Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 3769 | Serge | 1 | /* |
| 2 | * Copyright 2000-2011 Intel Corporation All Rights Reserved |
||
| 3 | * |
||
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
||
| 5 | * you may not use this file except in compliance with the License. |
||
| 6 | * You may obtain a copy of the License at |
||
| 7 | * |
||
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
||
| 9 | * |
||
| 10 | * Unless required by applicable law or agreed to in writing, software |
||
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
||
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
||
| 13 | * See the License for the specific language governing permissions and |
||
| 14 | * limitations under the License. |
||
| 15 | */ |
||
| 16 | // 33 // Total instruction count |
||
| 17 | // 1 // Total kernel count |
||
| 18 | |||
| 19 | |||
| 20 | // Module name: common.inc |
||
| 21 | // |
||
| 22 | // Common header file for all Video-Processing kernels |
||
| 23 | // |
||
| 24 | |||
| 25 | .default_execution_size (16) |
||
| 26 | .default_register_type :ub |
||
| 27 | |||
| 28 | .reg_count_total 128 |
||
| 29 | .reg_count_payload 7 |
||
| 30 | |||
| 31 | //========== Common constants ========== |
||
| 32 | |||
| 33 | |||
| 34 | //========== Macros ========== |
||
| 35 | |||
| 36 | |||
| 37 | //Fast Jump, For more details see "Set_Layer_N.asm" |
||
| 38 | |||
| 39 | |||
| 40 | //========== Defines ==================== |
||
| 41 | |||
| 42 | //========== Static Parameters (Common To All) ========== |
||
| 43 | //r1 |
||
| 44 | |||
| 45 | |||
| 46 | //r2 |
||
| 47 | |||
| 48 | // e.g. byte0 byte1 byte2 |
||
| 49 | // YUYV 0 1 3 |
||
| 50 | // YVYU 0 3 1 |
||
| 51 | |||
| 52 | //Color Pipe (IECP) parameters |
||
| 53 | |||
| 54 | |||
| 55 | //r4 |
||
| 56 | |||
| 57 | // e.g. byte0 byte1 byte2 |
||
| 58 | // YUYV 0 1 3 |
||
| 59 | // YVYU 0 3 1 |
||
| 60 | |||
| 61 | |||
| 62 | //========== Inline parameters (Common To All) =========== |
||
| 63 | |||
| 64 | |||
| 65 | //============== Binding Index Table=========== |
||
| 66 | //Common between DNDI and DNUV |
||
| 67 | |||
| 68 | |||
| 69 | //================= Common Message Descriptor ===== |
||
| 70 | // Message descriptor for thread spawning |
||
| 71 | // Message Descriptors |
||
| 72 | // = 000 0001 (min message len 1 ) 0,0000 (resp len 0 -add later) |
||
| 73 | // 0000,0000,0000 |
||
| 74 | // 0001(Spawn a root thread),0001 (Root thread spawn thread) |
||
| 75 | // = 0x02000011 |
||
| 76 | // Thread Spawner Message Descriptor |
||
| 77 | |||
| 78 | |||
| 79 | // Message descriptor for atomic operation add |
||
| 80 | // Message Descriptors |
||
| 81 | // = 000 0110 (min message len 6 ) 0,0000 (resp len 0 -add later) |
||
| 82 | // 1(header present)001,10(typed atomic operation)0(return enabled)0(slot group, low 8 bits),0111 (AOP_Add) |
||
| 83 | // 0000,0000 (Binding table index, added later) |
||
| 84 | // = 0x02000011 |
||
| 85 | |||
| 86 | // Atomic Operation Add Message Descriptor |
||
| 87 | |||
| 88 | |||
| 89 | // Message descriptor for dataport media write |
||
| 90 | // Message Descriptors |
||
| 91 | // = 000 0001 (min message len 1 - add later) 00000 (resp len 0) |
||
| 92 | // 1 (header present 1) 0 1010 (media block write) 000000 |
||
| 93 | // 00000000 (binding table index - set later) |
||
| 94 | // = 0x020A8000 |
||
| 95 | |||
| 96 | |||
| 97 | // Message Length defines |
||
| 98 | |||
| 99 | |||
| 100 | // Response Length defines |
||
| 101 | |||
| 102 | |||
| 103 | // Block Width and Height Size defines |
||
| 104 | |||
| 105 | |||
| 106 | // Extended Message Descriptors |
||
| 107 | |||
| 108 | |||
| 109 | // Common message descriptors: |
||
| 110 | |||
| 111 | |||
| 112 | //===================== Math Function Control =================================== |
||
| 113 | |||
| 114 | |||
| 115 | //============ Message Registers =============== |
||
| 116 | // buf4 starts from r28 |
||
| 117 | |||
| 118 | |||
| 119 | //#define mMSGHDR_EOT r43 // Dummy Message Register for EOT |
||
| 120 | |||
| 121 | |||
| 122 | .declare mubMSGPAYLOAD Base=r30 ElementSize=1 SrcRegion=<16;16,1> Type=ub |
||
| 123 | .declare muwMSGPAYLOAD Base=r30 ElementSize=2 SrcRegion=<16;16,1> Type=uw |
||
| 124 | .declare mudMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=ud |
||
| 125 | .declare mfMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 126 | |||
| 127 | //=================== End of thread instruction =========================== |
||
| 128 | |||
| 129 | |||
| 130 | //=====================Pointers Used===================================== |
||
| 131 | |||
| 132 | |||
| 133 | //======================================================================= |
||
| 134 | |||
| 135 | |||
| 136 | //r9-r17 |
||
| 137 | // Define temp space for any usages |
||
| 138 | |||
| 139 | |||
| 140 | // Common Buffers |
||
| 141 | |||
| 142 | |||
| 143 | // End of common.inc |
||
| 144 | |||
| 145 | // FileName: DI_Save_PA_16x4.asm |
||
| 146 | // Author: Vivek Kumar |
||
| 147 | // Description: Save two 16x4 blocks of DI output in Packed format |
||
| 148 | |||
| 149 | |||
| 150 | // FileName: DNDI.inc |
||
| 151 | // Author: Vivek Kumar |
||
| 152 | // Description: Include file for DN, DI and DNDI |
||
| 153 | // Inputs: DI_ENABLE, DN_ENABLE, DN_PLANAR, DN_PACKED |
||
| 154 | |||
| 155 | |||
| 156 | // End of common.inc |
||
| 157 | |||
| 158 | |||
| 159 | //Interface: |
||
| 160 | //Static Parameters: |
||
| 161 | //r1 |
||
| 162 | |||
| 163 | |||
| 164 | //====================== Binding table (Explicit To DNDI)========================================= |
||
| 165 | |||
| 166 | |||
| 167 | .declare mudMSGHDR_DNDI Base=r18 ElementSize=4 Type=ud |
||
| 168 | .declare mdMSGHDR_DNDI Base=r18 ElementSize=4 Type=d |
||
| 169 | .declare mwMSGHDR_DNDI Base=r18 ElementSize=2 Type=w |
||
| 170 | |||
| 171 | |||
| 172 | .declare mudMSGHDR_STMM Base=r20 ElementSize=4 Type=ud |
||
| 173 | |||
| 174 | |||
| 175 | .declare mudMSGHDR_HIST Base=r22 ElementSize=4 Type=ud |
||
| 176 | |||
| 177 | |||
| 178 | .declare mudMSGHDR_ENC_STATS Base=r24 ElementSize=4 Type=ud |
||
| 179 | .declare muwMSGHDR_ENC_STATS Base=r24 ElementSize=2 Type=uw |
||
| 180 | .declare mubMSGHDR_ENC_STATS Base=r24 ElementSize=1 Type=ub |
||
| 181 | |||
| 182 | |||
| 183 | .declare mudMSGHDR_DN_OUT Base=r31.0 ElementSize=4 Type=ud |
||
| 184 | .declare mdMSGHDR_DN_OUT Base=r31.0 ElementSize=4 Type=d |
||
| 185 | .declare mubMSGHDR_DN_OUT Base=r31.0 ElementSize=1 Type=ub |
||
| 186 | |||
| 187 | |||
| 188 | .declare mudMSGHDR_UVCOPY Base=r36 ElementSize=4 Type=ud |
||
| 189 | .declare mdMSGHDR_UVCOPY Base=r36 ElementSize=4 Type=d |
||
| 190 | .declare mudMSGHDR_UCOPY Base=r36 ElementSize=4 Type=ud |
||
| 191 | .declare mudMSGHDR_VCOPY Base=r38 ElementSize=4 Type=ud |
||
| 192 | |||
| 193 | |||
| 194 | .declare mudMSGHDR_DI_OUT1 Base=r18.0 ElementSize=4 Type=ud |
||
| 195 | .declare mubMSGHDR_DI_OUT1 Base=r18.0 ElementSize=1 Type=ub |
||
| 196 | |||
| 197 | |||
| 198 | .declare mudMSGHDR_DI_OUT2 Base=r23.0 ElementSize=4 Type=ud |
||
| 199 | .declare mubMSGHDR_DI_OUT2 Base=r23.0 ElementSize=1 Type=ub |
||
| 200 | |||
| 201 | //r45 |
||
| 202 | //Use r45 as message header, so no need to "mov" the data. |
||
| 203 | |||
| 204 | .declare mudDN_Y_OUT Base=r45.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 205 | |||
| 206 | // Message response (Denoised & DI-ed pixels & statistics); Use buffer 5 |
||
| 207 | .declare udDNDI_RESP Base=r46.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 208 | .declare uwDNDI_RESP Base=r46.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw |
||
| 209 | .declare ubDNDI_RESP Base=r46.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 210 | |||
| 211 | // Message response (UV Copy); Use buffer 5 |
||
| 212 | .declare udDNDI_UV_RESP Base=r58.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 213 | .declare ubDNDI_UV_RESP Base=r58.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 214 | |||
| 215 | //Temp GRFs: For 42X to 422 Conversion |
||
| 216 | .declare uwDNDI_UVCOPY_TEMP Base=r10.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw //8 GRFs |
||
| 217 | .declare ubDNDI_UVCOPY_TEMP Base=r10.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub //8 GRFs |
||
| 218 | //--------------------------------------------------------------------------- |
||
| 219 | // Message descriptors |
||
| 220 | //--------------------------------------------------------------------------- |
||
| 221 | // Extended message descriptor |
||
| 222 | // Message descriptor for sampler read |
||
| 223 | // = 000 0010 (message len 2) 00000 (resp len - set later, 12 or 5 or 11) |
||
| 224 | // 1 (header present 1) 0 11 (SIMD32/64 mode) |
||
| 225 | // 1000 (message type) 0000 (DI state index) |
||
| 226 | // 00000000 (binding table index - set later) |
||
| 227 | // = 0x040b8000 |
||
| 228 | |||
| 229 | |||
| 230 | // Attention: The Message Length is The Number of GRFs with Data Only, without the Header |
||
| 231 | |||
| 232 | |||
| 233 | //--------------------------------------------------------------------------- |
||
| 234 | // VDI Return Data format |
||
| 235 | //--------------------------------------------------------------------------- |
||
| 236 | // Defines for DI enabled |
||
| 237 | |||
| 238 | |||
| 239 | // Defines for DI disabled |
||
| 240 | |||
| 241 | |||
| 242 | add (4) a0.4<1>:uw r2.28<4;4,1>:ub 608:w // Initial Y,U,V offset in YUV422 block; it starts at m20 |
||
| 243 | |||
| 244 | mov (8) r27.0<1>:ud r0.0<8;8,1>:ud |
||
| 245 | shl (1) r27.0<1>:d r7.0<0;1,0>:w 1:w { NoDDClr } // H. block origin need to be doubled |
||
| 246 | mov (1) r27.1<1>:d r7.1<0;1,0>:w { NoDDClr, NoDDChk } // Block origin |
||
| 247 | mov (1) r27.2<1>:ud 0x3001F:ud { NoDDChk } // Block width and height (32x8) |
||
| 248 | |||
| 249 | //prepare the message headers |
||
| 250 | mov (8) r18.0<1>:ud r27<8;8,1>:ud |
||
| 251 | mov (8) r23.0<1>:ud r27<8;8,1>:ud |
||
| 252 | |||
| 253 | // Pack 2nd field Y |
||
| 254 | mov (16) r[a0.4, 0]<2> ubDNDI_RESP(0,0) { NoDDClr } |
||
| 255 | mov (16) r[a0.4, 32]<2> ubDNDI_RESP(0,16) { NoDDClr } |
||
| 256 | mov (16) r[a0.4, 64]<2> ubDNDI_RESP(0,32) { NoDDClr } |
||
| 257 | mov (16) r[a0.4, 96]<2> ubDNDI_RESP(0,48) { NoDDClr } |
||
| 258 | // Pack 2nd field U |
||
| 259 | mov (8) r[a0.5, 0]<4> ubDNDI_RESP(2,1)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 260 | mov (8) r[a0.5, 32]<4> ubDNDI_RESP(2,17)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 261 | mov (8) r[a0.5, 64]<4> ubDNDI_RESP(2,33)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 262 | mov (8) r[a0.5, 96]<4> ubDNDI_RESP(2,49)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 263 | // Pack 2nd field V |
||
| 264 | mov (8) r[a0.6, 0]<4> ubDNDI_RESP(2,0)<16;8,2> { NoDDChk } //Vpixels |
||
| 265 | mov (8) r[a0.6, 32]<4> ubDNDI_RESP(2,16)<16;8,2> { NoDDChk } //Vpixels |
||
| 266 | mov (8) r[a0.6, 64]<4> ubDNDI_RESP(2,32)<16;8,2> { NoDDChk } //Vpixels |
||
| 267 | mov (8) r[a0.6, 96]<4> ubDNDI_RESP(2,48)<16;8,2> { NoDDChk } //Vpixels |
||
| 268 | |||
| 269 | // Pack 1st field Y |
||
| 270 | mov (16) r[a0.4, 160]<2> ubDNDI_RESP(4,0) { NoDDClr } |
||
| 271 | mov (16) r[a0.4, 192]<2> ubDNDI_RESP(4,16) { NoDDClr } |
||
| 272 | mov (16) r[a0.4, 224]<2> ubDNDI_RESP(4,32) { NoDDClr } |
||
| 273 | mov (16) r[a0.4, 256]<2> ubDNDI_RESP(4,48) { NoDDClr } |
||
| 274 | // Pack 1st field U |
||
| 275 | mov (8) r[a0.5, 160]<4> ubDNDI_RESP(6,1)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 276 | mov (8) r[a0.5, 192]<4> ubDNDI_RESP(6,17)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 277 | mov (8) r[a0.5, 224]<4> ubDNDI_RESP(6,33)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 278 | mov (8) r[a0.5, 256]<4> ubDNDI_RESP(6,49)<16;8,2> { NoDDClr, NoDDChk } //U pixels |
||
| 279 | // Pack 1st field V |
||
| 280 | mov (8) r[a0.6, 160]<4> ubDNDI_RESP(6,0)<16;8,2> { NoDDChk } //Vpixels |
||
| 281 | mov (8) r[a0.6, 192]<4> ubDNDI_RESP(6,16)<16;8,2> { NoDDChk } //Vpixels |
||
| 282 | mov (8) r[a0.6, 224]<4> ubDNDI_RESP(6,32)<16;8,2> { NoDDChk } //Vpixels |
||
| 283 | mov (8) r[a0.6, 256]<4> ubDNDI_RESP(6,48)<16;8,2> { NoDDChk } //Vpixels |
||
| 284 | |||
| 285 | //save the previous frame |
||
| 286 | send (8) null<1>:d r18.0 0x5 0xA0A801B:ud |
||
| 287 | |||
| 288 | //save the current frame |
||
| 289 | send (8) null<1>:d r23.0 0x5 0xA0A801E:ud |