src/mesa/drivers/dri/r300/r300_blit.c

   1 /*
   2  * Copyright (C) 2009 Maciej Cencora <m.cencora@gmail.com>
   3  *
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining
   7  * a copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sublicense, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial
  16  * portions of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  */
  27
  28 #include "radeon_common.h"
  29 #include "r300_context.h"
  30
  31 #include "r300_blit.h"
  32 #include "r300_cmdbuf.h"
  33 #include "r300_emit.h"
  34 #include "r300_tex.h"
  35 #include "compiler/radeon_compiler.h"
  36 #include "compiler/radeon_opcodes.h"
  37
  38 static void vp_ins_outs(struct r300_vertex_program_compiler *c)
  39 {
  40     c->code->inputs[VERT_ATTRIB_POS] = 0;
  41     c->code->inputs[VERT_ATTRIB_TEX0] = 1;
  42     c->code->outputs[VERT_RESULT_HPOS] = 0;
  43     c->code->outputs[VERT_RESULT_TEX0] = 1;
  44 }
  45
  46 static void fp_allocate_hw_inputs(
  47     struct r300_fragment_program_compiler * c,
  48     void (*allocate)(void * data, unsigned input, unsigned hwreg),
  49     void * mydata)
  50 {
  51     allocate(mydata, FRAG_ATTRIB_TEX0, 0);
  52 }
  53
  54 static void create_vertex_program(struct r300_context *r300)
  55 {
  56     struct r300_vertex_program_compiler compiler;
  57     struct rc_instruction *inst;
  58
  59     rc_init(&compiler.Base);
  60
  61     inst = rc_insert_new_instruction(&compiler.Base, compiler.Base.Program.Instructions.Prev);
  62     inst->U.I.Opcode = RC_OPCODE_MOV;
  63     inst->U.I.DstReg.File = RC_FILE_OUTPUT;
  64     inst->U.I.DstReg.Index = VERT_RESULT_HPOS;
  65     inst->U.I.DstReg.RelAddr = 0;
  66     inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
  67     inst->U.I.SrcReg[0].Abs = 0;
  68     inst->U.I.SrcReg[0].File = RC_FILE_INPUT;
  69     inst->U.I.SrcReg[0].Index = VERT_ATTRIB_POS;
  70     inst->U.I.SrcReg[0].Negate = 0;
  71     inst->U.I.SrcReg[0].RelAddr = 0;
  72     inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;
  73
  74     inst = rc_insert_new_instruction(&compiler.Base, compiler.Base.Program.Instructions.Prev);
  75     inst->U.I.Opcode = RC_OPCODE_MOV;
  76     inst->U.I.DstReg.File = RC_FILE_OUTPUT;
  77     inst->U.I.DstReg.Index = VERT_RESULT_TEX0;
  78     inst->U.I.DstReg.RelAddr = 0;
  79     inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
  80     inst->U.I.SrcReg[0].Abs = 0;
  81     inst->U.I.SrcReg[0].File = RC_FILE_INPUT;
  82     inst->U.I.SrcReg[0].Index = VERT_ATTRIB_TEX0;
  83     inst->U.I.SrcReg[0].Negate = 0;
  84     inst->U.I.SrcReg[0].RelAddr = 0;
  85     inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;
  86
  87     compiler.Base.Program.InputsRead = (1 << VERT_ATTRIB_POS) | (1 << VERT_ATTRIB_TEX0);
  88     compiler.RequiredOutputs = compiler.Base.Program.OutputsWritten = (1 << VERT_RESULT_HPOS) | (1 << VERT_RESULT_TEX0);
  89     compiler.SetHwInputOutput = vp_ins_outs;
  90     compiler.code = &r300->blit.vp_code;
  91
  92     r3xx_compile_vertex_program(&compiler);
  93 }
  94
  95 static void create_fragment_program(struct r300_context *r300)
  96 {
  97     struct r300_fragment_program_compiler compiler;
  98     struct rc_instruction *inst;
  99
 100     memset(&compiler, 0, sizeof(struct r300_fragment_program_compiler));
 101     rc_init(&compiler.Base);
 102
 103     inst = rc_insert_new_instruction(&compiler.Base, compiler.Base.Program.Instructions.Prev);
 104     inst->U.I.Opcode = RC_OPCODE_TEX;
 105     inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
 106     inst->U.I.TexSrcUnit = 0;
 107     inst->U.I.DstReg.File = RC_FILE_OUTPUT;
 108     inst->U.I.DstReg.Index = FRAG_RESULT_COLOR;
 109     inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 110     inst->U.I.SrcReg[0].Abs = 0;
 111     inst->U.I.SrcReg[0].File = RC_FILE_INPUT;
 112     inst->U.I.SrcReg[0].Index = FRAG_ATTRIB_TEX0;
 113     inst->U.I.SrcReg[0].Negate = 0;
 114     inst->U.I.SrcReg[0].RelAddr = 0;
 115     inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;
 116
 117     compiler.Base.Program.InputsRead = (1 << FRAG_ATTRIB_TEX0);
 118     compiler.OutputColor[0] = FRAG_RESULT_COLOR;
 119     compiler.OutputDepth = FRAG_RESULT_DEPTH;
 120     compiler.is_r500 = (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515);
 121     compiler.code = &r300->blit.fp_code;
 122     compiler.AllocateHwInputs = fp_allocate_hw_inputs;
 123
 124     r3xx_compile_fragment_program(&compiler);
 125 }
 126
 127 void r300_blit_init(struct r300_context *r300)
 128 {
 129     if (r300->options.hw_tcl_enabled)
 130         create_vertex_program(r300);
 131     create_fragment_program(r300);
 132 }
 133
 134 static void r300_emit_tx_setup(struct r300_context *r300,
 135                                gl_format mesa_format,
 136                                struct radeon_bo *bo,
 137                                intptr_t offset,
 138                                unsigned width,
 139                                unsigned height,
 140                                unsigned pitch)
 141 {
 142     BATCH_LOCALS(&r300->radeon);
 143
 144     assert(width <= 2048);
 145     assert(height <= 2048);
 146     assert(r300TranslateTexFormat(mesa_format) >= 0);
 147     assert(offset % 32 == 0);
 148
 149     BEGIN_BATCH(17);
 150     OUT_BATCH_REGVAL(R300_TX_FILTER0_0,
 151                      (R300_TX_CLAMP_TO_EDGE  << R300_TX_WRAP_S_SHIFT) |
 152                      (R300_TX_CLAMP_TO_EDGE  << R300_TX_WRAP_T_SHIFT) |
 153                      (R300_TX_CLAMP_TO_EDGE  << R300_TX_WRAP_R_SHIFT) |
 154                      R300_TX_MIN_FILTER_MIP_NONE |
 155                      R300_TX_MIN_FILTER_NEAREST |
 156                      R300_TX_MAG_FILTER_NEAREST |
 157                      (0 << 28));
 158     OUT_BATCH_REGVAL(R300_TX_FILTER1_0, 0);
 159     OUT_BATCH_REGVAL(R300_TX_SIZE_0,
 160                      ((width-1) << R300_TX_WIDTHMASK_SHIFT) |
 161                      ((height-1) << R300_TX_HEIGHTMASK_SHIFT) |
 162                      (0 << R300_TX_DEPTHMASK_SHIFT) |
 163                      (0 << R300_TX_MAX_MIP_LEVEL_SHIFT) |
 164                      R300_TX_SIZE_TXPITCH_EN);
 165
 166     OUT_BATCH_REGVAL(R300_TX_FORMAT_0, r300TranslateTexFormat(mesa_format));
 167     OUT_BATCH_REGVAL(R300_TX_FORMAT2_0, pitch - 1);
 168     OUT_BATCH_REGSEQ(R300_TX_OFFSET_0, 1);
 169     OUT_BATCH_RELOC(0, bo, offset, RADEON_GEM_DOMAIN_GTT|RADEON_GEM_DOMAIN_VRAM, 0, 0);
 170
 171     OUT_BATCH_REGSEQ(R300_TX_INVALTAGS, 2);
 172     OUT_BATCH(0);
 173     OUT_BATCH(1);
 174
 175     END_BATCH();
 176 }
 177
 178 #define EASY_US_FORMAT(FMT, C0, C1, C2, C3, SIGN) \
 179     (FMT  | R500_C0_SEL_##C0 | R500_C1_SEL_##C1 | \
 180     R500_C2_SEL_##C2 | R500_C3_SEL_##C3 | R500_OUT_SIGN(SIGN))
 181
 182 static uint32_t mesa_format_to_us_format(gl_format mesa_format)
 183 {
 184     switch(mesa_format)
 185     {
 186         case MESA_FORMAT_RGBA8888: // x
 187             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, A, B, G, R, 0);
 188         case MESA_FORMAT_RGB565: // x
 189         case MESA_FORMAT_ARGB1555: // x
 190         case MESA_FORMAT_RGBA8888_REV: // x
 191             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, R, G, B, A, 0);
 192         case MESA_FORMAT_ARGB8888: // x
 193             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, B, G, R, A, 0);
 194         case MESA_FORMAT_ARGB8888_REV:
 195             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, A, R, G, B, 0);
 196         case MESA_FORMAT_XRGB8888:
 197             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, A, R, G, B, 0);
 198
 199         case MESA_FORMAT_RGB332:
 200             return EASY_US_FORMAT(R500_OUT_FMT_C_3_3_2, A, R, G, B, 0);
 201
 202         case MESA_FORMAT_RGBA_FLOAT32:
 203             return EASY_US_FORMAT(R500_OUT_FMT_C4_32_FP, R, G, B, A, 0);
 204         case MESA_FORMAT_RGBA_FLOAT16:
 205             return EASY_US_FORMAT(R500_OUT_FMT_C4_16_FP, R, G, B, A, 0);
 206         case MESA_FORMAT_ALPHA_FLOAT32:
 207             return EASY_US_FORMAT(R500_OUT_FMT_C_32_FP, A, A, A, A, 0);
 208         case MESA_FORMAT_ALPHA_FLOAT16:
 209             return EASY_US_FORMAT(R500_OUT_FMT_C_16_FP, A, A, A, A, 0);
 210
 211         case MESA_FORMAT_SIGNED_RGBA8888:
 212             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, R, G, B, A, 0xf);
 213         case MESA_FORMAT_SIGNED_RGBA8888_REV:
 214             return EASY_US_FORMAT(R500_OUT_FMT_C4_8, A, B, G, R, 0xf);
 215         case MESA_FORMAT_SIGNED_RGBA_16:
 216             return EASY_US_FORMAT(R500_OUT_FMT_C4_16, R, G, B, A, 0xf);
 217
 218         default:
 219             fprintf(stderr, "Unsupported format %s for US output\n", _mesa_get_format_name(mesa_format));
 220             assert(0);
 221             return 0;
 222     }
 223 }
 224 #undef EASY_US_FORMAT
 225
 226 static void r500_emit_fp_setup(struct r300_context *r300,
 227                                struct r500_fragment_program_code *fp,
 228                                gl_format dst_format)
 229 {
 230     r500_emit_fp(r300, (uint32_t *)fp->inst, (fp->inst_end + 1) * 6, 0, 0, 0);
 231     BATCH_LOCALS(&r300->radeon);
 232
 233     BEGIN_BATCH(10);
 234     OUT_BATCH_REGSEQ(R500_US_CODE_ADDR, 3);
 235     OUT_BATCH(R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(fp->inst_end));
 236     OUT_BATCH(R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(fp->inst_end));
 237     OUT_BATCH(0);
 238     OUT_BATCH_REGVAL(R500_US_CONFIG, 0);
 239     OUT_BATCH_REGVAL(R500_US_OUT_FMT_0, mesa_format_to_us_format(dst_format));
 240     OUT_BATCH_REGVAL(R500_US_PIXSIZE, fp->max_temp_idx);
 241     END_BATCH();
 242 }
 243
 244 static void r500_emit_rs_setup(struct r300_context *r300)
 245 {
 246     BATCH_LOCALS(&r300->radeon);
 247
 248     BEGIN_BATCH(7);
 249     OUT_BATCH_REGSEQ(R300_RS_COUNT, 2);
 250     OUT_BATCH((4 << R300_IT_COUNT_SHIFT) | R300_HIRES_EN);
 251     OUT_BATCH(0);
 252     OUT_BATCH_REGVAL(R500_RS_INST_0,
 253                      (0 << R500_RS_INST_TEX_ID_SHIFT) |
 254                      (0 << R500_RS_INST_TEX_ADDR_SHIFT) |
 255                      R500_RS_INST_TEX_CN_WRITE |
 256                      R500_RS_INST_COL_CN_NO_WRITE);
 257     OUT_BATCH_REGVAL(R500_RS_IP_0,
 258                      (0 << R500_RS_IP_TEX_PTR_S_SHIFT) |
 259                      (1 << R500_RS_IP_TEX_PTR_T_SHIFT) |
 260                      (2 << R500_RS_IP_TEX_PTR_R_SHIFT) |
 261                      (3 << R500_RS_IP_TEX_PTR_Q_SHIFT));
 262     END_BATCH();
 263 }
 264
 265 static void r300_emit_fp_setup(struct r300_context *r300,
 266                                struct r300_fragment_program_code *code,
 267                                gl_format dst_format)
 268 {
 269     unsigned i;
 270     BATCH_LOCALS(&r300->radeon);
 271
 272     BEGIN_BATCH((code->alu.length + 1) * 4 + code->tex.length + 1 + 11);
 273
 274     OUT_BATCH_REGSEQ(R300_US_ALU_RGB_INST_0, code->alu.length);
 275     for (i = 0; i < code->alu.length; i++) {
 276         OUT_BATCH(code->alu.inst[i].rgb_inst);
 277     }
 278     OUT_BATCH_REGSEQ(R300_US_ALU_RGB_ADDR_0, code->alu.length);
 279     for (i = 0; i < code->alu.length; i++) {
 280         OUT_BATCH(code->alu.inst[i].rgb_addr);
 281     }
 282     OUT_BATCH_REGSEQ(R300_US_ALU_ALPHA_INST_0, code->alu.length);
 283     for (i = 0; i < code->alu.length; i++) {
 284         OUT_BATCH(code->alu.inst[i].alpha_inst);
 285     }
 286     OUT_BATCH_REGSEQ(R300_US_ALU_ALPHA_ADDR_0, code->alu.length);
 287     for (i = 0; i < code->alu.length; i++) {
 288         OUT_BATCH(code->alu.inst[i].alpha_addr);
 289     }
 290
 291     OUT_BATCH_REGSEQ(R300_US_TEX_INST_0, code->tex.length);
 292     OUT_BATCH_TABLE(code->tex.inst, code->tex.length);
 293
 294     OUT_BATCH_REGSEQ(R300_US_CONFIG, 3);
 295     OUT_BATCH(R300_PFS_CNTL_FIRST_NODE_HAS_TEX);
 296     OUT_BATCH(code->pixsize);
 297     OUT_BATCH(code->code_offset);
 298     OUT_BATCH_REGSEQ(R300_US_CODE_ADDR_0, 4);
 299     OUT_BATCH_TABLE(code->code_addr, 4);
 300     OUT_BATCH_REGVAL(R500_US_OUT_FMT_0, mesa_format_to_us_format(dst_format));
 301     END_BATCH();
 302 }
 303
 304 static void r300_emit_rs_setup(struct r300_context *r300)
 305 {
 306     BATCH_LOCALS(&r300->radeon);
 307
 308     BEGIN_BATCH(7);
 309     OUT_BATCH_REGSEQ(R300_RS_COUNT, 2);
 310     OUT_BATCH((4 << R300_IT_COUNT_SHIFT) | R300_HIRES_EN);
 311     OUT_BATCH(0);
 312     OUT_BATCH_REGVAL(R300_RS_INST_0,
 313                      R300_RS_INST_TEX_ID(0) |
 314                      R300_RS_INST_TEX_ADDR(0) |
 315                      R300_RS_INST_TEX_CN_WRITE);
 316     OUT_BATCH_REGVAL(R300_RS_IP_0,
 317                      R300_RS_TEX_PTR(0) |
 318                      R300_RS_SEL_S(R300_RS_SEL_C0) |
 319                      R300_RS_SEL_T(R300_RS_SEL_C1) |
 320                      R300_RS_SEL_R(R300_RS_SEL_K0) |
 321                      R300_RS_SEL_Q(R300_RS_SEL_K1));
 322     END_BATCH();
 323 }
 324
 325 static void emit_pvs_setup(struct r300_context *r300,
 326                            uint32_t *vp_code,
 327                            unsigned vp_len)
 328 {
 329     BATCH_LOCALS(&r300->radeon);
 330
 331     r300_emit_vpu(r300, vp_code, vp_len * 4, R300_PVS_CODE_START);
 332
 333     BEGIN_BATCH(4);
 334     OUT_BATCH_REGSEQ(R300_VAP_PVS_CODE_CNTL_0, 3);
 335     OUT_BATCH((0 << R300_PVS_FIRST_INST_SHIFT) |
 336               ((vp_len - 1)  << R300_PVS_XYZW_VALID_INST_SHIFT) |
 337               ((vp_len - 1)<< R300_PVS_LAST_INST_SHIFT));
 338     OUT_BATCH(0);
 339     OUT_BATCH((vp_len - 1) << R300_PVS_LAST_VTX_SRC_INST_SHIFT);
 340     END_BATCH();
 341 }
 342
 343 static void emit_vap_setup(struct r300_context *r300)
 344 {
 345     int tex_offset;
 346     BATCH_LOCALS(&r300->radeon);
 347
 348     if (r300->options.hw_tcl_enabled)
 349         tex_offset = 1;
 350     else
 351         tex_offset = 6;
 352
 353     BEGIN_BATCH(12);
 354     OUT_BATCH_REGSEQ(R300_SE_VTE_CNTL, 2);
 355     OUT_BATCH(R300_VTX_XY_FMT | R300_VTX_Z_FMT);
 356     OUT_BATCH(4);
 357
 358     OUT_BATCH_REGVAL(R300_VAP_PSC_SGN_NORM_CNTL, 0xaaaaaaaa);
 359     OUT_BATCH_REGVAL(R300_VAP_PROG_STREAM_CNTL_0,
 360                      ((R300_DATA_TYPE_FLOAT_2 | (0 << R300_DST_VEC_LOC_SHIFT)) << 0) |
 361                      (((tex_offset << R300_DST_VEC_LOC_SHIFT) | R300_DATA_TYPE_FLOAT_2 | R300_LAST_VEC) << 16));
 362     OUT_BATCH_REGVAL(R300_VAP_PROG_STREAM_CNTL_EXT_0,
 363                     ((((R300_SWIZZLE_SELECT_X << R300_SWIZZLE_SELECT_X_SHIFT) |
 364                        (R300_SWIZZLE_SELECT_Y << R300_SWIZZLE_SELECT_Y_SHIFT) |
 365                        (R300_SWIZZLE_SELECT_FP_ZERO << R300_SWIZZLE_SELECT_Z_SHIFT) |
 366                        (R300_SWIZZLE_SELECT_FP_ONE << R300_SWIZZLE_SELECT_W_SHIFT) |
 367                        (0xf << R300_WRITE_ENA_SHIFT) ) << 0) |
 368                      (((R300_SWIZZLE_SELECT_X << R300_SWIZZLE_SELECT_X_SHIFT) |
 369                        (R300_SWIZZLE_SELECT_Y << R300_SWIZZLE_SELECT_Y_SHIFT) |
 370                        (R300_SWIZZLE_SELECT_FP_ZERO << R300_SWIZZLE_SELECT_Z_SHIFT) |
 371                        (R300_SWIZZLE_SELECT_FP_ONE << R300_SWIZZLE_SELECT_W_SHIFT) |
 372                        (0xf << R300_WRITE_ENA_SHIFT) ) << 16) ) );
 373     OUT_BATCH_REGSEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
 374     OUT_BATCH(R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT);
 375     OUT_BATCH(R300_VAP_OUTPUT_VTX_FMT_1__4_COMPONENTS);
 376     END_BATCH();
 377 }
 378
 379 static GLboolean validate_buffers(struct r300_context *r300,
 380                                   struct radeon_bo *src_bo,
 381                                   struct radeon_bo *dst_bo)
 382 {
 383     int ret;
 384
 385     radeon_cs_space_reset_bos(r300->radeon.cmdbuf.cs);
 386
 387     ret = radeon_cs_space_check_with_bo(r300->radeon.cmdbuf.cs,
 388                                         src_bo, RADEON_GEM_DOMAIN_VRAM, 0);
 389     if (ret)
 390         return GL_FALSE;
 391
 392     ret = radeon_cs_space_check_with_bo(r300->radeon.cmdbuf.cs,
 393                                         dst_bo, 0, RADEON_GEM_DOMAIN_VRAM);
 394     if (ret)
 395         return GL_FALSE;
 396
 397     ret = radeon_cs_space_check_with_bo(r300->radeon.cmdbuf.cs,
 398                                         first_elem(&r300->radeon.dma.reserved)->bo,
 399                                         RADEON_GEM_DOMAIN_GTT, 0);
 400     if (ret)
 401         return GL_FALSE;
 402
 403     return GL_TRUE;
 404 }
 405
 406 /**
 407  * Calculate texcoords for given image region.
 408  * Output values are [minx, maxx, miny, maxy]
 409  */
 410 static void calc_tex_coords(float img_width, float img_height,
 411                             float x, float y,
 412                             float reg_width, float reg_height,
 413                             unsigned flip_y, float *buf)
 414 {
 415     buf[0] = x / img_width;
 416     buf[1] = buf[0] + reg_width / img_width;
 417     buf[2] = y / img_height;
 418     buf[3] = buf[2] + reg_height / img_height;
 419     if (flip_y)
 420     {
 421         buf[2] = 1.0 - buf[2];
 422         buf[3] = 1.0 - buf[3];
 423     }
 424 }
 425
 426 static void emit_draw_packet(struct r300_context *r300,
 427                              unsigned src_width, unsigned src_height,
 428                              unsigned src_x_offset, unsigned src_y_offset,
 429                              unsigned dst_x_offset, unsigned dst_y_offset,
 430                              unsigned reg_width, unsigned reg_height,
 431                              unsigned flip_y)
 432 {
 433     float texcoords[4];
 434
 435     calc_tex_coords(src_width, src_height,
 436                     src_x_offset, src_y_offset,
 437                     reg_width, reg_height,
 438                     flip_y, texcoords);
 439
 440     float verts[] = { dst_x_offset, dst_y_offset,
 441                       texcoords[0], texcoords[2],
 442                       dst_x_offset, dst_y_offset + reg_height,
 443                       texcoords[0], texcoords[3],
 444                       dst_x_offset + reg_width, dst_y_offset + reg_height,
 445                       texcoords[1], texcoords[3],
 446                       dst_x_offset + reg_width, dst_y_offset,
 447                       texcoords[1], texcoords[2] };
 448
 449     BATCH_LOCALS(&r300->radeon);
 450
 451     BEGIN_BATCH(19);
 452     OUT_BATCH_PACKET3(R300_PACKET3_3D_DRAW_IMMD_2, 16);
 453     OUT_BATCH(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED |
 454               (4 << 16) | R300_VAP_VF_CNTL__PRIM_QUADS);
 455     OUT_BATCH_TABLE(verts, 16);
 456     END_BATCH();
 457 }
 458
 459 static void other_stuff(struct r300_context *r300)
 460 {
 461     BATCH_LOCALS(&r300->radeon);
 462
 463     BEGIN_BATCH(13);
 464     OUT_BATCH_REGVAL(R300_GA_POLY_MODE,
 465                      R300_GA_POLY_MODE_FRONT_PTYPE_TRI | R300_GA_POLY_MODE_BACK_PTYPE_TRI);
 466     OUT_BATCH_REGVAL(R300_SU_CULL_MODE, R300_FRONT_FACE_CCW);
 467     OUT_BATCH_REGVAL(R300_FG_FOG_BLEND, 0);
 468     OUT_BATCH_REGVAL(R300_FG_ALPHA_FUNC, 0);
 469     OUT_BATCH_REGSEQ(R300_RB3D_CBLEND, 2);
 470     OUT_BATCH(0x0);
 471     OUT_BATCH(0x0);
 472     OUT_BATCH_REGVAL(R300_ZB_CNTL, 0);
 473     END_BATCH();
 474     if (r300->options.hw_tcl_enabled) {
 475         BEGIN_BATCH(2);
 476         OUT_BATCH_REGVAL(R300_VAP_CLIP_CNTL, R300_CLIP_DISABLE);
 477         END_BATCH();
 478     }
 479 }
 480
 481 static void emit_cb_setup(struct r300_context *r300,
 482                           struct radeon_bo *bo,
 483                           intptr_t offset,
 484                           gl_format mesa_format,
 485                           unsigned pitch,
 486                           unsigned width,
 487                           unsigned height)
 488 {
 489     BATCH_LOCALS(&r300->radeon);
 490
 491     unsigned x1, y1, x2, y2;
 492     x1 = 0;
 493     y1 = 0;
 494     x2 = width - 1;
 495     y2 = height - 1;
 496
 497     if (r300->radeon.radeonScreen->chip_family < CHIP_FAMILY_RV515) {
 498         x1 += R300_SCISSORS_OFFSET;
 499         y1 += R300_SCISSORS_OFFSET;
 500         x2 += R300_SCISSORS_OFFSET;
 501         y2 += R300_SCISSORS_OFFSET;
 502     }
 503
 504     r300_emit_cb_setup(r300, bo, offset, mesa_format,
 505                        _mesa_get_format_bytes(mesa_format),
 506                        _mesa_format_row_stride(mesa_format, pitch));
 507
 508     BEGIN_BATCH_NO_AUTOSTATE(5);
 509     OUT_BATCH_REGSEQ(R300_SC_SCISSORS_TL, 2);
 510     OUT_BATCH((x1 << R300_SCISSORS_X_SHIFT)|(y1 << R300_SCISSORS_Y_SHIFT));
 511     OUT_BATCH((x2 << R300_SCISSORS_X_SHIFT)|(y2 << R300_SCISSORS_Y_SHIFT));
 512     OUT_BATCH_REGVAL(R300_RB3D_CCTL, 0);
 513     END_BATCH();
 514 }
 515
 516 unsigned r300_check_blit(gl_format dst_format)
 517 {
 518     switch (dst_format) {
 519         case MESA_FORMAT_RGB565:
 520         case MESA_FORMAT_ARGB1555:
 521         case MESA_FORMAT_RGBA8888:
 522         case MESA_FORMAT_RGBA8888_REV:
 523         case MESA_FORMAT_ARGB8888:
 524         case MESA_FORMAT_ARGB8888_REV:
 525         case MESA_FORMAT_XRGB8888:
 526             break;
 527         default:
 528             return 0;
 529     }
 530
 531     if (_mesa_get_format_bits(dst_format, GL_DEPTH_BITS) > 0)
 532         return 0;
 533
 534     return 1;
 535 }
 536
 537 /**
 538  * Copy a region of [@a width x @a height] pixels from source buffer
 539  * to destination buffer.
 540  * @param[in] r300 r300 context
 541  * @param[in] src_bo source radeon buffer object
 542  * @param[in] src_offset offset of the source image in the @a src_bo
 543  * @param[in] src_mesaformat source image format
 544  * @param[in] src_pitch aligned source image width
 545  * @param[in] src_width source image width
 546  * @param[in] src_height source image height
 547  * @param[in] src_x_offset x offset in the source image
 548  * @param[in] src_y_offset y offset in the source image
 549  * @param[in] dst_bo destination radeon buffer object
 550  * @param[in] dst_offset offset of the destination image in the @a dst_bo
 551  * @param[in] dst_mesaformat destination image format
 552  * @param[in] dst_pitch aligned destination image width
 553  * @param[in] dst_width destination image width
 554  * @param[in] dst_height destination image height
 555  * @param[in] dst_x_offset x offset in the destination image
 556  * @param[in] dst_y_offset y offset in the destination image
 557  * @param[in] width region width
 558  * @param[in] height region height
 559  * @param[in] flip_y set if y coords of the source image need to be flipped
 560  */
 561 unsigned r300_blit(GLcontext *ctx,
 562                    struct radeon_bo *src_bo,
 563                    intptr_t src_offset,
 564                    gl_format src_mesaformat,
 565                    unsigned src_pitch,
 566                    unsigned src_width,
 567                    unsigned src_height,
 568                    unsigned src_x_offset,
 569                    unsigned src_y_offset,
 570                    struct radeon_bo *dst_bo,
 571                    intptr_t dst_offset,
 572                    gl_format dst_mesaformat,
 573                    unsigned dst_pitch,
 574                    unsigned dst_width,
 575                    unsigned dst_height,
 576                    unsigned dst_x_offset,
 577                    unsigned dst_y_offset,
 578                    unsigned reg_width,
 579                    unsigned reg_height,
 580                    unsigned flip_y)
 581 {
 582     r300ContextPtr r300 = R300_CONTEXT(ctx);
 583
 584     if (!r300_check_blit(dst_mesaformat))
 585         return 0;
 586
 587     /* Make sure that colorbuffer has even width - hw limitation */
 588     if (dst_pitch % 2 > 0)
 589         ++dst_pitch;
 590
 591     /* Rendering to small buffer doesn't work.
 592      * Looks like a hw limitation.
 593      */
 594     if (dst_pitch < 32)
 595         return 0;
 596
 597     /* Need to clamp the region size to make sure
 598      * we don't read outside of the source buffer
 599      * or write outside of the destination buffer.
 600      */
 601     if (reg_width + src_x_offset > src_width)
 602         reg_width = src_width - src_x_offset;
 603     if (reg_height + src_y_offset > src_height)
 604         reg_height = src_height - src_y_offset;
 605     if (reg_width + dst_x_offset > dst_width)
 606         reg_width = dst_width - dst_x_offset;
 607     if (reg_height + dst_y_offset > dst_height)
 608         reg_height = dst_height - dst_y_offset;
 609
 610     if (src_bo == dst_bo) {
 611         return 0;
 612     }
 613
 614     if (src_offset % 32 || dst_offset % 32) {
 615         return GL_FALSE;
 616     }
 617
 618     if (0) {
 619         fprintf(stderr, "src: size [%d x %d], pitch %d, "
 620                 "offset [%d x %d], format %s, bo %p\n",
 621                 src_width, src_height, src_pitch,
 622                 src_x_offset, src_y_offset,
 623                 _mesa_get_format_name(src_mesaformat),
 624                 src_bo);
 625         fprintf(stderr, "dst: pitch %d, offset[%d x %d], format %s, bo %p\n",
 626                 dst_pitch, dst_x_offset, dst_y_offset,
 627                 _mesa_get_format_name(dst_mesaformat), dst_bo);
 628         fprintf(stderr, "region: %d x %d\n", reg_width, reg_height);
 629     }
 630
 631     /* Flush is needed to make sure that source buffer has correct data */
 632     radeonFlush(r300->radeon.glCtx);
 633
 634     if (!validate_buffers(r300, src_bo, dst_bo))
 635         return 0;
 636
 637     rcommonEnsureCmdBufSpace(&r300->radeon, 200, __FUNCTION__);
 638
 639     other_stuff(r300);
 640
 641     r300_emit_tx_setup(r300, src_mesaformat, src_bo, src_offset, src_width, src_height, src_pitch);
 642
 643     if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
 644         r500_emit_fp_setup(r300, &r300->blit.fp_code.code.r500, dst_mesaformat);
 645         r500_emit_rs_setup(r300);
 646     } else {
 647         r300_emit_fp_setup(r300, &r300->blit.fp_code.code.r300, dst_mesaformat);
 648         r300_emit_rs_setup(r300);
 649     }
 650
 651     if (r300->options.hw_tcl_enabled)
 652         emit_pvs_setup(r300, r300->blit.vp_code.body.d, 2);
 653
 654     emit_vap_setup(r300);
 655
 656     emit_cb_setup(r300, dst_bo, dst_offset, dst_mesaformat, dst_pitch, dst_width, dst_height);
 657
 658     emit_draw_packet(r300, src_width, src_height,
 659                      src_x_offset, src_y_offset,
 660                      dst_x_offset, dst_y_offset,
 661                      reg_width, reg_height,
 662                      flip_y);
 663
 664     r300EmitCacheFlush(r300);
 665
 666     radeonFlush(r300->radeon.glCtx);
 667
 668     return 1;
 669 }