
|
else
|
| 737 |
|
|
intraLeftValChroma1 <= update(intraLeftValChroma1,totalVer+1,outputVector[3]);
|
| 738 |
|
|
end
|
| 739 |
|
|
end
|
| 740 |
|
|
|
| 741 |
|
|
if(pixelVer==3 && (blockVer==3 || (blockVer[0]==1 && outChromaFlag==1) || (outstatefifo.first()==Intra4x4 && outChromaFlag==0)))
|
| 742 |
|
|
begin
|
| 743 |
|
|
if(outChromaFlag==0)
|
| 744 |
|
|
begin
|
| 745 |
|
|
intraTopVal <= update(intraTopVal,zeroExtend(blockHor),{outputVector[3],outputVector[2],outputVector[1],outputVector[0]});
|
| 746 |
|
|
if(outstatefifo.first()==Intra4x4)
|
| 747 |
|
|
intra4x4typeTop <= update(intra4x4typeTop,blockHor,cur_intra4x4_pred_mode);
|
| 748 |
|
|
else if(outstatefifo.first()==Intra)
|
| 749 |
|
|
intra4x4typeTop <= update(intra4x4typeTop,blockHor,13);
|
| 750 |
|
|
else
|
| 751 |
|
|
intra4x4typeTop <= update(intra4x4typeTop,blockHor,14);
|
| 752 |
|
|
end
|
| 753 |
|
|
else
|
| 754 |
|
|
begin
|
| 755 |
|
|
if(outBlockNum[2]==0)
|
| 756 |
|
|
begin
|
| 757 |
|
|
Vector#(4,Bit#(16)) intraTopValChroma0Next = intraTopValChroma0;
|
| 758 |
|
|
intraTopValChroma0Next[{blockHor[0],1'b0}] = {outputVector[1],outputVector[0]};
|
| 759 |
|
|
intraTopValChroma0Next[{blockHor[0],1'b1}] = {outputVector[3],outputVector[2]};
|
| 760 |
|
|
intraTopValChroma0 <= intraTopValChroma0Next;
|
| 761 |
|
|
end
|
| 762 |
|
|
else
|
| 763 |
|
|
begin
|
| 764 |
|
|
Vector#(4,Bit#(16)) intraTopValChroma1Next = intraTopValChroma1;
|
| 765 |
|
|
intraTopValChroma1Next[{blockHor[0],1'b0}] = {outputVector[1],outputVector[0]};
|
| 766 |
|
|
intraTopValChroma1Next[{blockHor[0],1'b1}] = {outputVector[3],outputVector[2]};
|
| 767 |
|
|
intraTopValChroma1 <= intraTopValChroma1Next;
|
| 768 |
|
|
end
|
| 769 |
|
|
end
|
| 770 |
|
|
end
|
| 771 |
|
|
|
| 772 |
|
|
if(outChromaFlag==1 && outBlockNum==7)
|
| 773 |
|
|
begin
|
| 774 |
|
|
Bit#(PicWidthSz) tempStoreAddr = truncate(currMbHor);
|
| 775 |
|
|
InterBlockMv outBlockMv = interOutBlockMvfifo.first();
|
| 776 |
|
|
if(outBlockMv matches tagged BlockMv .bdata)
|
| 777 |
|
|
begin
|
| 778 |
|
|
outBlockMv = (BlockMv {refIdx:bdata.refIdx,mvhor:bdata.mvhor,mvver:bdata.mvver,nonZeroTransCoeff:(interTopNonZeroTransCoeff[pixelVer]?1:0)});
|
| 779 |
|
|
interOutBlockMvfifo.deq();
|
| 780 |
|
|
end
|
| 781 |
|
|
else if(pixelVer==3)
|
| 782 |
|
|
interOutBlockMvfifo.deq();
|
| 783 |
|
|
if(pixelVer==3 && picWidth>1)
|
| 784 |
|
|
interMemReqQdelay <= StoreReq {addr:{tempStoreAddr,pixelVer},data:pack(outBlockMv)};
|
| 785 |
|
|
else
|
| 786 |
|
|
interMemReqQ.enq(StoreReq {addr:{tempStoreAddr,pixelVer},data:pack(outBlockMv)});
|
| 787 |
|
|
if(pixelVer>0)
|
| 788 |
|
|
begin
|
| 789 |
|
|
Bit#(4) intra4x4typeTopStore = ((outstatefifo.first()==Inter) ? 14 : ((outstatefifo.first()!=Intra4x4) ? 13: intra4x4typeTop[(pixelVer-1)]));
|
| 790 |
|
|
Bit#(32) intraTopValStore = intraTopVal[(pixelVer-1)];
|
| 791 |
|
|
Bit#(16) intraTopValChroma0Store = intraTopValChroma0[(pixelVer-1)];
|
| 792 |
|
|
Bit#(16) intraTopValChroma1Store = (pixelVer<3 ? intraTopValChroma1[(pixelVer-1)] : {outputVector[1],outputVector[0]});
|
| 793 |
|
|
Bit#(68) intraStore = {intra4x4typeTopStore,intraTopValChroma1Store,intraTopValChroma0Store,intraTopValStore};
|
| 794 |
|
|
intraMemReqQ.enq(StoreReq {addr:{tempStoreAddr,(pixelVer-1)},data:intraStore});
|
| 795 |
|
|
if(pixelVer==3)
|
| 796 |
|
|
begin
|
| 797 |
|
|
intra4x4typeTopStore = ((outstatefifo.first()==Inter) ? 14 : ((outstatefifo.first()!=Intra4x4) ? 13: intra4x4typeTop[3]));
|
| 798 |
|
|
intraTopValStore = intraTopVal[3];
|
| 799 |
|
|
intraTopValChroma0Store = intraTopValChroma0[3];
|
| 800 |
|
|
intraTopValChroma1Store = {outputVector[3],outputVector[2]};
|
| 801 |
|
|
intraStore = {intra4x4typeTopStore,intraTopValChroma1Store,intraTopValChroma0Store,intraTopValStore};
|
| 802 |
|
|
intraMemReqQdelay <= StoreReq {addr:{tempStoreAddr,2'b11},data:intraStore};
|
| 803 |
|
|
end
|
| 804 |
|
|
end
|
| 805 |
|
|
end
|
| 806 |
|
|
outPixelNum <= outPixelNum+4;
|
| 807 |
|
|
if(outPixelNum == 12)
|
| 808 |
|
|
begin
|
| 809 |
|
|
if(outChromaFlag==0)
|
| 810 |
|
|
begin
|
| 811 |
|
|
outBlockNum <= outBlockNum+1;
|
| 812 |
|
|
if(outBlockNum == 15)
|
| 813 |
|
|
outChromaFlag <= 1;
|
| 814 |
|
|
if(nextoutputfifo.first() == Intra4x4)
|
| 815 |
|
|
nextoutputfifo.deq();
|
| 816 |
|
|
end
|
| 817 |
|
|
else
|
| 818 |
|
|
begin
|
| 819 |
|
|
if(outBlockNum == 7)
|
| 820 |
|
|
begin
|
| 821 |
|
|
outBlockNum <= 0;
|
| 822 |
|
|
outChromaFlag <= 0;
|
| 823 |
|
|
currMb <= currMb+1;
|
| 824 |
|
|
currMbHor <= currMbHor+1;
|
| 825 |
|
|
interCurrMbDiff <= interCurrMbDiff-1;
|
| 826 |
|
|
outstatefifo.deq;
|
| 827 |
|
|
intrastate <= Start;
|
| 828 |
|
|
if(truncate(currMbHor)==picWidth-1 && currMbVer==picHeight-1)
|
| 829 |
|
|
interpolator.endOfFrame();
|
| 830 |
|
|
nextoutputfifo.deq();
|
| 831 |
|
|
end
|
| 832 |
|
|
else
|
| 833 |
|
|
outBlockNum <= outBlockNum+1;
|
| 834 |
|
|
end
|
| 835 |
|
|
end
|
| 836 |
|
|
end
|
| 837 |
|
|
endrule
|
| 838 |
|
|
|
| 839 |
|
|
|
| 840 |
|
|
rule currMbHorUpdate( !(currMbHor
|
| 841 |
|
|
Bit#(PicAreaSz) temp = zeroExtend(picWidth);
|
| 842 |
|
|
if((currMbHor >> 3) >= temp)
|
| 843 |
|
|
begin
|
| 844 |
|
|
currMbHor <= currMbHor - (temp << 3);
|
| 845 |
|
|
currMbVer <= currMbVer + 8;
|
| 846 |
|
|
end
|
| 847 |
|
|
else
|
| 848 |
|
|
begin
|
| 849 |
|
|
currMbHor <= currMbHor - temp;
|
| 850 |
|
|
currMbVer <= currMbVer + 1;
|
| 851 |
|
|
end
|
| 852 |
|
|
//$display( "Trace Prediction: currMbHorUpdate %h %h", currMbHor, currMbVer);
|
| 853 |
|
|
endrule
|
| 854 |
|
|
|
| 855 |
|
|
|
| 856 |
|
|
// inter prediction rules
|
| 857 |
|
|
|
| 858 |
|
|
rule interSendReq ( interReqCount>0 && currMbHor
|
| 859 |
|
|
Bit#(PicAreaSz) currMbHorTemp = currMbHor+zeroExtend(interCurrMbDiff)-1;
|
| 860 |
|
|
Bit#(PicAreaSz) currMbTemp = currMb+zeroExtend(interCurrMbDiff)-1;
|
| 861 |
|
|
if( currMbHorTemp >= zeroExtend(picWidth) )
|
| 862 |
|
|
currMbHorTemp = currMbHorTemp-zeroExtend(picWidth);
|
| 863 |
|
|
Bit#(PicWidthSz) temp2 = truncate(currMbHorTemp);
|
| 864 |
|
|
Bit#(TAdd#(PicWidthSz,2)) temp = 0;
|
| 865 |
|
|
Bool noMoreReq = False;
|
| 866 |
|
|
if( currMbTemp < zeroExtend(picWidth) )
|
| 867 |
|
|
noMoreReq = True;
|
| 868 |
|
|
else
|
| 869 |
|
|
begin
|
| 870 |
|
|
if(interReqCount<5)
|
| 871 |
|
|
begin
|
| 872 |
|
|
Bit#(2) temp3 = truncate(interReqCount-1);
|
| 873 |
|
|
temp = {temp2,temp3};
|
| 874 |
|
|
end
|
| 875 |
|
|
else if(interReqCount==5)
|
| 876 |
|
|
begin
|
| 877 |
|
|
if((currMbHorTemp+1)
|
| 878 |
|
|
temp = {(temp2+1),2'b00};
|
| 879 |
|
|
else if(currMbHorTemp>0 && currMbTemp-firstMb>zeroExtend(picWidth))
|
| 880 |
|
|
temp = {(temp2-1),2'b11};
|
| 881 |
|
|
else
|
| 882 |
|
|
noMoreReq = True;
|
| 883 |
|
|
end
|
| 884 |
|
|
else if(interReqCount==6)
|
| 885 |
|
|
begin
|
| 886 |
|
|
if((currMbHorTemp+1)0 && currMbTemp-firstMb>zeroExtend(picWidth))
|
| 887 |
|
|
temp = {(temp2-1),2'b11};
|
| 888 |
|
|
else
|
| 889 |
|
|
noMoreReq = True;
|
| 890 |
|
|
end
|
| 891 |
|
|
else
|
| 892 |
|
|
noMoreReq = True;
|
| 893 |
|
|
end
|
| 894 |
|
|
if(!noMoreReq)
|
| 895 |
|
|
begin
|
| 896 |
62 |
jamey.hick |
interMemReqQ.enq(tagged LoadReq temp);
|
| 897 |
2 |
jamey.hick |
interReqCount <= interReqCount+1;
|
| 898 |
|
|
//$display( "TRACE Prediction: interSendReq addr %0d",temp);///////////////////////
|
| 899 |
|
|
end
|
| 900 |
|
|
else
|
| 901 |
|
|
interReqCount <= 0;
|
| 902 |
|
|
$display( "Trace Prediction: interSendReq %h %h %h", interstate, interReqCount, temp);
|
| 903 |
|
|
endrule
|
| 904 |
|
|
|
| 905 |
|
|
|
| 906 |
|
|
rule interReceiveNoResp ( interRespCount>0 && currMbHor
|
| 907 |
|
|
Bit#(PicAreaSz) currMbHorTemp = currMbHor+zeroExtend(interCurrMbDiff)-1;
|
| 908 |
|
|
if( currMbHorTemp >= zeroExtend(picWidth) )
|
| 909 |
|
|
currMbHorTemp = currMbHorTemp-zeroExtend(picWidth);
|
| 910 |
|
|
interRespCount <= 0;
|
| 911 |
|
|
interStepCount <= 1;
|
| 912 |
|
|
interIPStepCount <= 1;
|
| 913 |
|
|
if(currMbHorTemp == 0)
|
| 914 |
|
|
begin
|
| 915 |
62 |
jamey.hick |
interLeftVal <= replicate(tagged NotInter 0);
|
| 916 |
|
|
interTopLeftVal <= replicate(tagged NotInter 0);
|
| 917 |
2 |
jamey.hick |
end
|
| 918 |
|
|
$display( "Trace Prediction: interReceiveNoResp %h %h", interstate, interRespCount);
|
| 919 |
|
|
endrule
|
| 920 |
|
|
|
| 921 |
|
|
|
| 922 |
|
|
rule interReceiveResp ( interRespCount>0 && interRespCount<7 && currMbHor
|
| 923 |
|
|
Bit#(PicAreaSz) currMbHorTemp = currMbHor+zeroExtend(interCurrMbDiff)-1;
|
| 924 |
|
|
Bit#(PicAreaSz) currMbTemp = currMb+zeroExtend(interCurrMbDiff)-1;
|
| 925 |
|
|
if( currMbHorTemp >= zeroExtend(picWidth) )
|
| 926 |
|
|
currMbHorTemp = currMbHorTemp-zeroExtend(picWidth);
|
| 927 |
|
|
Bool noMoreResp = False;
|
| 928 |
|
|
Bit#(2) temp2bit = 0;
|
| 929 |
|
|
InterBlockMv unpackedData = unpack(data);
|
| 930 |
|
|
Vector#(5,InterBlockMv) interTopValNext = interTopVal;
|
| 931 |
|
|
Vector#(4,InterBlockMv) interTopLeftValNext = interTopLeftVal;
|
| 932 |
|
|
if(interRespCount<5)
|
| 933 |
|
|
begin
|
| 934 |
|
|
temp2bit = truncate(interRespCount-1);
|
| 935 |
|
|
interTopValNext[temp2bit] = unpackedData;
|
| 936 |
|
|
if((interRespCount==4 || (interRespCount==1 && (interstate==InterPskip || interstate==InterP16x16 || interstate==InterP16x8)))
|
| 937 |
|
|
&& (!((currMbHorTemp+1)0 && currMbTemp-firstMb>zeroExtend(picWidth))))
|
| 938 |
|
|
noMoreResp = True;
|
| 939 |
|
|
end
|
| 940 |
|
|
else if(interRespCount==5)
|
| 941 |
|
|
begin
|
| 942 |
|
|
if((currMbHorTemp+1)
|
| 943 |
|
|
begin
|
| 944 |
|
|
interTopValNext[4] = unpackedData;
|
| 945 |
|
|
if(!(currMbHorTemp>0 && currMbTemp-firstMb>zeroExtend(picWidth)))
|
| 946 |
|
|
noMoreResp = True;
|
| 947 |
|
|
end
|
| 948 |
|
|
else
|
| 949 |
|
|
begin
|
| 950 |
|
|
interTopLeftValNext[0] = unpackedData;
|
| 951 |
|
|
noMoreResp = True;
|
| 952 |
|
|
end
|
| 953 |
|
|
end
|
| 954 |
|
|
else
|
| 955 |
|
|
begin
|
| 956 |
|
|
interTopLeftValNext[0] = unpackedData;
|
| 957 |
|
|
noMoreResp = True;
|
| 958 |
|
|
end
|
| 959 |
|
|
interMemRespQ.deq();
|
| 960 |
|
|
//$display( "TRACE Prediction: interReceiveResp data %h",data);///////////////////////
|
| 961 |
|
|
if(!noMoreResp)
|
| 962 |
|
|
interRespCount <= interRespCount+1;
|
| 963 |
|
|
else
|
| 964 |
|
|
begin
|
| 965 |
|
|
interRespCount <= 0;
|
| 966 |
|
|
interStepCount <= 1;
|
| 967 |
|
|
interIPStepCount <= 1;
|
| 968 |
|
|
if(currMbHorTemp == 0)
|
| 969 |
|
|
begin
|
| 970 |
62 |
jamey.hick |
interLeftVal <= replicate(tagged NotInter 0);
|
| 971 |
|
|
interTopLeftValNext = replicate(tagged NotInter 0);
|
| 972 |
2 |
jamey.hick |
end
|
| 973 |
|
|
end
|
| 974 |
|
|
interTopVal <= interTopValNext;
|
| 975 |
|
|
interTopLeftVal <= interTopLeftValNext;
|
| 976 |
|
|
$display( "Trace Prediction: interReceiveResp %h %h %h", interstate, interRespCount, data);
|
| 977 |
|
|
endrule
|
| 978 |
|
|
|
| 979 |
|
|
|
| 980 |
|
|
rule interProcessStep ( interStepCount>0 && currMbHor
|
| 981 |
|
|
Bit#(PicAreaSz) currMbTemp = currMb+zeroExtend(interCurrMbDiff)-1;
|
| 982 |
|
|
Bit#(2) blockHor = {interMbPartNum[0],interSubMbPartNum[0]};
|
| 983 |
|
|
Bit#(2) blockVer = {interMbPartNum[1],interSubMbPartNum[1]};
|
| 984 |
33 |
jamey.hick |
|
| 985 |
|
|
if(interStepCount == 1)
|
| 986 |
|
|
begin
|
| 987 |
|
|
Bit#(3) partWidth = 0;
|
| 988 |
|
|
Bit#(3) partHeight = 0;
|
| 989 |
|
|
Bit#(3) numPart = 1;
|
| 990 |
|
|
Bit#(3) numSubPart = 1;
|
| 991 |
|
|
Bit#(2) subMbType = 0;
|
| 992 |
|
|
Bool noBlockC = False;
|
| 993 |
|
|
Bool calcmv = False;
|
| 994 |
|
|
Bool leftmv = False;
|
| 995 |
|
|
if(interstate==InterPskip || interstate==InterP16x16)
|
| 996 |
|
|
begin
|
| 997 |
|
|
partWidth = 4;
|
| 998 |
|
|
partHeight = 4;
|
| 999 |
|
|
numPart = 1;
|
| 1000 |
|
|
calcmv = (interMbPartNum==0 && interSubMbPartNum==0);
|
| 1001 |
|
|
leftmv = (blockHor>0);
|
| 1002 |
|
|
end
|
| 1003 |
|
|
else if(interstate==InterP16x8)
|
| 1004 |
|
|
begin
|
| 1005 |
|
|
partWidth = 4;
|
| 1006 |
|
|
partHeight = 2;
|
| 1007 |
|
|
numPart = 2;
|
| 1008 |
|
|
if(interMbPartNum==2)
|
| 1009 |
|
|
noBlockC = True;
|
| 1010 |
|
|
calcmv = (interMbPartNum[0]==0 && interSubMbPartNum==0);
|
| 1011 |
|
|
leftmv = (blockHor>0);
|
| 1012 |
|
|
end
|
| 1013 |
|
|
else if(interstate==InterP8x16)
|
| 1014 |
|
|
begin
|
| 1015 |
|
|
partWidth = 2;
|
| 1016 |
|
|
partHeight = 4;
|
| 1017 |
|
|
numPart = 2;
|
| 1018 |
|
|
calcmv = (interMbPartNum[1]==0 && interSubMbPartNum==0);
|
| 1019 |
|
|
leftmv = !(blockVer>0);
|
| 1020 |
|
|
end
|
| 1021 |
|
|
else if(interstate==InterP8x8 || interstate==InterP8x8ref0)
|
| 1022 |
|
|
begin
|
| 1023 |
|
|
numPart = 4;
|
| 1024 |
|
|
subMbType = interSubMbTypeVector[interMbPartNum];
|
| 1025 |
|
|
numSubPart = numSubMbPart(subMbType);
|
| 1026 |
|
|
case(subMbType)
|
| 1027 |
|
|
0:
|
| 1028 |
|
|
begin
|
| 1029 |
|
|
partWidth = 2;
|
| 1030 |
|
|
partHeight = 2;
|
| 1031 |
|
|
if(interMbPartNum==3)
|
| 1032 |
|
|
noBlockC = True;
|
| 1033 |
|
|
calcmv = (interSubMbPartNum==0);
|
| 1034 |
|
|
leftmv = (blockHor[0]>0);
|
| 1035 |
|
|
end
|
| 1036 |
|
|
1:
|
| 1037 |
|
|
begin
|
| 1038 |
|
|
partWidth = 2;
|
| 1039 |
|
|
partHeight = 1;
|
| 1040 |
|
|
if(interSubMbPartNum==2)
|
| 1041 |
|
|
noBlockC = True;
|
| 1042 |
|
|
calcmv = (interSubMbPartNum[0]==0);
|
| 1043 |
|
|
leftmv = True;
|
| 1044 |
|
|
end
|
| 1045 |
|
|
2:
|
| 1046 |
|
|
begin
|
| 1047 |
|
|
partWidth = 1;
|
| 1048 |
|
|
partHeight = 2;
|
| 1049 |
|
|
calcmv = (interSubMbPartNum[1]==0);
|
| 1050 |
|
|
leftmv = False;
|
| 1051 |
|
|
end
|
| 1052 |
|
|
3:
|
| 1053 |
|
|
begin
|
| 1054 |
|
|
partWidth = 1;
|
| 1055 |
|
|
partHeight = 1;
|
| 1056 |
|
|
if(interSubMbPartNum==3)
|
| 1057 |
|
|
noBlockC = True;
|
| 1058 |
|
|
calcmv = True;
|
| 1059 |
|
|
end
|
| 1060 |
|
|
endcase
|
| 1061 |
|
|
end
|
| 1062 |
|
|
else
|
| 1063 |
|
|
$display( "ERROR Prediction: interProcessStep unexpected interstate");
|
| 1064 |
|
|
|
| 1065 |
|
|
Bit#(4) refIndex = ((interstate==InterPskip||interstate==InterP8x8ref0) ? 0 : interRefIdxVector[interMbPartNum]);
|
| 1066 |
62 |
jamey.hick |
Vector#(3,InterBlockMv) blockABC = replicate(tagged NotInter 0);
|
| 1067 |
33 |
jamey.hick |
if( currMbTemp-firstMb==0 && blockHor==0 )
|
| 1068 |
62 |
jamey.hick |
blockABC[0] = (tagged NotInter 0);
|
| 1069 |
33 |
jamey.hick |
else
|
| 1070 |
|
|
blockABC[0] = interLeftVal[blockVer];
|
| 1071 |
|
|
if( currMbTemp-firstMb
|
| 1072 |
62 |
jamey.hick |
blockABC[1] = (tagged NotInter 0);
|
| 1073 |
33 |
jamey.hick |
else
|
| 1074 |
|
|
blockABC[1] = interTopVal[blockHor];
|
| 1075 |
|
|
blockABC[2] = interTopVal[{1'b0,blockHor}+partWidth];
|
| 1076 |
62 |
jamey.hick |
if(noBlockC || blockABC[2]==(tagged NotInter 0))
|
| 1077 |
33 |
jamey.hick |
blockABC[2] = interTopLeftVal[blockVer];
|
| 1078 |
|
|
partWidthR <= partWidth;
|
| 1079 |
|
|
partHeightR <= partHeight;
|
| 1080 |
|
|
numPartR <= numPart;
|
| 1081 |
|
|
numSubPartR <= numSubPart;
|
| 1082 |
|
|
subMbTypeR <= subMbType;
|
| 1083 |
|
|
calcmvR <= calcmv;
|
| 1084 |
|
|
leftmvR <= leftmv;
|
| 1085 |
|
|
refIndexR <= refIndex;
|
| 1086 |
|
|
blockABCR <= blockABC;
|
| 1087 |
|
|
interStepCount <= 2;
|
| 1088 |
|
|
end
|
| 1089 |
|
|
else if(interStepCount==2)
|
| 1090 |
|
|
begin
|
| 1091 |
|
|
Bit#(3) partWidth = partWidthR;
|
| 1092 |
|
|
Bit#(3) partHeight = partHeightR;
|
| 1093 |
|
|
Bit#(3) numPart = numPartR;
|
| 1094 |
|
|
Bit#(3) numSubPart = numSubPartR;
|
| 1095 |
|
|
Bit#(2) subMbType = subMbTypeR;
|
| 1096 |
|
|
Bool calcmv = calcmvR;
|
| 1097 |
|
|
Bool leftmv = leftmvR;
|
| 1098 |
|
|
Bit#(4) refIndex = refIndexR;
|
| 1099 |
|
|
Vector#(3,InterBlockMv) blockABC = blockABCR;
|
| 1100 |
|
|
Bit#(14) mvhorfinal = 0;
|
| 1101 |
|
|
Bit#(12) mvverfinal = 0;
|
| 1102 |
|
|
Bit#(5) interNewestMvNext = 0;
|
| 1103 |
|
|
if(calcmv)//motion vector caculation
|
| 1104 |
|
|
begin
|
| 1105 |
|
|
Vector#(3,Int#(14)) mvhorABC = replicate(0);
|
| 1106 |
|
|
Vector#(3,Int#(12)) mvverABC = replicate(0);
|
| 1107 |
|
|
Bit#(2) validCount = 0;
|
| 1108 |
|
|
Bit#(14) mvhorPred = 0;
|
| 1109 |
|
|
Bit#(12) mvverPred = 0;
|
| 1110 |
|
|
for(Integer ii=0; ii<3; ii=ii+1)
|
| 1111 |
|
|
begin
|
| 1112 |
|
|
if(blockABC[ii] matches tagged BlockMv .xdata)
|
| 1113 |
|
|
begin
|
| 1114 |
|
|
mvhorABC[ii] = unpack(xdata.mvhor);
|
| 1115 |
|
|
mvverABC[ii] = unpack(xdata.mvver);
|
| 1116 |
|
|
if(xdata.refIdx == refIndex)
|
| 1117 |
|
|
begin
|
| 1118 |
|
|
validCount = validCount+1;
|
| 1119 |
|
|
mvhorPred = xdata.mvhor;
|
| 1120 |
|
|
mvverPred = xdata.mvver;
|
| 1121 |
|
|
end
|
| 1122 |
|
|
end
|
| 1123 |
|
|
else
|
| 1124 |
|
|
begin
|
| 1125 |
|
|
mvhorABC[ii] = 0;
|
| 1126 |
|
|
mvverABC[ii] = 0;
|
| 1127 |
|
|
end
|
| 1128 |
|
|
end
|
| 1129 |
|
|
if(validCount != 1)//median
|
| 1130 |
|
|
begin
|
| 1131 |
|
|
if(mvhorABC[0]>mvhorABC[1] && mvhorABC[0]>mvhorABC[2])
|
| 1132 |
|
|
mvhorPred = pack((mvhorABC[1]>mvhorABC[2]) ? mvhorABC[1] : mvhorABC[2]);
|
| 1133 |
|
|
else if(mvhorABC[0]
|
| 1134 |
|
|
mvhorPred = pack((mvhorABC[1]
|
| 1135 |
|
|
else
|
| 1136 |
|
|
mvhorPred = pack(mvhorABC[0]);
|
| 1137 |
|
|
if(mvverABC[0]>mvverABC[1] && mvverABC[0]>mvverABC[2])
|
| 1138 |
|
|
mvverPred = pack((mvverABC[1]>mvverABC[2]) ? mvverABC[1] : mvverABC[2]);
|
| 1139 |
|
|
else if(mvverABC[0]
|
| 1140 |
|
|
mvverPred = pack((mvverABC[1]
|
| 1141 |
|
|
else
|
| 1142 |
|
|
mvverPred = pack(mvverABC[0]);
|
| 1143 |
|
|
end
|
| 1144 |
|
|
if(interstate==InterPskip)
|
| 1145 |
|
|
begin
|
| 1146 |
|
|
for(Integer ii=0; ii<2; ii=ii+1)
|
| 1147 |
|
|
begin
|
| 1148 |
|
|
if(blockABC[ii] matches tagged BlockMv .xdata)
|
| 1149 |
|
|
begin
|
| 1150 |
|
|
if(xdata.refIdx==0 && xdata.mvhor==0 && xdata.mvver==0)
|
| 1151 |
|
|
begin
|
| 1152 |
|
|
mvhorPred = 0;
|
| 1153 |
|
|
mvverPred = 0;
|
| 1154 |
|
|
end
|
| 1155 |
|
|
end
|
| 1156 |
62 |
jamey.hick |
else if(blockABC[ii] matches tagged NotInter 0)
|
| 1157 |
33 |
jamey.hick |
begin
|
| 1158 |
|
|
mvhorPred = 0;
|
| 1159 |
|
|
mvverPred = 0;
|
| 1160 |
|
|
end
|
| 1161 |
|
|
end
|
| 1162 |
|
|
end
|
| 1163 |
|
|
else if(interstate==InterP16x8 || interstate==InterP8x16)
|
| 1164 |
|
|
begin
|
| 1165 |
|
|
InterBlockMv blockCheck;
|
| 1166 |
|
|
if(interstate==InterP16x8)
|
| 1167 |
|
|
begin
|
| 1168 |
|
|
if(interMbPartNum==0)
|
| 1169 |
|
|
blockCheck = blockABC[1];
|
| 1170 |
|
|
else
|
| 1171 |
|
|
blockCheck = blockABC[0];
|
| 1172 |
|
|
end
|
| 1173 |
|
|
else
|
| 1174 |
|
|
begin
|
| 1175 |
|
|
if(interMbPartNum==0)
|
| 1176 |
|
|
blockCheck = blockABC[0];
|
| 1177 |
|
|
else
|
| 1178 |
|
|
blockCheck = blockABC[2];
|
| 1179 |
|
|
end
|
| 1180 |
|
|
if(blockCheck matches tagged BlockMv .xdata &&& xdata.refIdx==refIndex)
|
| 1181 |
|
|
begin
|
| 1182 |
|
|
mvhorPred = xdata.mvhor;
|
| 1183 |
|
|
mvverPred = xdata.mvver;
|
| 1184 |
|
|
end
|
| 1185 |
|
|
end
|
| 1186 |
|
|
mvhorfinal = mvhorPred;
|
| 1187 |
|
|
mvverfinal = mvverPred;
|
| 1188 |
|
|
if(interstate!=InterPskip)
|
| 1189 |
|
|
begin
|
| 1190 |
|
|
mvhorfinal = truncate(tpl_1(interMvDiff.first()) + signExtend(mvhorPred));
|
| 1191 |
|
|
mvverfinal = truncate(tpl_2(interMvDiff.first()) + signExtend(mvverPred));
|
| 1192 |
|
|
interMvDiff.deq();
|
| 1193 |
|
|
end
|
| 1194 |
|
|
interMvFile.upd({interMbPartNum,interSubMbPartNum},tuple2(mvhorfinal,mvverfinal));
|
| 1195 |
|
|
interNewestMvNext = zeroExtend({interMbPartNum,interSubMbPartNum})+1;
|
| 1196 |
|
|
$display( "Trace Prediction: interProcessStep %h %h %h %h %h %h %h %h %h", interstate, interStepCount, interMbPartNum, interSubMbPartNum, pack(blockABC[0]), pack(blockABC[1]), pack(blockABC[2]), mvhorPred, mvverPred);
|
| 1197 |
|
|
end
|
| 1198 |
|
|
else
|
| 1199 |
|
|
begin
|
| 1200 |
|
|
if(leftmv)
|
| 1201 |
|
|
begin
|
| 1202 |
|
|
if(blockABC[0] matches tagged BlockMv .xdata)
|
| 1203 |
|
|
begin
|
| 1204 |
|
|
mvhorfinal = unpack(xdata.mvhor);
|
| 1205 |
|
|
mvverfinal = unpack(xdata.mvver);
|
| 1206 |
|
|
end
|
| 1207 |
|
|
else
|
| 1208 |
|
|
$display( "ERROR Prediction: interProcessStep unexpected blockABC[0]");
|
| 1209 |
|
|
end
|
| 1210 |
|
|
else
|
| 1211 |
|
|
begin
|
| 1212 |
|
|
if(blockABC[1] matches tagged BlockMv .xdata)
|
| 1213 |
|
|
begin
|
| 1214 |
|
|
mvhorfinal = unpack(xdata.mvhor);
|
| 1215 |
|
|
mvverfinal = unpack(xdata.mvver);
|
| 1216 |
|
|
end
|
| 1217 |
|
|
else
|
| 1218 |
|
|
$display( "ERROR Prediction: interProcessStep unexpected blockABC[1]");
|
| 1219 |
|
|
end
|
| 1220 |
|
|
end
|
| 1221 |
|
|
|
| 1222 |
|
|
mvhorfinalR <= mvhorfinal;
|
| 1223 |
|
|
mvverfinalR <= mvverfinal;
|
| 1224 |
|
|
interNewestMvNextR <= interNewestMvNext;
|
| 1225 |
|
|
interStepCount <= 3;
|
| 1226 |
|
|
|
| 1227 |
|
|
end
|
| 1228 |
|
|
else // stepCount == 3
|
| 1229 |
|
|
begin
|
| 1230 |
|
|
Bit#(2) tempBShor = 0;//bS calculation
|
| 1231 |
|
|
Bit#(2) tempBSver = 0;
|
| 1232 |
|
|
Bool allDone = False;
|
| 1233 |
|
|
Bit#(4) refIndex = refIndexR;
|
| 1234 |
|
|
Bit#(14) mvhorfinal = mvhorfinalR;
|
| 1235 |
|
|
Bit#(12) mvverfinal = mvverfinalR;
|
| 1236 |
|
|
Bit#(5) interNewestMvNext = interNewestMvNextR;
|
| 1237 |
|
|
|
| 1238 |
|
|
if(interLeftVal[blockVer] matches tagged BlockMv .xdata)
|
| 1239 |
2 |
jamey.hick |
begin
|
| 1240 |
33 |
jamey.hick |
if(xdata.nonZeroTransCoeff == 1)
|
| 1241 |
|
|
tempBShor = 2;
|
| 1242 |
|
|
else
|
| 1243 |
|
|
begin
|
| 1244 |
|
|
if(xdata.refIdx!=refIndex || absDiffGEFour14(mvhorfinal,xdata.mvhor) || absDiffGEFour12(mvverfinal,xdata.mvver))
|
| 1245 |
|
|
tempBShor = 1;
|
| 1246 |
|
|
else
|
| 1247 |
|
|
tempBShor = 0;
|
| 1248 |
|
|
end
|
| 1249 |
2 |
jamey.hick |
end
|
| 1250 |
33 |
jamey.hick |
else
|
| 1251 |
|
|
tempBShor = 3;
|
| 1252 |
|
|
if(interTopVal[blockHor] matches tagged BlockMv .xdata)
|
| 1253 |
2 |
jamey.hick |
begin
|
| 1254 |
33 |
jamey.hick |
if(xdata.nonZeroTransCoeff == 1)
|
| 1255 |
|
|
tempBSver = 2;
|
| 1256 |
|
|
else
|
| 1257 |
|
|
begin
|
| 1258 |
|
|
if(xdata.refIdx!=refIndex || absDiffGEFour14(mvhorfinal,xdata.mvhor) || absDiffGEFour12(mvverfinal,xdata.mvver))
|
| 1259 |
|
|
tempBSver = 1;
|
| 1260 |
|
|
else
|
| 1261 |
|
|
tempBSver = 0;
|
| 1262 |
|
|
end
|
| 1263 |
2 |
jamey.hick |
end
|
| 1264 |
33 |
jamey.hick |
else
|
| 1265 |
|
|
tempBSver = 3;
|
| 1266 |
|
|
interBSfifo.enq(tuple2(tempBShor,tempBSver));
|
| 1267 |
|
|
Vector#(5,InterBlockMv) interTopValNext = interTopVal;//update inter*Val
|
| 1268 |
|
|
Vector#(4,InterBlockMv) interLeftValNext = interLeftVal;
|
| 1269 |
|
|
Vector#(4,InterBlockMv) interTopLeftValNext = interTopLeftVal;
|
| 1270 |
|
|
interLeftValNext[blockVer] = (BlockMv {refIdx:refIndex,mvhor:mvhorfinal,mvver:mvverfinal,nonZeroTransCoeff:0});
|
| 1271 |
|
|
interTopValNext[blockHor] = (BlockMv {refIdx:refIndex,mvhor:mvhorfinal,mvver:mvverfinal,nonZeroTransCoeff:0});
|
| 1272 |
|
|
interTopLeftValNext[blockVer] = interTopVal[blockHor];
|
| 1273 |
|
|
interTopVal <= interTopValNext;
|
| 1274 |
|
|
interLeftVal <= interLeftValNext;
|
| 1275 |
|
|
interTopLeftVal <= interTopLeftValNext;
|
| 1276 |
|
|
if(blockVer == 3)
|
| 1277 |
|
|
interOutBlockMvfifo.enq(BlockMv {refIdx:refIndex,mvhor:mvhorfinal,mvver:mvverfinal,nonZeroTransCoeff:0});
|
| 1278 |
|
|
if(interSubMbPartNum == 3)//next step
|
| 1279 |
2 |
jamey.hick |
begin
|
| 1280 |
33 |
jamey.hick |
interSubMbPartNum <= 0;
|
| 1281 |
|
|
if(interMbPartNum == 3)
|
| 1282 |
|
|
begin
|
| 1283 |
|
|
interMbPartNum <= 0;
|
| 1284 |
|
|
allDone = True;
|
| 1285 |
|
|
interNewestMvNext = 16;
|
| 1286 |
|
|
end
|
| 1287 |
2 |
jamey.hick |
else
|
| 1288 |
33 |
jamey.hick |
interMbPartNum <= interMbPartNum+1;
|
| 1289 |
2 |
jamey.hick |
end
|
| 1290 |
33 |
jamey.hick |
else
|
| 1291 |
|
|
interSubMbPartNum <= interSubMbPartNum+1;
|
| 1292 |
|
|
if(interNewestMvNext > 0)
|
| 1293 |
|
|
interNewestMv <= interNewestMvNext;
|
| 1294 |
|
|
|
| 1295 |
|
|
// Check to see if we are done.
|
| 1296 |
|
|
|
| 1297 |
|
|
if(allDone)
|
| 1298 |
|
|
interStepCount <= 0;
|
| 1299 |
2 |
jamey.hick |
else
|
| 1300 |
33 |
jamey.hick |
interStepCount <= 1;
|
| 1301 |
|
|
|
| 1302 |
|
|
$display( "Trace Prediction: interProcessStep final %h %h %h %h %h %h %h",interstate,interStepCount,interMbPartNum,interSubMbPartNum,mvhorfinal,mvverfinal,interNewestMvNext);
|
| 1303 |
|
|
|
| 1304 |
|
|
end
|
| 1305 |
2 |
jamey.hick |
endrule
|
| 1306 |
|
|
|
| 1307 |
|
|
|
| 1308 |
|
|
rule interIPProcessStep ( interIPStepCount>0 && currMbHorzeroExtend({interIPMbPartNum,interIPSubMbPartNum}) );
|
| 1309 |
|
|
Bit#(PicAreaSz) currMbHorTemp = currMbHor+zeroExtend(interCurrMbDiff)-1;
|
| 1310 |
|
|
Bit#(PicHeightSz) currMbVerTemp = currMbVer;
|
| 1311 |
|
|
if( currMbHorTemp >= zeroExtend(picWidth) )
|
| 1312 |
|
|
begin
|
| 1313 |
|
|
currMbHorTemp = currMbHorTemp-zeroExtend(picWidth);
|
| 1314 |
|
|
currMbVerTemp = currMbVerTemp+1;
|
| 1315 |
|
|
end
|
| 1316 |
|
|
Bit#(2) blockHor = {interIPMbPartNum[0],interIPSubMbPartNum[0]};
|
| 1317 |
|
|
Bit#(2) blockVer = {interIPMbPartNum[1],interIPSubMbPartNum[1]};
|
| 1318 |
|
|
Bit#(3) numPart = 1;
|
| 1319 |
|
|
Bit#(3) numSubPart = 1;
|
| 1320 |
|
|
Bit#(2) subMbType = 0;
|
| 1321 |
|
|
if(interstate==InterPskip || interstate==InterP16x16)
|
| 1322 |
|
|
numPart = 1;
|
| 1323 |
|
|
else if(interstate==InterP16x8)
|
| 1324 |
|
|
numPart = 2;
|
| 1325 |
|
|
else if(interstate==InterP8x16)
|
| 1326 |
|
|
numPart = 2;
|
| 1327 |
|
|
else if(interstate==InterP8x8 || interstate==InterP8x8ref0)
|
| 1328 |
|
|
begin
|
| 1329 |
|
|
numPart = 4;
|
| 1330 |
|
|
subMbType = interSubMbTypeVector[interIPMbPartNum];
|
| 1331 |
|
|
numSubPart = numSubMbPart(subMbType);
|
| 1332 |
|
|
end
|
| 1333 |
|
|
else
|
| 1334 |
|
|
$display( "ERROR Prediction: interIPProcessStep unexpected interstate");
|
| 1335 |
|
|
Bit#(4) refIndex = ((interstate==InterPskip||interstate==InterP8x8ref0) ? 0 : interRefIdxVector[interIPMbPartNum]);
|
| 1336 |
|
|
Bit#(PicWidthSz) currMbHorT = truncate(currMbHorTemp);
|
| 1337 |
|
|
Bit#(TAdd#(PicWidthSz,2)) horTemp = {currMbHorT,blockHor};
|
| 1338 |
|
|
Bit#(TAdd#(PicHeightSz,4)) verTemp = {currMbVerTemp,blockVer,2'b00};
|
| 1339 |
|
|
IPBlockType btTemp = IP16x16;
|
| 1340 |
|
|
if(interstate==InterPskip || interstate==InterP16x16)
|
| 1341 |
|
|
btTemp = IP16x16;
|
| 1342 |
|
|
else if(interstate==InterP16x8)
|
| 1343 |
|
|
btTemp = IP16x8;
|
| 1344 |
|
|
else if(interstate==InterP8x16)
|
| 1345 |
|
|
btTemp = IP8x16;
|
| 1346 |
|
|
else
|
| 1347 |
|
|
begin
|
| 1348 |
|
|
case(subMbType)
|
| 1349 |
|
|
0: btTemp = IP8x8;
|
| 1350 |
|
|
1: btTemp = IP8x4;
|
| 1351 |
|
|
2: btTemp = IP4x8;
|
| 1352 |
|
|
3: btTemp = IP4x4;
|
| 1353 |
|
|
endcase
|
| 1354 |
|
|
end
|
| 1355 |
|
|
Bit#(14) mvhorTemp = tpl_1(interMvFile.sub({interIPMbPartNum,interIPSubMbPartNum}));
|
| 1356 |
|
|
Bit#(12) mvverTemp = tpl_2(interMvFile.sub({interIPMbPartNum,interIPSubMbPartNum}));
|
| 1357 |
|
|
if(interIPStepCount == 1)
|
| 1358 |
|
|
begin
|
| 1359 |
|
|
if(!(interstate==InterP8x8 || interstate==InterP8x8ref0))
|
| 1360 |
|
|
begin
|
| 1361 |
|
|
numPart = 4;
|
| 1362 |
|
|
Bit#(2) interIPMbPartNumTemp = interIPMbPartNum;
|
| 1363 |
|
|
if(btTemp==IP16x16)
|
| 1364 |
|
|
interIPMbPartNumTemp = 0;
|
| 1365 |
|
|
else if(btTemp==IP16x8 && interIPMbPartNumTemp[0]==1)
|
| 1366 |
|
|
interIPMbPartNumTemp = interIPMbPartNumTemp-1;
|
| 1367 |
|
|
else if(btTemp==IP8x16 && interIPMbPartNumTemp[1]==1)
|
| 1368 |
|
|
interIPMbPartNumTemp = interIPMbPartNumTemp-2;
|
| 1369 |
|
|
refIndex = ((interstate==InterPskip||interstate==InterP8x8ref0) ? 0 : interRefIdxVector[interIPMbPartNumTemp]);
|
| 1370 |
|
|
btTemp = IP8x8;
|
| 1371 |
|
|
mvhorTemp = tpl_1(interMvFile.sub({interIPMbPartNumTemp,2'b00}));
|
| 1372 |
|
|
mvverTemp = tpl_2(interMvFile.sub({interIPMbPartNumTemp,2'b00}));
|
| 1373 |
|
|
interpolator.request(IPLuma {refIdx:refIndex,hor:horTemp,ver:verTemp,mvhor:mvhorTemp,mvver:mvverTemp,bt:btTemp});
|
| 1374 |
|
|
end
|
| 1375 |
|
|
else
|
| 1376 |
|
|
interpolator.request(IPLuma {refIdx:refIndex,hor:horTemp,ver:verTemp,mvhor:mvhorTemp,mvver:mvverTemp,bt:btTemp});
|
| 1377 |
|
|
end
|
| 1378 |
|
|
else
|
| 1379 |
|
|
interpolator.request(IPChroma {refIdx:refIndex,uv:interIPStepCount[0],hor:horTemp,ver:truncate(verTemp>>1),mvhor:mvhorTemp,mvver:mvverTemp,bt:btTemp});
|
| 1380 |
|
|
if(interIPSubMbPartNum >= truncate(numSubPart-1))
|
| 1381 |
|
|
begin
|
| 1382 |
|
|
interIPSubMbPartNum <= 0;
|
| 1383 |
|
|
if(interIPMbPartNum >= truncate(numPart-1))
|
| 1384 |
|
|
begin
|
| 1385 |
|
|
interIPMbPartNum <= 0;
|
| 1386 |
|
|
interIPStepCount <= interIPStepCount+1;
|
| 1387 |
|
|
end
|
| 1388 |
|
|
else
|
| 1389 |
|
|
begin
|
| 1390 |
|
|
if(btTemp == IP16x8)
|
| 1391 |
|
|
interIPMbPartNum <= 2;
|
| 1392 |
|
|
else
|
| 1393 |
|
|
interIPMbPartNum <= interIPMbPartNum+1;
|
| 1394 |
|
|
end
|
| 1395 |
|
|
end
|
| 1396 |
|
|
else
|
| 1397 |
|
|
begin
|
| 1398 |
|
|
if(subMbType == 1)
|
| 1399 |
|
|
interIPSubMbPartNum <= 2;
|
| 1400 |
|
|
else
|
| 1401 |
|
|
interIPSubMbPartNum <= interIPSubMbPartNum+1;
|
| 1402 |
|
|
end
|
| 1403 |
|
|
$display( "Trace Prediction: interIPProcessStep %h %h %h %h %h %h %h %h %h %h", interstate, interIPStepCount, interIPMbPartNum, interIPSubMbPartNum, refIndex, horTemp, verTemp, mvhorTemp, mvverTemp, pack(btTemp));
|
| 1404 |
|
|
endrule
|
| 1405 |
|
|
|
| 1406 |
|
|
|
| 1407 |
|
|
rule interDone ( interstate!=Start && interReqCount==0 && interRespCount==0 && interStepCount==0 && interIPStepCount==0 );
|
| 1408 |
|
|
interstate <= Start;
|
| 1409 |
|
|
//$display( "Trace Prediction: interOutputTransfer %h %h", interstate, interOutputCount);
|
| 1410 |
|
|
endrule
|
| 1411 |
|
|
|
| 1412 |
|
|
|
| 1413 |
|
|
rule interOutputTransfer ( True );
|
| 1414 |
|
|
predictedfifo.enq(interpolator.first());
|
| 1415 |
|
|
interpolator.deq();
|
| 1416 |
|
|
//$display( "Trace Prediction: interOutputTransfer %h %h", interstate, interOutputCount);
|
| 1417 |
|
|
endrule
|
| 1418 |
|
|
|
| 1419 |
|
|
|
| 1420 |
|
|
|
| 1421 |
|
|
// intra prediction rules
|
| 1422 |
|
|
|
| 1423 |
|
|
rule intraSendReq ( intraReqCount>0 && currMbHor
|
| 1424 |
|
|
Bit#(PicWidthSz) temp2 = truncate(currMbHor);
|
| 1425 |
|
|
Bit#(TAdd#(PicWidthSz,2)) temp = 0;
|
| 1426 |
|
|
Bit#(1) noMoreReq = 0;
|
| 1427 |
|
|
if( currMb-firstMb < zeroExtend(picWidth) )
|
| 1428 |
|
|
noMoreReq = 1;
|
| 1429 |
|
|
else
|
| 1430 |
|
|
begin
|
| 1431 |
|
|
if(intraReqCount<5)
|
| 1432 |
|
|
begin
|
| 1433 |
|
|
Bit#(2) temp3 = truncate(intraReqCount-1);
|
| 1434 |
|
|
temp = {temp2,temp3};
|
| 1435 |
|
|
end
|
| 1436 |
|
|
else if(intraReqCount==5)
|
| 1437 |
|
|
begin
|
| 1438 |
|
|
if((currMbHor+1)
|
| 1439 |
|
|
temp = {(temp2+1),2'b00};
|
| 1440 |
|
|
else if(currMbHor>0 && currMb-firstMb>zeroExtend(picWidth))
|
| 1441 |
|
|
temp = {(temp2-1),2'b11};
|
| 1442 |
|
|
else
|
| 1443 |
|
|
noMoreReq = 1;
|
| 1444 |
|
|
end
|
| 1445 |
|
|
else if(intraReqCount==6)
|
| 1446 |
|
|
begin
|
| 1447 |
|
|
if((currMbHor+1)0 && currMb-firstMb>zeroExtend(picWidth))
|
| 1448 |
|
|
temp = {(temp2-1),2'b11};
|
| 1449 |
|
|
else
|
| 1450 |
|
|
noMoreReq = 1;
|
| 1451 |
|
|
end
|
| 1452 |
|
|
else
|
| 1453 |
|
|
noMoreReq = 1;
|
| 1454 |
|
|
end
|
| 1455 |
|
|
if(noMoreReq == 0)
|
| 1456 |
|
|
begin
|
| 1457 |
62 |
jamey.hick |
intraMemReqQ.enq(tagged LoadReq temp);
|
| 1458 |
2 |
jamey.hick |
intraReqCount <= intraReqCount+1;
|
| 1459 |
|
|
//$display( "TRACE Prediction: intraSendReq addr %0d",temp);///////////////////////
|
| 1460 |
|
|
end
|
| 1461 |
|
|
else
|
| 1462 |
|
|
intraReqCount <= 0;
|
| 1463 |
|
|
$display( "Trace Prediction: intraSendReq");
|
| 1464 |
|
|
endrule
|
| 1465 |
|
|
|
| 1466 |
|
|
|
| 1467 |
|
|
rule intraReceiveNoResp ( intraRespCount>0 && currMbHor
|
| 1468 |
|
|
intra4x4typeTop <= replicate(15);
|
| 1469 |
|
|
intraRespCount <= 0;
|
| 1470 |
|
|
intraStepCount <= 1;
|
| 1471 |
|
|
blockNum <= 0;
|
| 1472 |
|
|
pixelNum <= 0;
|
| 1473 |
62 |
jamey.hick |
interOutBlockMvfifo.enq(tagged NotInter 1);
|
| 1474 |
2 |
jamey.hick |
$display( "Trace Prediction: intraReceiveNoResp");
|
| 1475 |
|
|
endrule
|
| 1476 |
|
|
|
| 1477 |
|
|
|
| 1478 |
|
|
rule intraReceiveResp ( intraRespCount>0 && intraRespCount<7 && currMbHor
|
| 1479 |
|
|
Bit#(1) noMoreResp = 0;
|
| 1480 |
|
|
Bit#(2) temp2bit = 0;
|
| 1481 |
|
|
if(intraRespCount<5)
|
| 1482 |
|
|
begin
|
| 1483 |
|
|
temp2bit = truncate(intraRespCount-1);
|
| 1484 |
|
|
intra4x4typeTop <= update(intra4x4typeTop, temp2bit, data[67:64]);
|
| 1485 |
|
|
if(intraRespCount==4)
|
| 1486 |
|
|
begin
|
| 1487 |
|
|
Vector#(5,Bit#(32)) intraTopValTemp = intraTopVal;
|
| 1488 |
|
|
intraTopValTemp[3] = data[31:0];
|
| 1489 |
|
|
intraTopValTemp[4] = {data[31:24],data[31:24],data[31:24],data[31:24]};
|
| 1490 |
|
|
intraTopVal <= intraTopValTemp;
|
| 1491 |
|
|
if(!((currMbHor+1)0 && currMb-firstMb>zeroExtend(picWidth)))
|
| 1492 |
|
|
noMoreResp = 1;
|
| 1493 |
|
|
end
|
| 1494 |
|
|
else
|
| 1495 |
|
|
intraTopVal <= update(intraTopVal, intraRespCount-1, data[31:0]);
|
| 1496 |
|
|
intraTopValChroma0 <= update(intraTopValChroma0, temp2bit, data[47:32]);
|
| 1497 |
|
|
intraTopValChroma1 <= update(intraTopValChroma1, temp2bit, data[63:48]);
|
| 1498 |
|
|
end
|
| 1499 |
|
|
else if(intraRespCount==5)
|
| 1500 |
|
|
begin
|
| 1501 |
|
|
if((currMbHor+1)
|
| 1502 |
|
|
begin
|
| 1503 |
|
|
if(!(data[67:64]==15 || (data[67:64]==14 && ppsconstrained_intra_pred_flag==1)))
|
| 1504 |
|
|
intraTopVal <= update(intraTopVal, 4, data[31:0]);
|
| 1505 |
|
|
if(!(currMbHor>0 && currMb-firstMb>zeroExtend(picWidth)))
|
| 1506 |
|
|
noMoreResp = 1;
|
| 1507 |
|
|
end
|
| 1508 |
|
|
else
|
| 1509 |
|
|
begin
|
| 1510 |
|
|
Bit#(40) temp2 = intraLeftVal[0];
|
| 1511 |
|
|
intraLeftVal <= update(intraLeftVal, 0, {temp2[39:8],data[31:24]});
|
| 1512 |
|
|
intraLeftValChroma0 <= update(intraLeftValChroma0, 0, data[47:40]);
|
| 1513 |
|
|
intraLeftValChroma1 <= update(intraLeftValChroma1, 0, data[63:56]);
|
| 1514 |
|
|
noMoreResp = 1;
|
| 1515 |
|
|
end
|
| 1516 |
|
|
end
|
| 1517 |
|
|
else
|
| 1518 |
|
|
begin
|
| 1519 |
|
|
Bit#(40) temp2 = intraLeftVal[0];
|
| 1520 |
|
|
intraLeftVal <= update(intraLeftVal, 0, {temp2[39:8],data[31:24]});
|
| 1521 |
|
|
intraLeftValChroma0 <= update(intraLeftValChroma0, 0, data[47:40]);
|
| 1522 |
|
|
intraLeftValChroma1 <= update(intraLeftValChroma1, 0, data[63:56]);
|
| 1523 |
|
|
noMoreResp = 1;
|
| 1524 |
|
|
end
|
| 1525 |
|
|
intraMemRespQ.deq();
|
| 1526 |
|
|
//$display( "TRACE Prediction: intraReceiveResp data %h",data);///////////////////////
|
| 1527 |
|
|
if(noMoreResp == 0)
|
| 1528 |
|
|
intraRespCount <= intraRespCount+1;
|
| 1529 |
|
|
else
|
| 1530 |
|
|
begin
|
| 1531 |
|
|
intraRespCount <= 0;
|
| 1532 |
|
|
intraStepCount <= 1;
|
| 1533 |
|
|
blockNum <= 0;
|
| 1534 |
|
|
pixelNum <= 0;
|
| 1535 |
62 |
jamey.hick |
interOutBlockMvfifo.enq(tagged NotInter 1);
|
| 1536 |
2 |
jamey.hick |
end
|
| 1537 |
|
|
$display( "Trace Prediction: intraReceiveResp");
|
| 1538 |
|
|
endrule
|
| 1539 |
|
|
|
| 1540 |
|
|
|
| 1541 |
|
|
rule intraPredTypeStep ( intraStepCount==1 && !nextoutputfifo.notEmpty());
|
| 1542 |
|
|
Bit#(2) blockHor = {blockNum[2],blockNum[0]};
|
| 1543 |
|
|
Bit#(2) blockVer = {blockNum[3],blockNum[1]};
|
| 1544 |
|
|
Bit#(4) topType = select(intra4x4typeTop, blockHor);
|
| 1545 |
|
|
Bit#(4) leftType;
|
| 1546 |
|
|
if(currMbHor!=0 || blockNum!=0)
|
| 1547 |
|
|
leftType = select(intra4x4typeLeft, blockVer);
|
| 1548 |
|
|
else
|
| 1549 |
|
|
begin
|
| 1550 |
|
|
leftType = 15;
|
| 1551 |
|
|
intra4x4typeLeft <= replicate(15);
|
| 1552 |
|
|
end
|
| 1553 |
|
|
if(intrastate!=Intra4x4)
|
| 1554 |
|
|
begin
|
| 1555 |
|
|
intraStepCount <= intraStepCount+1;
|
| 1556 |
|
|
nextoutputfifo.enq(NonSkipMB);
|
| 1557 |
|
|
end
|
| 1558 |
|
|
else
|
| 1559 |
|
|
begin
|
| 1560 |
|
|
Bit#(1) topAvailable;
|
| 1561 |
|
|
Bit#(1) leftAvailable;
|
| 1562 |
|
|
if(topType==15 || (topType==14 && ppsconstrained_intra_pred_flag==1))
|
| 1563 |
|
|
topAvailable = 0;
|
| 1564 |
|
|
else
|
| 1565 |
|
|
topAvailable = 1;
|
| 1566 |
|
|
if(leftType==15 || (leftType==14 && ppsconstrained_intra_pred_flag==1))
|
| 1567 |
|
|
leftAvailable = 0;
|
| 1568 |
|
|
else
|
| 1569 |
|
|
leftAvailable = 1;
|
| 1570 |
|
|
Bit#(4) predType = 0;
|
| 1571 |
|
|
Bit#(4) remType = rem_intra4x4_pred_mode.first();
|
| 1572 |
|
|
Bit#(4) curType = 0;
|
| 1573 |
|
|
rem_intra4x4_pred_mode.deq();
|
| 1574 |
|
|
if(topAvailable==0 || leftAvailable==0)
|
| 1575 |
|
|
predType = 2;
|
| 1576 |
|
|
else
|
| 1577 |
|
|
begin
|
| 1578 |
|
|
Bit#(4) topType2 = topType;
|
| 1579 |
|
|
Bit#(4) leftType2 = leftType;
|
| 1580 |
|
|
if(topType>8)
|
| 1581 |
|
|
topType2 = 2;
|
| 1582 |
|
|
if(leftType>8)
|
| 1583 |
|
|
leftType2 = 2;
|
| 1584 |
|
|
if(topType2 > leftType2)
|
| 1585 |
|
|
predType = leftType2;
|
| 1586 |
|
|
else
|
| 1587 |
|
|
predType = topType2;
|
| 1588 |
|
|
end
|
| 1589 |
|
|
if(remType[3] == 1)
|
| 1590 |
|
|
curType = predType;
|
| 1591 |
|
|
else if(remType < predType)
|
| 1592 |
|
|
curType = remType;
|
| 1593 |
|
|
else
|
| 1594 |
|
|
curType = remType+1;
|
| 1595 |
|
|
cur_intra4x4_pred_mode <= curType;
|
| 1596 |
|
|
intraStepCount <= intraStepCount+1;
|
| 1597 |
|
|
if(blockNum == 15)
|
| 1598 |
|
|
nextoutputfifo.enq(Intra4x4PlusChroma);
|
| 1599 |
|
|
else
|
| 1600 |
|
|
nextoutputfifo.enq(Intra4x4);
|
| 1601 |
|
|
$display( "TRACE Prediction: intraPredTypeStep currMbHor currMbVer blockNum topType leftType predType remType curType %0d %0d %0d %0d %0d %0d %0d %0d",currMbHor,currMbVer,blockNum,topType,leftType,predType,remType,curType);//////////////////
|
| 1602 |
|
|
end
|
| 1603 |
|
|
//$display( "Trace Prediction: intraPredTypeStep");
|
| 1604 |
|
|
endrule
|
| 1605 |
|
|
|
| 1606 |
|
|
|
| 1607 |
|
|
rule intraProcessStep ( intraStepCount>1 );
|
| 1608 |
|
|
$display( "TRACE Prediction: intraProcessStep %0d %0d", blockNum, pixelNum);////////////////////
|
| 1609 |
|
|
//$display( "TRACE Prediction: intraProcessStep intraTopVal %h %h %h %h %h",intraTopVal[4],intraTopVal[3],intraTopVal[2],intraTopVal[1],intraTopVal[0]);/////////////////
|
| 1610 |
|
|
Bit#(1) outFlag = 0;
|
| 1611 |
|
|
Bit#(4) nextIntraStepCount = intraStepCount+1;
|
| 1612 |
|
|
Bit#(2) blockHor = {blockNum[2],blockNum[0]};
|
| 1613 |
|
|
Bit#(2) blockVer = {blockNum[3],blockNum[1]};
|
| 1614 |
|
|
Bit#(2) pixelVer = {pixelNum[3],pixelNum[2]};
|
| 1615 |
|
|
Vector#(4,Bit#(8)) predVector = replicate(0);
|
| 1616 |
|
|
|
| 1617 |
|
|
Bit#(4) topType = select(intra4x4typeTop, blockHor);
|
| 1618 |
|
|
Bit#(4) leftType = select(intra4x4typeLeft, blockVer);
|
| 1619 |
|
|
Bit#(1) topAvailable;
|
| 1620 |
|
|
Bit#(1) leftAvailable;
|
| 1621 |
|
|
if(topType==15 || (topType==14 && ppsconstrained_intra_pred_flag==1))
|
| 1622 |
|
|
topAvailable = 0;
|
| 1623 |
|
|
else
|
| 1624 |
|
|
topAvailable = 1;
|
| 1625 |
|
|
if(leftType==15 || (leftType==14 && ppsconstrained_intra_pred_flag==1))
|
| 1626 |
|
|
leftAvailable = 0;
|
| 1627 |
|
|
else
|
| 1628 |
|
|
leftAvailable = 1;
|
| 1629 |
|
|
if(blockNum==0 && pixelNum==0 && intraChromaFlag==0)
|
| 1630 |
|
|
begin
|
| 1631 |
|
|
intraChromaTopAvailable <= topAvailable;
|
| 1632 |
|
|
intraChromaLeftAvailable <= leftAvailable;
|
| 1633 |
|
|
end
|
| 1634 |
|
|
if(intrastate==Intra4x4 && intraChromaFlag==0)
|
| 1635 |
|
|
begin
|
| 1636 |
|
|
if(intraStepCount==2)
|
| 1637 |
|
|
begin
|
| 1638 |
|
|
outFlag = 1;
|
| 1639 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,blockVer);
|
| 1640 |
|
|
Bit#(32) topMidValSet = select(intraTopVal,zeroExtend(blockHor));
|
| 1641 |
|
|
Bit#(32) topRightValSet = select(intraTopVal,{1'b0,blockHor}+1);
|
| 1642 |
|
|
Bit#(72) topValSet;
|
| 1643 |
|
|
if((blockNum[3:2]==3 && blockNum[0]==1) || blockNum[1:0]==3)
|
| 1644 |
|
|
topValSet = {topMidValSet[31:24],topMidValSet[31:24],topMidValSet[31:24],topMidValSet[31:24],topMidValSet,leftValSet[7:0]};
|
| 1645 |
|
|
else
|
| 1646 |
|
|
topValSet = {topRightValSet,topMidValSet,leftValSet[7:0]};
|
| 1647 |
|
|
$display( "TRACE Prediction: intraProcessStep intra4x4 %0d %0d %h %h", cur_intra4x4_pred_mode, blockNum, leftValSet, topValSet);////////////////////
|
| 1648 |
|
|
Bit#(4) topSelect1 = 0;
|
| 1649 |
|
|
Bit#(4) topSelect2 = 0;
|
| 1650 |
|
|
Bit#(4) topSelect3 = 0;
|
| 1651 |
|
|
Bit#(3) leftSelect1 = 0;
|
| 1652 |
|
|
Bit#(3) leftSelect2 = 0;
|
| 1653 |
|
|
Bit#(3) leftSelect3 = 0;
|
| 1654 |
|
|
Bit#(10) tempVal1 = 0;
|
| 1655 |
|
|
Bit#(10) tempVal2 = 0;
|
| 1656 |
|
|
Bit#(10) tempVal3 = 0;
|
| 1657 |
|
|
case(cur_intra4x4_pred_mode)
|
| 1658 |
|
|
0://vertical
|
| 1659 |
|
|
begin
|
| 1660 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1661 |
|
|
begin
|
| 1662 |
|
|
topSelect1 = fromInteger(pixelHor);
|
| 1663 |
|
|
Bit#(8) topVal = intra4x4SelectTop(topValSet,topSelect1);
|
| 1664 |
|
|
predVector[pixelHor] = topVal;
|
| 1665 |
|
|
end
|
| 1666 |
|
|
end
|
| 1667 |
|
|
1://horizontal
|
| 1668 |
|
|
begin
|
| 1669 |
|
|
leftSelect1 = zeroExtend(pixelVer);
|
| 1670 |
|
|
Bit#(8) leftVal = intra4x4SelectLeft(leftValSet,leftSelect1);
|
| 1671 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1672 |
|
|
predVector[pixelHor] = leftVal;
|
| 1673 |
|
|
end
|
| 1674 |
|
|
2://dc
|
| 1675 |
|
|
begin
|
| 1676 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1677 |
|
|
begin
|
| 1678 |
|
|
Bit#(10) tempTopSum = zeroExtend(topValSet[15:8])+zeroExtend(topValSet[23:16])+zeroExtend(topValSet[31:24])+zeroExtend(topValSet[39:32]) + 2;
|
| 1679 |
|
|
Bit#(10) tempLeftSum = zeroExtend(leftValSet[15:8])+zeroExtend(leftValSet[23:16])+zeroExtend(leftValSet[31:24])+zeroExtend(leftValSet[39:32]) + 2;
|
| 1680 |
|
|
Bit#(11) tempTotalSum = zeroExtend(tempTopSum)+zeroExtend(tempLeftSum);
|
| 1681 |
|
|
Bit#(8) topSum = tempTopSum[9:2];
|
| 1682 |
|
|
Bit#(8) leftSum = tempLeftSum[9:2];
|
| 1683 |
|
|
Bit#(8) totalSum = tempTotalSum[10:3];
|
| 1684 |
|
|
if(topAvailable==1 && leftAvailable==1)
|
| 1685 |
|
|
predVector[pixelHor] = totalSum;
|
| 1686 |
|
|
else if(topAvailable==1)
|
| 1687 |
|
|
predVector[pixelHor] = topSum;
|
| 1688 |
|
|
else if(leftAvailable==1)
|
| 1689 |
|
|
predVector[pixelHor] = leftSum;
|
| 1690 |
|
|
else
|
| 1691 |
|
|
predVector[pixelHor] = 8'b10000000;
|
| 1692 |
|
|
end
|
| 1693 |
|
|
end
|
| 1694 |
|
|
3://diagonal down left
|
| 1695 |
|
|
begin
|
| 1696 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1697 |
|
|
begin
|
| 1698 |
|
|
Bit#(4) selectNum = fromInteger(pixelHor)+zeroExtend(pixelVer);
|
| 1699 |
|
|
if(pixelHor==3 && pixelVer==3)
|
| 1700 |
|
|
begin
|
| 1701 |
|
|
topSelect1 = 6;
|
| 1702 |
|
|
topSelect2 = 7;
|
| 1703 |
|
|
topSelect3 = 7;
|
| 1704 |
|
|
end
|
| 1705 |
|
|
else
|
| 1706 |
|
|
begin
|
| 1707 |
|
|
topSelect1 = selectNum;
|
| 1708 |
|
|
topSelect2 = selectNum+1;
|
| 1709 |
|
|
topSelect3 = selectNum+2;
|
| 1710 |
|
|
end
|
| 1711 |
|
|
tempVal1 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1712 |
|
|
tempVal2 = zeroExtend(intra4x4SelectTop(topValSet,topSelect2));
|
| 1713 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect3));
|
| 1714 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1715 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1716 |
|
|
end
|
| 1717 |
|
|
end
|
| 1718 |
|
|
4://diagonal down right
|
| 1719 |
|
|
begin
|
| 1720 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1721 |
|
|
begin
|
| 1722 |
|
|
if(fromInteger(pixelHor) > pixelVer)
|
| 1723 |
|
|
begin
|
| 1724 |
|
|
topSelect3 = fromInteger(pixelHor)-zeroExtend(pixelVer);
|
| 1725 |
|
|
topSelect2 = topSelect3-1;
|
| 1726 |
|
|
topSelect1 = topSelect3-2;
|
| 1727 |
|
|
tempVal1 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1728 |
|
|
tempVal2 = zeroExtend(intra4x4SelectTop(topValSet,topSelect2));
|
| 1729 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect3));
|
| 1730 |
|
|
end
|
| 1731 |
|
|
else if(fromInteger(pixelHor) < pixelVer)
|
| 1732 |
|
|
begin
|
| 1733 |
|
|
leftSelect3 = zeroExtend(pixelVer)-fromInteger(pixelHor);
|
| 1734 |
|
|
leftSelect2 = leftSelect3-1;
|
| 1735 |
|
|
leftSelect1 = leftSelect3-2;
|
| 1736 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1737 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1738 |
|
|
tempVal3 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect3));
|
| 1739 |
|
|
end
|
| 1740 |
|
|
else
|
| 1741 |
|
|
begin
|
| 1742 |
|
|
leftSelect1 = 0;
|
| 1743 |
|
|
leftSelect2 = -1;
|
| 1744 |
|
|
topSelect1 = 0;
|
| 1745 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1746 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1747 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1748 |
|
|
end
|
| 1749 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1750 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1751 |
|
|
end
|
| 1752 |
|
|
end
|
| 1753 |
|
|
5://vertical right
|
| 1754 |
|
|
begin
|
| 1755 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1756 |
|
|
begin
|
| 1757 |
|
|
Bit#(4) tempPixelHor = fromInteger(pixelHor);
|
| 1758 |
|
|
Bit#(4) zVR = (tempPixelHor<<1)-zeroExtend(pixelVer);
|
| 1759 |
|
|
if(zVR<=6 && zVR>=0)
|
| 1760 |
|
|
begin
|
| 1761 |
|
|
topSelect3 = fromInteger(pixelHor)-zeroExtend(pixelVer>>1);
|
| 1762 |
|
|
topSelect2 = topSelect3-1;
|
| 1763 |
|
|
if(zVR==1 || zVR==3 || zVR==5)
|
| 1764 |
|
|
topSelect1 = topSelect3-2;
|
| 1765 |
|
|
else
|
| 1766 |
|
|
topSelect1 = topSelect3;
|
| 1767 |
|
|
tempVal1 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1768 |
|
|
tempVal2 = zeroExtend(intra4x4SelectTop(topValSet,topSelect2));
|
| 1769 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect3));
|
| 1770 |
|
|
end
|
| 1771 |
|
|
else if(zVR==-1)
|
| 1772 |
|
|
begin
|
| 1773 |
|
|
leftSelect1 = 0;
|
| 1774 |
|
|
leftSelect2 = -1;
|
| 1775 |
|
|
topSelect1 = 0;
|
| 1776 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1777 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1778 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1779 |
|
|
end
|
| 1780 |
|
|
else
|
| 1781 |
|
|
begin
|
| 1782 |
|
|
leftSelect1 = zeroExtend(pixelVer)-1;
|
| 1783 |
|
|
leftSelect2 = leftSelect1-1;
|
| 1784 |
|
|
leftSelect3 = leftSelect1-2;
|
| 1785 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1786 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1787 |
|
|
tempVal3 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect3));
|
| 1788 |
|
|
end
|
| 1789 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1790 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1791 |
|
|
end
|
| 1792 |
|
|
end
|
| 1793 |
|
|
6://horizontal down
|
| 1794 |
|
|
begin
|
| 1795 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1796 |
|
|
begin
|
| 1797 |
|
|
Bit#(4) tempPixelVer = zeroExtend(pixelVer);
|
| 1798 |
|
|
Bit#(4) zHD = (tempPixelVer<<1)-fromInteger(pixelHor);
|
| 1799 |
|
|
if(zHD<=6 && zHD>=0)
|
| 1800 |
|
|
begin
|
| 1801 |
|
|
leftSelect3 = zeroExtend(pixelVer)-fromInteger(pixelHor/2);
|
| 1802 |
|
|
leftSelect2 = leftSelect3-1;
|
| 1803 |
|
|
if(zHD==1 || zHD==3 || zHD==5)
|
| 1804 |
|
|
leftSelect1 = leftSelect3-2;
|
| 1805 |
|
|
else
|
| 1806 |
|
|
leftSelect1 = leftSelect3;
|
| 1807 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1808 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1809 |
|
|
tempVal3 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect3));
|
| 1810 |
|
|
end
|
| 1811 |
|
|
else if(zHD==-1)
|
| 1812 |
|
|
begin
|
| 1813 |
|
|
leftSelect1 = 0;
|
| 1814 |
|
|
leftSelect2 = -1;
|
| 1815 |
|
|
topSelect1 = 0;
|
| 1816 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1817 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1818 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1819 |
|
|
end
|
| 1820 |
|
|
else
|
| 1821 |
|
|
begin
|
| 1822 |
|
|
topSelect1 = fromInteger(pixelHor)-1;
|
| 1823 |
|
|
topSelect2 = topSelect1-1;
|
| 1824 |
|
|
topSelect3 = topSelect1-2;
|
| 1825 |
|
|
tempVal1 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1826 |
|
|
tempVal2 = zeroExtend(intra4x4SelectTop(topValSet,topSelect2));
|
| 1827 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect3));
|
| 1828 |
|
|
end
|
| 1829 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1830 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1831 |
|
|
end
|
| 1832 |
|
|
end
|
| 1833 |
|
|
7://vertical left
|
| 1834 |
|
|
begin
|
| 1835 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1836 |
|
|
begin
|
| 1837 |
|
|
topSelect1 = fromInteger(pixelHor)+zeroExtend(pixelVer>>1);
|
| 1838 |
|
|
topSelect2 = topSelect1+1;
|
| 1839 |
|
|
if(pixelVer==1 || pixelVer==3)
|
| 1840 |
|
|
topSelect3 = topSelect1+2;
|
| 1841 |
|
|
else
|
| 1842 |
|
|
topSelect3 = topSelect1;
|
| 1843 |
|
|
tempVal1 = zeroExtend(intra4x4SelectTop(topValSet,topSelect1));
|
| 1844 |
|
|
tempVal2 = zeroExtend(intra4x4SelectTop(topValSet,topSelect2));
|
| 1845 |
|
|
tempVal3 = zeroExtend(intra4x4SelectTop(topValSet,topSelect3));
|
| 1846 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1847 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1848 |
|
|
end
|
| 1849 |
|
|
end
|
| 1850 |
|
|
8://horizontal up
|
| 1851 |
|
|
begin
|
| 1852 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1853 |
|
|
begin
|
| 1854 |
|
|
Bit#(4) tempPixelVer = zeroExtend(pixelVer);
|
| 1855 |
|
|
Bit#(4) zHU = (tempPixelVer<<1)+fromInteger(pixelHor);
|
| 1856 |
|
|
if(zHU<=4)
|
| 1857 |
|
|
begin
|
| 1858 |
|
|
leftSelect1 = zeroExtend(pixelVer)+fromInteger(pixelHor/2);
|
| 1859 |
|
|
leftSelect2 = leftSelect1+1;
|
| 1860 |
|
|
if(zHU==1 || zHU==3)
|
| 1861 |
|
|
leftSelect3 = leftSelect1+2;
|
| 1862 |
|
|
else
|
| 1863 |
|
|
leftSelect3 = leftSelect1;
|
| 1864 |
|
|
end
|
| 1865 |
|
|
else
|
| 1866 |
|
|
begin
|
| 1867 |
|
|
if(zHU==5)
|
| 1868 |
|
|
leftSelect1 = 2;
|
| 1869 |
|
|
else
|
| 1870 |
|
|
leftSelect1 = 3;
|
| 1871 |
|
|
leftSelect2 = 3;
|
| 1872 |
|
|
leftSelect3 = 3;
|
| 1873 |
|
|
end
|
| 1874 |
|
|
tempVal1 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect1));
|
| 1875 |
|
|
tempVal2 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect2));
|
| 1876 |
|
|
tempVal3 = zeroExtend(intra4x4SelectLeft(leftValSet,leftSelect3));
|
| 1877 |
|
|
Bit#(10) predVal = tempVal1 + (tempVal2<<1) + tempVal3 + 2;
|
| 1878 |
|
|
predVector[pixelHor] = predVal[9:2];
|
| 1879 |
|
|
end
|
| 1880 |
|
|
end
|
| 1881 |
|
|
default: $display( "ERROR Prediction: intraProcessStep intra4x4 unknown cur_intra4x4_pred_mode");
|
| 1882 |
|
|
endcase
|
| 1883 |
|
|
end
|
| 1884 |
|
|
else
|
| 1885 |
|
|
$display( "ERROR Prediction: intraProcessStep intra4x4 unknown intraStepCount");
|
| 1886 |
|
|
end
|
| 1887 |
|
|
else if(intrastate==Intra16x16 && intraChromaFlag==0)
|
| 1888 |
|
|
begin
|
| 1889 |
|
|
//$display( "TRACE Prediction: intraProcessStep intra16x16 %0d %0d %0d %h", intra16x16_pred_mode, currMb, blockNum, select(intraTopVal,blockHor));/////////////////
|
| 1890 |
|
|
case(intra16x16_pred_mode)
|
| 1891 |
|
|
0://vertical
|
| 1892 |
|
|
begin
|
| 1893 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1894 |
|
|
begin
|
| 1895 |
|
|
Bit#(32) topValSet = select(intraTopVal,blockHor);
|
| 1896 |
|
|
Bit#(8) topVal = select32to8(topValSet,fromInteger(pixelHor));
|
| 1897 |
|
|
predVector[pixelHor] = topVal;
|
| 1898 |
|
|
end
|
| 1899 |
|
|
outFlag = 1;
|
| 1900 |
|
|
end
|
| 1901 |
|
|
1://horizontal
|
| 1902 |
|
|
begin
|
| 1903 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,blockVer);
|
| 1904 |
|
|
Bit#(8) leftVal = intra4x4SelectLeft(leftValSet,zeroExtend(pixelVer));
|
| 1905 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1906 |
|
|
predVector[pixelHor] = leftVal;
|
| 1907 |
|
|
outFlag = 1;
|
| 1908 |
|
|
end
|
| 1909 |
|
|
2://dc
|
| 1910 |
|
|
begin
|
| 1911 |
|
|
case(intraStepCount)
|
| 1912 |
|
|
2:
|
| 1913 |
|
|
begin
|
| 1914 |
|
|
if(topAvailable == 1)
|
| 1915 |
|
|
begin
|
| 1916 |
|
|
Bit#(32) topValSet = select(intraTopVal,0);
|
| 1917 |
|
|
intraSumA <= zeroExtend(topValSet[7:0])+zeroExtend(topValSet[15:8])+zeroExtend(topValSet[23:16])+zeroExtend(topValSet[31:24]);
|
| 1918 |
|
|
end
|
| 1919 |
|
|
else
|
| 1920 |
|
|
begin
|
| 1921 |
|
|
intraSumA <= 0;
|
| 1922 |
|
|
nextIntraStepCount = 6;
|
| 1923 |
|
|
end
|
| 1924 |
|
|
end
|
| 1925 |
|
|
3:
|
| 1926 |
|
|
begin
|
| 1927 |
|
|
Bit#(32) topValSet = select(intraTopVal,1);
|
| 1928 |
|
|
intraSumA <= intraSumA+zeroExtend(topValSet[7:0])+zeroExtend(topValSet[15:8])+zeroExtend(topValSet[23:16])+zeroExtend(topValSet[31:24]);
|
| 1929 |
|
|
end
|
| 1930 |
|
|
4:
|
| 1931 |
|
|
begin
|
| 1932 |
|
|
Bit#(32) topValSet = select(intraTopVal,2);
|
| 1933 |
|
|
intraSumA <= intraSumA+zeroExtend(topValSet[7:0])+zeroExtend(topValSet[15:8])+zeroExtend(topValSet[23:16])+zeroExtend(topValSet[31:24]);
|
| 1934 |
|
|
end
|
| 1935 |
|
|
5:
|
| 1936 |
|
|
begin
|
| 1937 |
|
|
Bit#(32) topValSet = select(intraTopVal,3);
|
| 1938 |
|
|
intraSumA <= intraSumA+zeroExtend(topValSet[7:0])+zeroExtend(topValSet[15:8])+zeroExtend(topValSet[23:16])+zeroExtend(topValSet[31:24])+8;
|
| 1939 |
|
|
end
|
| 1940 |
|
|
6:
|
| 1941 |
|
|
begin
|
| 1942 |
|
|
if(leftAvailable == 1)
|
| 1943 |
|
|
begin
|
| 1944 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,0);
|
| 1945 |
|
|
intraSumA <= intraSumA+zeroExtend(leftValSet[15:8])+zeroExtend(leftValSet[23:16])+zeroExtend(leftValSet[31:24])+zeroExtend(leftValSet[39:32]);
|
| 1946 |
|
|
end
|
| 1947 |
|
|
else
|
| 1948 |
|
|
nextIntraStepCount = 10;
|
| 1949 |
|
|
end
|
| 1950 |
|
|
7:
|
| 1951 |
|
|
begin
|
| 1952 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,1);
|
| 1953 |
|
|
intraSumA <= intraSumA+zeroExtend(leftValSet[15:8])+zeroExtend(leftValSet[23:16])+zeroExtend(leftValSet[31:24])+zeroExtend(leftValSet[39:32]);
|
| 1954 |
|
|
end
|
| 1955 |
|
|
8:
|
| 1956 |
|
|
begin
|
| 1957 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,2);
|
| 1958 |
|
|
intraSumA <= intraSumA+zeroExtend(leftValSet[15:8])+zeroExtend(leftValSet[23:16])+zeroExtend(leftValSet[31:24])+zeroExtend(leftValSet[39:32]);
|
| 1959 |
|
|
end
|
| 1960 |
|
|
9:
|
| 1961 |
|
|
begin
|
| 1962 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,3);
|
| 1963 |
|
|
intraSumA <= intraSumA+zeroExtend(leftValSet[15:8])+zeroExtend(leftValSet[23:16])+zeroExtend(leftValSet[31:24])+zeroExtend(leftValSet[39:32])+8;
|
| 1964 |
|
|
end
|
| 1965 |
|
|
10:
|
| 1966 |
|
|
begin
|
| 1967 |
|
|
if(leftAvailable == 1 && topAvailable == 1)
|
| 1968 |
|
|
intraSumA <= intraSumA >> 5;
|
| 1969 |
|
|
else if(leftAvailable == 1 || topAvailable == 1)
|
| 1970 |
|
|
intraSumA <= intraSumA >> 4;
|
| 1971 |
|
|
else
|
| 1972 |
|
|
intraSumA <= 128;
|
| 1973 |
|
|
end
|
| 1974 |
|
|
11:
|
| 1975 |
|
|
begin
|
| 1976 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 1977 |
|
|
predVector[pixelHor] = intraSumA[7:0];
|
| 1978 |
|
|
outFlag = 1;
|
| 1979 |
|
|
end
|
| 1980 |
|
|
default: $display( "ERROR Prediction: intraProcessStep intra16x16 DC unknown intraStepCount");
|
| 1981 |
|
|
endcase
|
| 1982 |
|
|
end
|
| 1983 |
|
|
3://plane
|
| 1984 |
|
|
begin
|
| 1985 |
|
|
if(intraStepCount == 2)
|
| 1986 |
|
|
begin
|
| 1987 |
|
|
Bit#(32) topValSet = select(intraTopVal,3);
|
| 1988 |
|
|
Bit#(8) topVal = select32to8(topValSet,3);
|
| 1989 |
|
|
Bit#(40) leftValSet = select(intraLeftVal,3);
|
| 1990 |
|
|
Bit#(8) leftVal = intra4x4SelectLeft(leftValSet,3);
|
| 1991 |
|
|
Bit#(13) tempVal = zeroExtend(topVal) + zeroExtend(leftVal);
|
| 1992 |
|
|
intraSumA <= tempVal << 4;
|
| 1993 |
|
|
intraSumB <= 0;
|
| 1994 |
|
|
intraSumC <= 0;
|
| 1995 |
|
|
end
|
| 1996 |
|
|
else if(intraStepCount < 11)
|
| 1997 |
|
|
begin
|
| 1998 |
|
|
Bit#(4) xyPlusOne = intraStepCount-2;
|
| 1999 |
|
|
Bit#(4) xyPlusEight = intraStepCount+5;
|
| 2000 |
|
|
Bit#(4) sixMinusXY = 9-intraStepCount;
|
| 2001 |
|
|
Bit#(32) topValSet1 = select(intraTopVal,xyPlusEight[3:2]);
|
| 2002 |
|
|
Bit#(8) topVal1 = select32to8(topValSet1,xyPlusEight[1:0]);
|
| 2003 |
|
|
Bit#(40) leftValSet1 = select(intraLeftVal,xyPlusEight[3:2]);
|
| 2004 |
|
|
Bit#(8) leftVal1 = intra4x4SelectLeft(leftValSet1,zeroExtend(xyPlusEight[1:0]));
|
| 2005 |
|
|
Bit#(32) topValSet2=0;
|
| 2006 |
|
|
Bit#(8) topVal2;
|
| 2007 |
|
|
Bit#(40) leftValSet2;
|
| 2008 |
|
|
Bit#(8) leftVal2;
|
| 2009 |
|
|
if(intraStepCount==10)
|
| 2010 |
|
|
begin
|
| 2011 |
|
|
leftValSet2 = select(intraLeftVal,0);
|
| 2012 |
|
|
leftVal2 = intra4x4SelectLeft(leftValSet2,-1);
|
| 2013 |
|
|
topVal2 = leftVal2;
|
| 2014 |
|
|
end
|
| 2015 |
|
|
else
|
| 2016 |
|
|
begin
|
| 2017 |
|
|
topValSet2 = select(intraTopVal,sixMinusXY[3:2]);
|
| 2018 |
|
|
topVal2 = select32to8(topValSet2,sixMinusXY[1:0]);
|
| 2019 |
|
|
leftValSet2 = select(intraLeftVal,sixMinusXY[3:2]);
|
| 2020 |
|
|
leftVal2 = intra4x4SelectLeft(leftValSet2,zeroExtend(sixMinusXY[1:0]));
|
| 2021 |
|
|
end
|
| 2022 |
|
|
Bit#(15) diffH = zeroExtend(topVal1) - zeroExtend(topVal2);
|
| 2023 |
|
|
Bit#(15) diffV = zeroExtend(leftVal1) - zeroExtend(leftVal2);
|
| 2024 |
|
|
intraSumB <= intraSumB + (zeroExtend(xyPlusOne) * diffH);
|
| 2025 |
|
|
intraSumC <= intraSumC + (zeroExtend(xyPlusOne) * diffV);
|
| 2026 |
|
|
end
|
| 2027 |
|
|
else if(intraStepCount == 11)
|
| 2028 |
|
|
begin
|
| 2029 |
|
|
Bit#(18) tempSumB = (5*signExtend(intraSumB)) + 32;
|
| 2030 |
|
|
Bit#(18) tempSumC = (5*signExtend(intraSumC)) + 32;
|
| 2031 |
|
|
intraSumB <= signExtend(tempSumB[17:6]);
|
| 2032 |
|
|
intraSumC <= signExtend(tempSumC[17:6]);
|
| 2033 |
|
|
end
|
| 2034 |
|
|
else if(intraStepCount == 12)
|
| 2035 |
|
|
begin
|
| 2036 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 2037 |
|
|
begin
|
| 2038 |
|
|
Bit#(5) positionHor = {1'b0,blockHor,fromInteger(pixelHor)};
|
| 2039 |
|
|
Bit#(5) positionVer = {1'b0,blockVer,pixelVer};
|
| 2040 |
|
|
Bit#(16) tempProductB = signExtend(intraSumB) * signExtend(positionHor-7);
|
| 2041 |
|
|
Bit#(16) tempProductC = signExtend(intraSumC) * signExtend(positionVer-7);
|
| 2042 |
|
|
Bit#(16) tempTotal = tempProductB + tempProductC + zeroExtend(intraSumA) + 16;
|
| 2043 |
|
|
if(tempTotal[15]==1)
|
| 2044 |
|
|
predVector[pixelHor] = 0;
|
| 2045 |
|
|
else if(tempTotal[14:5] > 255)
|
| 2046 |
|
|
predVector[pixelHor] = 255;
|
| 2047 |
|
|
else
|
| 2048 |
|
|
predVector[pixelHor] = tempTotal[12:5];
|
| 2049 |
|
|
end
|
| 2050 |
|
|
outFlag = 1;
|
| 2051 |
|
|
end
|
| 2052 |
|
|
else
|
| 2053 |
|
|
$display( "ERROR Prediction: intraProcessStep intra16x16 plane unknown intraStepCount");
|
| 2054 |
|
|
end
|
| 2055 |
|
|
endcase
|
| 2056 |
|
|
end
|
| 2057 |
|
|
else if(intraChromaFlag==1)
|
| 2058 |
|
|
begin
|
| 2059 |
|
|
//$display( "TRACE Prediction: intraProcessStep intraChroma %0d %0d %0d %0d %0d %0d %h %h %h %h %h %h %h %h",intra_chroma_pred_mode.first(),intraChromaTopAvailable,intraChromaLeftAvailable,currMb,blockNum,pixelNum,pack(intraLeftValChroma0),pack(intraTopValChroma0),pack(intraLeftValChroma1),pack(intraTopValChroma1),intraLeftValChroma0[0],intraTopValChroma0[3][15:8],intraLeftValChroma1[0],intraTopValChroma1[3][15:8]);///////////////////
|
| 2060 |
|
|
Vector#(9,Bit#(8)) tempLeftVec;
|
| 2061 |
|
|
Vector#(4,Bit#(16)) tempTopVec;
|
| 2062 |
|
|
if(blockNum[2] == 0)
|
| 2063 |
|
|
begin
|
| 2064 |
|
|
tempLeftVec = intraLeftValChroma0;
|
| 2065 |
|
|
tempTopVec = intraTopValChroma0;
|
| 2066 |
|
|
end
|
| 2067 |
|
|
else
|
| 2068 |
|
|
begin
|
| 2069 |
|
|
tempLeftVec = intraLeftValChroma1;
|
| 2070 |
|
|
tempTopVec = intraTopValChroma1;
|
| 2071 |
|
|
end
|
| 2072 |
|
|
case(intra_chroma_pred_mode.first())
|
| 2073 |
|
|
0://dc
|
| 2074 |
|
|
begin
|
| 2075 |
|
|
if(intraStepCount == 2)
|
| 2076 |
|
|
begin
|
| 2077 |
|
|
Bit#(1) useTop=0;
|
| 2078 |
|
|
Bit#(1) useLeft=0;
|
| 2079 |
|
|
if(blockNum[1:0] == 0 || blockNum[1:0] == 3)
|
| 2080 |
|
|
begin
|
| 2081 |
|
|
useTop = intraChromaTopAvailable;
|
| 2082 |
|
|
useLeft = intraChromaLeftAvailable;
|
| 2083 |
|
|
end
|
| 2084 |
|
|
else if(blockNum[1:0] == 1)
|
| 2085 |
|
|
begin
|
| 2086 |
|
|
if(intraChromaTopAvailable == 1)
|
| 2087 |
|
|
useTop = 1;
|
| 2088 |
|
|
else if(intraChromaLeftAvailable == 1)
|
| 2089 |
|
|
useLeft = 1;
|
| 2090 |
|
|
end
|
| 2091 |
|
|
else if(blockNum[1:0] == 2)
|
| 2092 |
|
|
begin
|
| 2093 |
|
|
if(intraChromaLeftAvailable == 1)
|
| 2094 |
|
|
useLeft = 1;
|
| 2095 |
|
|
else if(intraChromaTopAvailable == 1)
|
| 2096 |
|
|
useTop = 1;
|
| 2097 |
|
|
end
|
| 2098 |
|
|
else
|
| 2099 |
|
|
$display( "ERROR Prediction: intraProcessStep intraChroma dc unknown blockNum");
|
| 2100 |
|
|
Bit#(10) topSum;
|
| 2101 |
|
|
Bit#(10) leftSum;
|
| 2102 |
|
|
Bit#(11) totalSum;
|
| 2103 |
|
|
if(blockHor[0] == 0)
|
| 2104 |
|
|
topSum = zeroExtend(tempTopVec[0][15:8])+zeroExtend(tempTopVec[0][7:0])+zeroExtend(tempTopVec[1][15:8])+zeroExtend(tempTopVec[1][7:0])+2;
|
| 2105 |
|
|
else
|
| 2106 |
|
|
topSum = zeroExtend(tempTopVec[2][15:8])+zeroExtend(tempTopVec[2][7:0])+zeroExtend(tempTopVec[3][15:8])+zeroExtend(tempTopVec[3][7:0])+2;
|
| 2107 |
|
|
if(blockVer[0] == 0)
|
| 2108 |
|
|
leftSum = zeroExtend(tempLeftVec[1])+zeroExtend(tempLeftVec[2])+zeroExtend(tempLeftVec[3])+zeroExtend(tempLeftVec[4])+2;
|
| 2109 |
|
|
else
|
| 2110 |
|
|
leftSum = zeroExtend(tempLeftVec[5])+zeroExtend(tempLeftVec[6])+zeroExtend(tempLeftVec[7])+zeroExtend(tempLeftVec[8])+2;
|
| 2111 |
|
|
totalSum = zeroExtend(topSum) + zeroExtend(leftSum);
|
| 2112 |
|
|
if(useTop==1 && useLeft==1)
|
| 2113 |
|
|
intraSumA <= zeroExtend(totalSum[10:3]);
|
| 2114 |
|
|
else if(useTop==1)
|
| 2115 |
|
|
intraSumA <= zeroExtend(topSum[9:2]);
|
| 2116 |
|
|
else if(useLeft==1)
|
| 2117 |
|
|
intraSumA <= zeroExtend(leftSum[9:2]);
|
| 2118 |
|
|
else
|
| 2119 |
|
|
intraSumA <= zeroExtend(8'b10000000);
|
| 2120 |
|
|
end
|
| 2121 |
|
|
else if(intraStepCount == 3)
|
| 2122 |
|
|
begin
|
| 2123 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 2124 |
|
|
predVector[pixelHor] = intraSumA[7:0];
|
| 2125 |
|
|
outFlag = 1;
|
| 2126 |
|
|
end
|
| 2127 |
|
|
else
|
| 2128 |
|
|
$display( "ERROR Prediction: intraProcessStep intraChroma dc unknown intraStepCount");
|
| 2129 |
|
|
end
|
| 2130 |
|
|
1://horizontal
|
| 2131 |
|
|
begin
|
| 2132 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 2133 |
|
|
begin
|
| 2134 |
|
|
Bit#(4) tempLeftIdx = {1'b0,blockVer[0],pixelVer} + 1;
|
| 2135 |
|
|
predVector[pixelHor] = select(tempLeftVec,tempLeftIdx);
|
| 2136 |
|
|
end
|
| 2137 |
|
|
outFlag = 1;
|
| 2138 |
|
|
end
|
| 2139 |
|
|
2://vertical
|
| 2140 |
|
|
begin
|
| 2141 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 2142 |
|
|
begin
|
| 2143 |
|
|
Bit#(2) pixelHorTemp = fromInteger(pixelHor);
|
| 2144 |
|
|
Bit#(16) tempTopVal = select(tempTopVec,{blockHor[0],pixelHorTemp[1]});
|
| 2145 |
|
|
if(pixelHorTemp[0] == 0)
|
| 2146 |
|
|
predVector[pixelHor] = tempTopVal[7:0];
|
| 2147 |
|
|
else
|
| 2148 |
|
|
predVector[pixelHor] = tempTopVal[15:8];
|
| 2149 |
|
|
end
|
| 2150 |
|
|
outFlag = 1;
|
| 2151 |
|
|
end
|
| 2152 |
|
|
3://plane
|
| 2153 |
|
|
begin
|
| 2154 |
|
|
if(intraStepCount == 2)
|
| 2155 |
|
|
begin
|
| 2156 |
|
|
Bit#(16) topValSet = tempTopVec[3];
|
| 2157 |
|
|
Bit#(8) topVal = topValSet[15:8];
|
| 2158 |
|
|
Bit#(8) leftVal = tempLeftVec[8];
|
| 2159 |
|
|
Bit#(13) tempVal = zeroExtend(topVal) + zeroExtend(leftVal);
|
| 2160 |
|
|
intraSumA <= tempVal << 4;
|
| 2161 |
|
|
intraSumB <= 0;
|
| 2162 |
|
|
intraSumC <= 0;
|
| 2163 |
|
|
end
|
| 2164 |
|
|
else if(intraStepCount < 7)
|
| 2165 |
|
|
begin
|
| 2166 |
|
|
Bit#(3) xyPlusOne = truncate(intraStepCount)-2;
|
| 2167 |
|
|
Bit#(3) xyPlusFour = truncate(intraStepCount)+1;
|
| 2168 |
|
|
Bit#(4) twoMinusXY = 5-intraStepCount;
|
| 2169 |
|
|
Bit#(16) topValSet1 = select(tempTopVec,xyPlusFour[2:1]);
|
| 2170 |
|
|
Bit#(8) topVal1 = select16to8(topValSet1,xyPlusFour[0]);
|
| 2171 |
|
|
Bit#(4) tempLeftIdx1 = {1'b0,xyPlusFour} + 1;
|
| 2172 |
|
|
Bit#(8) leftVal1 = select(tempLeftVec,tempLeftIdx1);
|
| 2173 |
|
|
|
| 2174 |
|
|
Bit#(16) topValSet2 = select(tempTopVec,twoMinusXY[2:1]);
|
| 2175 |
|
|
Bit#(8) topVal2;
|
| 2176 |
|
|
Bit#(8) leftVal2 = select(tempLeftVec,twoMinusXY+1);
|
| 2177 |
|
|
if(intraStepCount==6)
|
| 2178 |
|
|
topVal2 = leftVal2;
|
| 2179 |
|
|
else
|
| 2180 |
|
|
topVal2 = select16to8(topValSet2,twoMinusXY[0]);
|
| 2181 |
|
|
Bit#(15) diffH = zeroExtend(topVal1) - zeroExtend(topVal2);
|
| 2182 |
|
|
Bit#(15) diffV = zeroExtend(leftVal1) - zeroExtend(leftVal2);
|
| 2183 |
|
|
intraSumB <= intraSumB + (zeroExtend(xyPlusOne) * diffH);
|
| 2184 |
|
|
intraSumC <= intraSumC + (zeroExtend(xyPlusOne) * diffV);
|
| 2185 |
|
|
Int#(15) tempDisplayH = unpack(zeroExtend(xyPlusOne) * diffH);
|
| 2186 |
|
|
Int#(15) tempDisplayV = unpack(zeroExtend(xyPlusOne) * diffV);
|
| 2187 |
|
|
//$display( "TRACE Prediction: intraProcessStep intraChroma plane partH partV %0d %0d",tempDisplayH,tempDisplayV);////////////////////
|
| 2188 |
|
|
end
|
| 2189 |
|
|
else if(intraStepCount == 7)
|
| 2190 |
|
|
begin
|
| 2191 |
|
|
Int#(15) tempDisplayH = unpack(intraSumB);
|
| 2192 |
|
|
Int#(15) tempDisplayV = unpack(intraSumC);
|
| 2193 |
|
|
//$display( "TRACE Prediction: intraProcessStep intraChroma plane H V %0d %0d",tempDisplayH,tempDisplayV);////////////////////
|
| 2194 |
|
|
Bit#(19) tempSumB = (34*signExtend(intraSumB)) + 32;
|
| 2195 |
|
|
Bit#(19) tempSumC = (34*signExtend(intraSumC)) + 32;
|
| 2196 |
|
|
intraSumB <= signExtend(tempSumB[18:6]);
|
| 2197 |
|
|
intraSumC <= signExtend(tempSumC[18:6]);
|
| 2198 |
|
|
end
|
| 2199 |
|
|
else if(intraStepCount == 8)
|
| 2200 |
|
|
begin
|
| 2201 |
|
|
for(Integer pixelHor=0; pixelHor<4; pixelHor=pixelHor+1)
|
| 2202 |
|
|
begin
|
| 2203 |
|
|
Bit#(4) positionHor = {1'b0,blockHor[0],fromInteger(pixelHor)};
|
| 2204 |
|
|
Bit#(4) positionVer = {1'b0,blockVer[0],pixelVer};
|
| 2205 |
|
|
Bit#(17) tempProductB = signExtend(intraSumB) * signExtend(positionHor-3);
|
| 2206 |
|
|
Bit#(17) tempProductC = signExtend(intraSumC) * signExtend(positionVer-3);
|
| 2207 |
|
|
Bit#(17) tempTotal = tempProductB + tempProductC + zeroExtend(intraSumA) + 16;
|
| 2208 |
|
|
if(tempTotal[16]==1)
|
| 2209 |
|
|
predVector[pixelHor] = 0;
|
| 2210 |
|
|
else if(tempTotal[15:5] > 255)
|
| 2211 |
|
|
predVector[pixelHor] = 255;
|
| 2212 |
|
|
else
|
| 2213 |
|
|
predVector[pixelHor] = tempTotal[12:5];
|
| 2214 |
|
|
end
|
| 2215 |
|
|
outFlag = 1;
|
| 2216 |
|
|
end
|
| 2217 |
|
|
else
|
| 2218 |
|
|
$display( "ERROR Prediction: intraProcessStep intraChroma plane unknown intraStepCount");
|
| 2219 |
|
|
end
|
| 2220 |
|
|
endcase
|
| 2221 |
|
|
end
|
| 2222 |
|
|
else
|
| 2223 |
|
|
$display( "ERROR Prediction: intraProcessStep unknown intrastate");
|
| 2224 |
|
|
|
| 2225 |
|
|
if(outFlag==1)
|
| 2226 |
|
|
begin
|
| 2227 |
|
|
predictedfifo.enq(predVector);
|
| 2228 |
|
|
pixelNum <= pixelNum+4;
|
| 2229 |
|
|
if(pixelNum == 12)
|
| 2230 |
|
|
begin
|
| 2231 |
|
|
if(intraChromaFlag==0)
|
| 2232 |
|
|
begin
|
| 2233 |
|
|
blockNum <= blockNum+1;
|
| 2234 |
|
|
if(blockNum == 15)
|
| 2235 |
|
|
begin
|
| 2236 |
|
|
intraChromaFlag <= 1;
|
| 2237 |
|
|
intraStepCount <= 2;
|
| 2238 |
|
|
end
|
| 2239 |
|
|
else if(intrastate==Intra4x4)
|
| 2240 |
|
|
intraStepCount <= 1;
|
| 2241 |
|
|
end
|
| 2242 |
|
|
else
|
| 2243 |
|
|
begin
|
| 2244 |
|
|
if(blockNum == 7)
|
| 2245 |
|
|
begin
|
| 2246 |
|
|
blockNum <= 0;
|
| 2247 |
|
|
intraChromaFlag <= 0;
|
| 2248 |
|
|
intraStepCount <= 0;
|
| 2249 |
|
|
intra_chroma_pred_mode.deq();
|
| 2250 |
|
|
end
|
| 2251 |
|
|
else
|
| 2252 |
|
|
begin
|
| 2253 |
|
|
blockNum <= blockNum+1;
|
| 2254 |
|
|
if(intra_chroma_pred_mode.first()==0)
|
| 2255 |
|
|
intraStepCount <= 2;
|
| 2256 |
|
|
else if(blockNum==3)
|
| 2257 |
|
|
intraStepCount <= 2;
|
| 2258 |
|
|
end
|
| 2259 |
|
|
end
|
| 2260 |
|
|
end
|
| 2261 |
|
|
end
|
| 2262 |
|
|
else
|
| 2263 |
|
|
intraStepCount <= nextIntraStepCount;
|
| 2264 |
|
|
//$display( "Trace Prediction: intraProcessStep");
|
| 2265 |
|
|
endrule
|
| 2266 |
|
|
|
| 2267 |
|
|
|
| 2268 |
|
|
|
| 2269 |
|
|
interface Client mem_client_intra;
|
| 2270 |
|
|
interface Get request = fifoToGet(intraMemReqQ);
|
| 2271 |
|
|
interface Put response = fifoToPut(intraMemRespQ);
|
| 2272 |
|
|
endinterface
|
| 2273 |
|
|
interface Client mem_client_inter;
|
| 2274 |
|
|
interface Get request = fifoToGet(interMemReqQ);
|
| 2275 |
|
|
interface Put response = fifoToPut(interMemRespQ);
|
| 2276 |
|
|
endinterface
|
| 2277 |
|
|
interface Client mem_client_buffer = interpolator.mem_client;
|
| 2278 |
|
|
|
| 2279 |
|
|
interface Put ioin = fifoToPut(infifo);
|
| 2280 |
|
|
interface Put ioin_InverseTrans = fifoToPut(infifo_ITB);
|
| 2281 |
|
|
interface Get ioout = fifoToGet(outfifo);
|
| 2282 |
|
|
|
| 2283 |
|
|
|
| 2284 |
|
|
endmodule
|
| 2285 |
|
|
|
| 2286 |
|
|
endpackage
|
