workbench/hidds/nouveau/xf86-video-nouveau/nv30_shaders.c

   1 /*
   2  * Copyright 2007 Nouveau Project
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #include "nv30_shaders.h"
  24 #include "nv04_pushbuf.h"
  25
  26 void NV30_UploadFragProg(NVPtr pNv, nv_shader_t *shader, int *hw_offset)
  27 {
  28         uint32_t data, i;
  29         uint32_t *map;
  30
  31         shader->hw_id = *hw_offset;
  32
  33         nouveau_bo_map(pNv->shader_mem, NOUVEAU_BO_WR);
  34         map = pNv->shader_mem->map + *hw_offset;
  35         for (i = 0; i < shader->size; i++) {
  36                 data = shader->data[i];
  37 #if (X_BYTE_ORDER != X_LITTLE_ENDIAN)
  38                 data = ((data >> 16) | ((data & 0xffff) << 16));
  39 #endif
  40                 map[i] = data;
  41         }
  42         nouveau_bo_unmap(pNv->shader_mem);
  43
  44         *hw_offset += (shader->size * sizeof(uint32_t));
  45         *hw_offset = (*hw_offset + 63) & ~63;
  46 }
  47
  48 void NV40_UploadVtxProg(NVPtr pNv, nv_shader_t *shader, int *hw_id)
  49 {
  50         struct nouveau_channel *chan = pNv->chan;
  51         struct nouveau_grobj *curie = pNv->Nv3D;
  52         int i;
  53
  54         shader->hw_id = *hw_id;
  55
  56         BEGIN_RING(chan, curie, NV40TCL_VP_UPLOAD_FROM_ID, 1);
  57         OUT_RING  (chan, (shader->hw_id));
  58         for (i=0; i<shader->size; i+=4) {
  59                 BEGIN_RING(chan, curie, NV40TCL_VP_UPLOAD_INST(0), 4);
  60                 OUT_RING  (chan, shader->data[i + 0]);
  61                 OUT_RING  (chan, shader->data[i + 1]);
  62                 OUT_RING  (chan, shader->data[i + 2]);
  63                 OUT_RING  (chan, shader->data[i + 3]);
  64                 (*hw_id)++;
  65         }
  66 }
  67
  68 Bool
  69 NV30_LoadFragProg(ScrnInfoPtr pScrn, nv_shader_t *shader)
  70 {
  71         NVPtr pNv = NVPTR(pScrn);
  72         struct nouveau_channel *chan = pNv->chan;
  73         struct nouveau_grobj *rankine = pNv->Nv3D;
  74
  75         BEGIN_RING(chan, rankine, NV34TCL_FP_ACTIVE_PROGRAM, 1);
  76         if (OUT_RELOC(chan, pNv->shader_mem, shader->hw_id, NOUVEAU_BO_VRAM |
  77                       NOUVEAU_BO_RD | NOUVEAU_BO_LOW | NOUVEAU_BO_OR,
  78                       NV34TCL_FP_ACTIVE_PROGRAM_DMA0,
  79                       NV34TCL_FP_ACTIVE_PROGRAM_DMA1))
  80                 return FALSE;
  81         BEGIN_RING(chan, rankine, NV34TCL_FP_REG_CONTROL, 1);
  82         OUT_RING  (chan, (1 << 16)| 0xf);
  83         BEGIN_RING(chan, rankine, NV34TCL_MULTISAMPLE_CONTROL, 1);
  84         OUT_RING  (chan, 0xffff0000);
  85
  86         BEGIN_RING(chan, rankine, NV34TCL_FP_CONTROL,1);
  87         OUT_RING  (chan, (shader->card_priv.NV30FP.num_regs-1)/2);
  88
  89         return TRUE;
  90 }
  91
  92 void
  93 NV40_LoadVtxProg(ScrnInfoPtr pScrn, nv_shader_t *shader)
  94 {
  95         NVPtr pNv = NVPTR(pScrn);
  96         struct nouveau_channel *chan = pNv->chan;
  97         struct nouveau_grobj *curie = pNv->Nv3D;
  98
  99         BEGIN_RING(chan, curie, NV40TCL_VP_START_FROM_ID, 1);
 100         OUT_RING  (chan, (shader->hw_id));
 101
 102         BEGIN_RING(chan, curie, NV40TCL_VP_ATTRIB_EN, 2);
 103         OUT_RING  (chan, shader->card_priv.NV30VP.vp_in_reg);
 104         OUT_RING  (chan, shader->card_priv.NV30VP.vp_out_reg);
 105 }
 106
 107 Bool
 108 NV40_LoadFragProg(ScrnInfoPtr pScrn, nv_shader_t *shader)
 109 {
 110         NVPtr pNv = NVPTR(pScrn);
 111         struct nouveau_channel *chan = pNv->chan;
 112         struct nouveau_grobj *curie = pNv->Nv3D;
 113
 114         BEGIN_RING(chan, curie, NV40TCL_FP_ADDRESS, 1);
 115         if (OUT_RELOC(chan, pNv->shader_mem, shader->hw_id, NOUVEAU_BO_VRAM |
 116                       NOUVEAU_BO_GART | NOUVEAU_BO_RD | NOUVEAU_BO_LOW |
 117                       NOUVEAU_BO_OR,
 118                       NV40TCL_FP_ADDRESS_DMA0, NV40TCL_FP_ADDRESS_DMA1))
 119                 return FALSE;
 120         BEGIN_RING(chan, curie, NV40TCL_FP_CONTROL, 1);
 121         OUT_RING  (chan, shader->card_priv.NV30FP.num_regs <<
 122                          NV40TCL_FP_CONTROL_TEMP_COUNT_SHIFT);
 123
 124         return TRUE;
 125 }
 126
 127 /*******************************************************************************
 128  * NV40/G70 vertex shaders
 129  */
 130
 131 nv_shader_t nv40_vp_exa_render = {
 132         .card_priv.NV30VP.vp_in_reg  = 0x00000309,
 133         .card_priv.NV30VP.vp_out_reg = 0x0000c001,
 134         .size = (3*4),
 135         .data = {
 136                 /* MOV result.position, vertex.position */
 137                 0x40041c6c, 0x0040000d, 0x8106c083, 0x6041ff80,
 138                 /* MOV result.texcoord[0], vertex.texcoord[0] */
 139                 0x401f9c6c, 0x0040080d, 0x8106c083, 0x6041ff9c,
 140                 /* MOV result.texcoord[1], vertex.texcoord[1] */
 141                 0x401f9c6c, 0x0040090d, 0x8106c083, 0x6041ffa1,
 142         }
 143 };
 144
 145 /*******************************************************************************
 146  * NV30/NV40/G70 fragment shaders
 147  */
 148
 149 nv_shader_t nv30_fp_pass_col0 = {
 150         .card_priv.NV30FP.num_regs = 2,
 151         .size = (1*4),
 152         .data = {
 153                 /* MOV R0, fragment.color */
 154                 0x01403e81, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 155         }
 156 };
 157
 158 nv_shader_t nv30_fp_pass_tex0 = {
 159         .card_priv.NV30FP.num_regs = 2,
 160         .size = (2*4),
 161         .data = {
 162                 /* TEX R0, fragment.texcoord[0], texture[0], 2D */
 163                 0x17009e00, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 164                 /* MOV R0, R0 */
 165                 0x01401e81, 0x1c9dc800, 0x0001c800, 0x0001c800,
 166         }
 167 };
 168
 169 nv_shader_t nv30_fp_composite_mask = {
 170         .card_priv.NV30FP.num_regs = 2,
 171         .size = (3*4),
 172         .data = {
 173                 /* TEXC0 R1.w         , fragment.texcoord[1], texture[1], 2D */
 174                 0x1702b102, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 175                 /* TEX   R0 (NE0.wwww), fragment.texcoord[0], texture[0], 2D */
 176                 0x17009e00, 0x1ff5c801, 0x0001c800, 0x3fe1c800,
 177                 /* MUL   R0           , R0, R1.w */
 178                 0x02001e81, 0x1c9dc800, 0x0001fe04, 0x0001c800,
 179         }
 180 };
 181
 182 nv_shader_t nv30_fp_composite_mask_sa_ca = {
 183         .card_priv.NV30FP.num_regs = 2,
 184         .size = (3*4),
 185         .data = {
 186                 /* TEXC0 R1.w         , fragment.texcoord[0], texture[0], 2D */
 187                 0x17009102, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 188                 /* TEX   R0 (NE0.wwww), fragment.texcoord[1], texture[1], 2D */
 189                 0x1702be00, 0x1ff5c801, 0x0001c800, 0x3fe1c800,
 190                 /* MUL   R0           , R1,wwww, R0 */
 191                 0x02001e81, 0x1c9dfe04, 0x0001c800, 0x0001c800,
 192         }
 193 };
 194
 195 nv_shader_t nv30_fp_composite_mask_ca = {
 196         .card_priv.NV30FP.num_regs = 2,
 197         .size = (3*4),
 198         .data = {
 199                 /* TEXC0 R0           , fragment.texcoord[0], texture[0], 2D */
 200                 0x17009f00, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 201                 /* TEX   R1 (NE0.xyzw), fragment.texcoord[1], texture[1], 2D */
 202                 0x1702be02, 0x1c95c801, 0x0001c800, 0x3fe1c800,
 203                 /* MUL   R0           , R0, R1 */
 204                 0x02001e81, 0x1c9dc800, 0x0001c804, 0x0001c800,
 205         }
 206 };
 207
 208 nv_shader_t nv40_vp_video = {
 209         .card_priv.NV30VP.vp_in_reg  = 0x00000309,
 210         .card_priv.NV30VP.vp_out_reg = 0x0000c001,
 211         .size = (3*4),
 212         .data = {
 213                 /* MOV result.position, vertex.position */
 214                 0x40041c6c, 0x0040000d, 0x8106c083, 0x6041ff80,
 215                 /* MOV result.texcoord[0], vertex.texcoord[0] */
 216                 0x401f9c6c, 0x0040080d, 0x8106c083, 0x6041ff9c,
 217                 /* MOV result.texcoord[1], vertex.texcoord[1] */
 218                 0x401f9c6c, 0x0040090d, 0x8106c083, 0x6041ffa1,
 219         }
 220 };
 221
 222 nv_shader_t nv40_fp_yv12_bicubic = {
 223         .card_priv.NV30FP.num_regs = 4,
 224         .size = (29*4),
 225         .data = {
 226                 /* INST 0: MOVR R0.xy (TR0.xyzw), attrib.texcoord[0] */
 227                 0x01008600, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 228                 /* INST 1: ADDR R0.z (TR0.xyzw), R0.yyyy, { 0.50, 0.00, 0.00, 0.00 }.xxxx */
 229                 0x03000800, 0x1c9caa00, 0x00000002, 0x0001c800,
 230                 0x3f000000, 0x00000000, 0x00000000, 0x00000000,
 231                 /* INST 2: ADDR R1.x (TR0.xyzw), R0, { 0.50, 0.00, 0.00, 0.00 }.xxxx */
 232                 0x03000202, 0x1c9dc800, 0x00000002, 0x0001c800,
 233                 0x3f000000, 0x00000000, 0x00000000, 0x00000000,
 234                 /* INST 3: TEXRC0 R1.xyz (TR0.xyzw), R0.zzzz, texture[0] */
 235                 0x17000f82, 0x1c9d5400, 0x0001c800, 0x0001c800,
 236                 /* INST 4: MULR R2.yw (TR0.xyzw), R1.xxyy, { -1.00, 1.00, 0.00, 0.00 }.xxyy */
 237                 0x02001404, 0x1c9ca104, 0x0000a002, 0x0001c800,
 238                 0xbf800000, 0x3f800000, 0x00000000, 0x00000000,
 239                 /* INST 5: TEXR R3.xyz (TR0.xyzw), R1, texture[0] */
 240                 0x17000e86, 0x1c9dc804, 0x0001c800, 0x0001c800,
 241                 /* INST 6: MULR R2.xz (TR0.xyzw), R3.xxyy, { -1.00, 1.00, 0.00, 0.00 }.xxyy */
 242                 0x02000a04, 0x1c9ca10c, 0x0000a002, 0x0001c800,
 243                 0xbf800000, 0x3f800000, 0x00000000, 0x00000000,
 244                 /* INST 7: ADDR R2 (TR0.xyzw), R0.xyxy, R2 */
 245                 0x03001e04, 0x1c9c8800, 0x0001c808, 0x0001c800,
 246                 /* INST 8: TEXR R1.y (TR0.xyzw), R2.zwzz, -texture[1] */
 247                 0x17020402, 0x1c9d5c08, 0x0001c800, 0x0001c800,
 248                 /* INST 9: MADH R1.x (TR0.xyzw), -R1.zzzz, R1.yyyy, R1.yyyy */
 249                 0x04400282, 0x1c9f5504, 0x0000aa04, 0x0000aa04,
 250                 /* INST 10: TEXR R0.y (TR0.xyzw), R2.xwxw, -texture[1] */
 251                 0x17020400, 0x1c9d9808, 0x0001c800, 0x0001c800,
 252                 /* INST 11: MADH R0.w (TR0.xyzw), -R1.zzzz, R0.yyyy, R0.yyyy */
 253                 0x04401080, 0x1c9f5504, 0x0000aa00, 0x0000aa00,
 254                 /* INST 12: TEXR R0.x (TR0.xyzw), R2.zyxy, texture[1] */
 255                 0x17020200, 0x1c9c8c08, 0x0001c800, 0x0001c800,
 256                 /* INST 13: MADH R1.x (TR0.xyzw), R1.zzzz, R0, R1 */
 257                 0x04400282, 0x1c9d5504, 0x0001c800, 0x0001c904,
 258                 /* INST 14: TEXR R0.x (NE0.zzzz), R2, texture[1] */
 259                 0x17020200, 0x1555c808, 0x0001c800, 0x0001c800,
 260                 /* INST 15: MADH R0.x (TR0.xyzw), R1.zzzz, R0, R0.wwww */
 261                 0x04400280, 0x1c9d5504, 0x0001c800, 0x0001ff00,
 262                 /* INST 16: MADH R0.w (TR0.xyzw), -R3.zzzz, R1.xxxx, R1.xxxx */
 263                 0x04401080, 0x1c9f550c, 0x00000104, 0x00000104,
 264                 /* INST 17: TEXR R0.yz (TR0.xyzw), attrib.texcoord[1], abs(texture[2]) */
 265                 0x1704ac80, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 266                 /* INST 18: MADH R0.x (TR0.xyzw), R3.zzzz, R0, R0.wwww */
 267                 0x04400280, 0x1c9d550c, 0x0001c900, 0x0001ff00,
 268                 /* INST 19: MADH R1.xyz (TR0.xyzw), R0.xxxx, { 1.16, -0.87, 0.53, -1.08 }.xxxx, { 1.16, -0.87, 0.53, -1.08 }.yzww */
 269                 0x04400e82, 0x1c9c0100, 0x00000002, 0x0001f202,
 270                 0x3f9507c8, 0xbf5ee393, 0x3f078fef, 0xbf8a6762,
 271                 /* INST 20: MADH R1.xyz (TR0.xyzw), R0.yyyy, { 0.00, -0.39, 2.02, 0.00 }, R1 */
 272                 0x04400e82, 0x1c9cab00, 0x0001c802, 0x0001c904,
 273                 0x00000000, 0xbec890d6, 0x40011687, 0x00000000,
 274                 /* INST 21: MADH R0.xyz (TR0.xyzw), R0.zzzz, { 1.60, -0.81, 0.00, 0.00 }, R1 + END */
 275                 0x04400e81, 0x1c9d5500, 0x0001c802, 0x0001c904,
 276                 0x3fcc432d, 0xbf501a37, 0x00000000, 0x00000000,
 277         }
 278 };
 279
 280 nv_shader_t nv30_fp_yv12_bicubic = {
 281         .card_priv.NV30FP.num_regs = 4,
 282         .size = (24*4),
 283         .data = {
 284                 /* INST 0: MOVR R2.xy (TR0.xyzw), attrib.texcoord[0] */
 285                 0x01008604, 0x1c9dc801, 0x0001c800, 0x0001c800,
 286                 /* INST 1: ADDR R0.xy (TR0.xyzw), R2, { 0.50, 0.00, 0.00, 0.00 }.xxxx */
 287                 0x03000600, 0x1c9dc808, 0x00000002, 0x0001c800,
 288                 0x3f000000, 0x00000000, 0x00000000, 0x00000000,
 289                 /* INST 2: TEXR R3.xyz (TR0.xyzw), R0, texture[0] */
 290                 0x17000e06, 0x1c9dc800, 0x0001c800, 0x0001c800,
 291                 /* INST 3: TEXR R0.xyz (TR0.xyzw), R0.yyyy, texture[0] */
 292                 0x17000e00, 0x1c9caa00, 0x0001c800, 0x0001c800,
 293                 /* INST 4: MULR R1.xz (TR0.xyzw), R3.xxyy, { -1.00, 1.00, 0.00, 0.00 }.xxyy */
 294                 0x02000a02, 0x1c9ca00c, 0x0000a002, 0x0001c800,
 295                 0xbf800000, 0x3f800000, 0x00000000, 0x00000000,
 296                 /* INST 5: MULR R1.yw (TR0.xyzw), R0.xxyy, { -1.00, 1.00, 0.00, 0.00 }.xxyy */
 297                 0x02001402, 0x1c9ca000, 0x0000a002, 0x0001c800,
 298                 0xbf800000, 0x3f800000, 0x00000000, 0x00000000,
 299                 /* INST 6: ADDR R2 (TR0.xyzw), R2.xyxy, R1 */
 300                 0x03001e04, 0x1c9c8808, 0x0001c804, 0x0001c800,
 301                 /* INST 7: TEXR R0.x (TR0.xyzw), R2, texture[1] */
 302                 0x17020200, 0x1c9dc808, 0x0001c800, 0x0001c800,
 303                 /* INST 8: TEXR R1.y (TR0.xyzw), R2.xwxw, texture[1] */
 304                 0x17020402, 0x1c9d9808, 0x0001c800, 0x0001c800,
 305                 /* INST 9: TEXR R1.x (TR0.xyzw), R2.zyxy, texture[1] */
 306                 0x17020202, 0x1c9c8c08, 0x0001c800, 0x0001c800,
 307                 /* INST 10: LRPH R0.x (TR0.xyzw), R0.zzzz, R0, R1.yyyy */
 308                 0x1f400280, 0x1c9d5400, 0x0001c800, 0x0000aa04,
 309                 /* INST 11: TEXR R0.y (TR0.xyzw), R2.zwzz, texture[1] */
 310                 0x17020400, 0x1c9d5c08, 0x0001c800, 0x0001c800,
 311                 /* INST 12: LRPH R0.y (TR0.xyzw), R0.zzzz, R1.xxxx, R0 */
 312                 0x1f400480, 0x1c9d5400, 0x00000004, 0x0001c800,
 313                 /* INST 13: LRPH R0.x (TR0.xyzw), R3.zzzz, R0, R0.yyyy */
 314                 0x1f400280, 0x1c9d540c, 0x0001c900, 0x0000ab00,
 315                 /* INST 14: MADH R0.xyz (TR0.xyzw), R0.xxxx, { 1.16, -0.87, 0.53, -1.08 }.xxxx, { 1.16, -0.87, 0.53, -1.08 }.yzww */
 316                 0x04400e80, 0x1c9c0100, 0x00000002, 0x0001f202,
 317                 0x3f9507c8, 0xbf5ee393, 0x3f078fef, 0xbf8a6762,
 318                 /* INST 15: TEXR R1.yz (TR0.xyzw), attrib.texcoord[1], abs(texture[2]) */
 319                 0x1704ac02, 0x1c9dc801, 0x0001c800, 0x0001c800,
 320                 /* INST 16: MADH R0.xyz (TR0.xyzw), R1.yyyy, { 0.00, -0.39, 2.02, 0.00 }, R0 */
 321                 0x04400e80, 0x1c9caa04, 0x0001c802, 0x0001c900,
 322                 0x00000000, 0xbec890d6, 0x40011687, 0x00000000,
 323                 /* INST 17: MADH R0.xyz (TR0.xyzw), R1.zzzz, { 1.60, -0.81, 0.00, 0.00 }, R0 + END */
 324                 0x04400e81, 0x1c9d5404, 0x0001c802, 0x0001c900,
 325                 0x3fcc432d, 0xbf501a37, 0x00000000, 0x00000000,
 326         }
 327 };
 328
 329 nv_shader_t nv30_fp_yv12_bilinear = {
 330         .card_priv.NV30FP.num_regs = 2,
 331         .size = (8*4),
 332         .data = {
 333                 /* INST 0: TEXR R0.x (TR0.xyzw), attrib.texcoord[0], abs(texture[1]) */
 334                 0x17028200, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 335                 /* INST 1: MADR R1.xyz (TR0.xyzw), R0.xxxx, { 1.16, -0.87, 0.53, -1.08 }.xxxx, { 1.16, -0.87, 0.53, -1.08 }.yzww */
 336                 0x04000e02, 0x1c9c0000, 0x00000002, 0x0001f202,
 337                 0x3f9507c8, 0xbf5ee393, 0x3f078fef, 0xbf8a6762,
 338                 /* INST 2: TEXR R0.yz (TR0.xyzw), attrib.texcoord[1], abs(texture[2]) */
 339                 0x1704ac80, 0x1c9dc801, 0x0001c800, 0x3fe1c800,
 340                 /* INST 3: MADR R1.xyz (TR0.xyzw), R0.yyyy, { 0.00, -0.39, 2.02, 0.00 }, R1 */
 341                 0x04000e02, 0x1c9cab00, 0x0001c802, 0x0001c804,
 342                 0x00000000, 0xbec890d6, 0x40011687, 0x00000000,
 343                 /* INST 4: MADR R0.xyz (TR0.xyzw), R0.zzzz, { 1.60, -0.81, 0.00, 0.00 }, R1 + END */
 344                 0x04000e81, 0x1c9d5500, 0x0001c802, 0x0001c804,
 345                 0x3fcc432d, 0xbf501a37, 0x00000000, 0x00000000,
 346         }
 347 };
 348
 349