src/gallium/drivers/freedreno/a3xx/fd3_emit.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2013 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_string.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_helpers.h"
  33 #include "util/u_format.h"
  34
  35 #include "freedreno_resource.h"
  36
  37 #include "fd3_emit.h"
  38 #include "fd3_blend.h"
  39 #include "fd3_context.h"
  40 #include "fd3_program.h"
  41 #include "fd3_rasterizer.h"
  42 #include "fd3_texture.h"
  43 #include "fd3_util.h"
  44 #include "fd3_zsa.h"
  45
  46 /* regid:          base const register
  47  * prsc or dwords: buffer containing constant values
  48  * sizedwords:     size of const value buffer
  49  */
  50 void
  51 fd3_emit_constant(struct fd_ringbuffer *ring,
  52                 enum adreno_state_block sb,
  53                 uint32_t regid, uint32_t offset, uint32_t sizedwords,
  54                 const uint32_t *dwords, struct pipe_resource *prsc)
  55 {
  56         uint32_t i, sz;
  57         enum adreno_state_src src;
  58
  59         if (prsc) {
  60                 sz = 0;
  61                 src = SS_INDIRECT;
  62         } else {
  63                 sz = sizedwords;
  64                 src = SS_DIRECT;
  65         }
  66
  67         /* we have this sometimes, not others.. perhaps we could be clever
  68          * and figure out actually when we need to invalidate cache:
  69          */
  70         OUT_PKT0(ring, REG_A3XX_UCHE_CACHE_INVALIDATE0_REG, 2);
  71         OUT_RING(ring, A3XX_UCHE_CACHE_INVALIDATE0_REG_ADDR(0));
  72         OUT_RING(ring, A3XX_UCHE_CACHE_INVALIDATE1_REG_ADDR(0) |
  73                         A3XX_UCHE_CACHE_INVALIDATE1_REG_OPCODE(INVALIDATE) |
  74                         A3XX_UCHE_CACHE_INVALIDATE1_REG_ENTIRE_CACHE);
  75
  76         OUT_PKT3(ring, CP_LOAD_STATE, 2 + sz);
  77         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(regid/2) |
  78                         CP_LOAD_STATE_0_STATE_SRC(src) |
  79                         CP_LOAD_STATE_0_STATE_BLOCK(sb) |
  80                         CP_LOAD_STATE_0_NUM_UNIT(sizedwords/2));
  81         if (prsc) {
  82                 struct fd_bo *bo = fd_resource(prsc)->bo;
  83                 OUT_RELOC(ring, bo, offset,
  84                                 CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS));
  85         } else {
  86                 OUT_RING(ring, CP_LOAD_STATE_1_EXT_SRC_ADDR(0) |
  87                                 CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS));
  88                 dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
  89         }
  90         for (i = 0; i < sz; i++) {
  91                 OUT_RING(ring, dwords[i]);
  92         }
  93 }
  94
  95 static void
  96 emit_constants(struct fd_ringbuffer *ring,
  97                 enum adreno_state_block sb,
  98                 struct fd_constbuf_stateobj *constbuf,
  99                 struct fd3_shader_stateobj *shader)
 100 {
 101         uint32_t enabled_mask = constbuf->enabled_mask;
 102         uint32_t base = 0;
 103         unsigned i;
 104
 105         // XXX TODO only emit dirty consts.. but we need to keep track if
 106         // they are clobbered by a clear, gmem2mem, or mem2gmem..
 107         constbuf->dirty_mask = enabled_mask;
 108
 109         /* emit user constants: */
 110         while (enabled_mask) {
 111                 unsigned index = ffs(enabled_mask) - 1;
 112                 struct pipe_constant_buffer *cb = &constbuf->cb[index];
 113                 unsigned size = align(cb->buffer_size, 4) / 4; /* size in dwords */
 114
 115                 // I expect that size should be a multiple of vec4's:
 116                 assert(size == align(size, 4));
 117
 118                 /* gallium could have const-buffer still bound, even though the
 119                  * shader is not using it.  Writing consts above constlen (or
 120                  * rather, HLSQ_{VS,FS}_CONTROL_REG.CONSTLENGTH) will cause a
 121                  * hang.
 122                  */
 123                 if ((base / 4) >= shader->constlen)
 124                         break;
 125
 126                 if (constbuf->dirty_mask & (1 << index)) {
 127                         fd3_emit_constant(ring, sb, base,
 128                                         cb->buffer_offset, size,
 129                                         cb->user_buffer, cb->buffer);
 130                         constbuf->dirty_mask &= ~(1 << index);
 131                 }
 132
 133                 base += size;
 134                 enabled_mask &= ~(1 << index);
 135         }
 136
 137         /* emit shader immediates: */
 138         if (shader) {
 139                 for (i = 0; i < shader->immediates_count; i++) {
 140                         fd3_emit_constant(ring, sb,
 141                                         4 * (shader->first_immediate + i),
 142                                         0, 4, shader->immediates[i].val, NULL);
 143                 }
 144         }
 145 }
 146
 147 #define VERT_TEX_OFF    0
 148 #define FRAG_TEX_OFF    16
 149 #define BASETABLE_SZ    14
 150
 151 static void
 152 emit_textures(struct fd_ringbuffer *ring,
 153                 enum adreno_state_block sb,
 154                 struct fd_texture_stateobj *tex)
 155 {
 156         static const unsigned tex_off[] = {
 157                         [SB_VERT_TEX] = VERT_TEX_OFF,
 158                         [SB_FRAG_TEX] = FRAG_TEX_OFF,
 159         };
 160         static const enum adreno_state_block mipaddr[] = {
 161                         [SB_VERT_TEX] = SB_VERT_MIPADDR,
 162                         [SB_FRAG_TEX] = SB_FRAG_MIPADDR,
 163         };
 164         unsigned i, j;
 165
 166         assert(tex->num_samplers == tex->num_textures);  // TODO check..
 167
 168         if (!tex->num_samplers)
 169                 return;
 170
 171         /* output sampler state: */
 172         OUT_PKT3(ring, CP_LOAD_STATE, 2 + (2 * tex->num_samplers));
 173         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(tex_off[sb]) |
 174                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 175                         CP_LOAD_STATE_0_STATE_BLOCK(sb) |
 176                         CP_LOAD_STATE_0_NUM_UNIT(tex->num_samplers));
 177         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_SHADER) |
 178                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 179         for (i = 0; i < tex->num_samplers; i++) {
 180                 struct fd3_sampler_stateobj *sampler =
 181                                 fd3_sampler_stateobj(tex->samplers[i]);
 182                 OUT_RING(ring, sampler->texsamp0);
 183                 OUT_RING(ring, sampler->texsamp1);
 184         }
 185
 186         /* emit texture state: */
 187         OUT_PKT3(ring, CP_LOAD_STATE, 2 + (4 * tex->num_textures));
 188         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(tex_off[sb]) |
 189                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 190                         CP_LOAD_STATE_0_STATE_BLOCK(sb) |
 191                         CP_LOAD_STATE_0_NUM_UNIT(tex->num_textures));
 192         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 193                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 194         for (i = 0; i < tex->num_textures; i++) {
 195                 struct fd3_pipe_sampler_view *view =
 196                                 fd3_pipe_sampler_view(tex->textures[i]);
 197                 OUT_RING(ring, view->texconst0);
 198                 OUT_RING(ring, view->texconst1);
 199                 OUT_RING(ring, view->texconst2 |
 200                                 A3XX_TEX_CONST_2_INDX(BASETABLE_SZ * i));
 201                 OUT_RING(ring, view->texconst3);
 202         }
 203
 204         /* emit mipaddrs: */
 205         OUT_PKT3(ring, CP_LOAD_STATE, 2 + (BASETABLE_SZ * tex->num_textures));
 206         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(BASETABLE_SZ * tex_off[sb]) |
 207                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 208                         CP_LOAD_STATE_0_STATE_BLOCK(mipaddr[sb]) |
 209                         CP_LOAD_STATE_0_NUM_UNIT(BASETABLE_SZ * tex->num_textures));
 210         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 211                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 212         for (i = 0; i < tex->num_textures; i++) {
 213                 struct fd3_pipe_sampler_view *view =
 214                                 fd3_pipe_sampler_view(tex->textures[i]);
 215                 OUT_RELOC(ring, view->tex_resource->bo, 0, 0);
 216                 /* I think each entry is a ptr to mipmap level.. for now, just
 217                  * pad w/ null's until I get around to actually implementing
 218                  * mipmap support..
 219                  */
 220                 for (j = 1; j < BASETABLE_SZ; j++) {
 221                         OUT_RING(ring, 0x00000000);
 222                 }
 223         }
 224 }
 225
 226 static void
 227 emit_cache_flush(struct fd_ringbuffer *ring)
 228 {
 229         OUT_PKT3(ring, CP_EVENT_WRITE, 1);
 230         OUT_RING(ring, CACHE_FLUSH);
 231
 232         OUT_PKT3(ring, CP_DRAW_INDX, 3);
 233         OUT_RING(ring, 0x00000000);
 234         OUT_RING(ring, DRAW(DI_PT_POINTLIST, DI_SRC_SEL_AUTO_INDEX,
 235                         INDEX_SIZE_IGN, IGNORE_VISIBILITY));
 236         OUT_RING(ring, 0);                                      /* NumIndices */
 237
 238         OUT_PKT3(ring, CP_NOP, 4);
 239         OUT_RING(ring, 0x00000000);
 240         OUT_RING(ring, 0x00000000);
 241         OUT_RING(ring, 0x00000000);
 242         OUT_RING(ring, 0x00000000);
 243
 244         OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
 245         OUT_RING(ring, 0x00000000);
 246 }
 247
 248 /* emit texture state for mem->gmem restore operation.. eventually it would
 249  * be good to get rid of this and use normal CSO/etc state for more of these
 250  * special cases, but for now the compiler is not sufficient..
 251  */
 252 void
 253 fd3_emit_gmem_restore_tex(struct fd_ringbuffer *ring, struct pipe_surface *psurf)
 254 {
 255         struct fd_resource *rsc = fd_resource(psurf->texture);
 256
 257         /* output sampler state: */
 258         OUT_PKT3(ring, CP_LOAD_STATE, 4);
 259         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(FRAG_TEX_OFF) |
 260                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 261                         CP_LOAD_STATE_0_STATE_BLOCK(SB_FRAG_TEX) |
 262                         CP_LOAD_STATE_0_NUM_UNIT(1));
 263         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_SHADER) |
 264                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 265         OUT_RING(ring, A3XX_TEX_SAMP_0_XY_MAG(A3XX_TEX_NEAREST) |
 266                         A3XX_TEX_SAMP_0_XY_MIN(A3XX_TEX_NEAREST) |
 267                         A3XX_TEX_SAMP_0_WRAP_S(A3XX_TEX_CLAMP_TO_EDGE) |
 268                         A3XX_TEX_SAMP_0_WRAP_T(A3XX_TEX_CLAMP_TO_EDGE) |
 269                         A3XX_TEX_SAMP_0_WRAP_R(A3XX_TEX_REPEAT));
 270         OUT_RING(ring, 0x00000000);
 271
 272         /* emit texture state: */
 273         OUT_PKT3(ring, CP_LOAD_STATE, 6);
 274         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(FRAG_TEX_OFF) |
 275                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 276                         CP_LOAD_STATE_0_STATE_BLOCK(SB_FRAG_TEX) |
 277                         CP_LOAD_STATE_0_NUM_UNIT(1));
 278         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 279                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 280         OUT_RING(ring, A3XX_TEX_CONST_0_FMT(fd3_pipe2tex(psurf->format)) |
 281                         0x40000000 | // XXX
 282                         fd3_tex_swiz(psurf->format,  PIPE_SWIZZLE_BLUE, PIPE_SWIZZLE_GREEN,
 283                                         PIPE_SWIZZLE_RED, PIPE_SWIZZLE_ALPHA));
 284         OUT_RING(ring, A3XX_TEX_CONST_1_FETCHSIZE(fd3_pipe2fetchsize(psurf->format)) |
 285                         A3XX_TEX_CONST_1_WIDTH(psurf->width) |
 286                         A3XX_TEX_CONST_1_HEIGHT(psurf->height));
 287         OUT_RING(ring, A3XX_TEX_CONST_2_PITCH(rsc->pitch * rsc->cpp) |
 288                         A3XX_TEX_CONST_2_INDX(0));
 289         OUT_RING(ring, 0x00000000);
 290
 291         /* emit mipaddrs: */
 292         OUT_PKT3(ring, CP_LOAD_STATE, 3);
 293         OUT_RING(ring, CP_LOAD_STATE_0_DST_OFF(BASETABLE_SZ * FRAG_TEX_OFF) |
 294                         CP_LOAD_STATE_0_STATE_SRC(SS_DIRECT) |
 295                         CP_LOAD_STATE_0_STATE_BLOCK(SB_FRAG_MIPADDR) |
 296                         CP_LOAD_STATE_0_NUM_UNIT(1));
 297         OUT_RING(ring, CP_LOAD_STATE_1_STATE_TYPE(ST_CONSTANTS) |
 298                         CP_LOAD_STATE_1_EXT_SRC_ADDR(0));
 299         OUT_RELOC(ring, rsc->bo, 0, 0);
 300 }
 301
 302 void
 303 fd3_emit_vertex_bufs(struct fd_ringbuffer *ring,
 304                 struct fd_program_stateobj *prog,
 305                 struct fd3_vertex_buf *vbufs, uint32_t n)
 306 {
 307         struct fd3_shader_stateobj *vp = prog->vp;
 308         uint32_t i;
 309
 310         n = MIN2(n, vp->inputs_count);
 311
 312         for (i = 0; i < n; i++) {
 313                 struct pipe_resource *prsc = vbufs[i].prsc;
 314                 struct fd_resource *rsc = fd_resource(prsc);
 315                 enum a3xx_vtx_fmt fmt = fd3_pipe2vtx(vbufs[i].format);
 316                 bool switchnext = (i != (n - 1));
 317                 uint32_t fs = util_format_get_blocksize(vbufs[i].format);
 318
 319                 OUT_PKT0(ring, REG_A3XX_VFD_FETCH(i), 2);
 320                 OUT_RING(ring, A3XX_VFD_FETCH_INSTR_0_FETCHSIZE(fs - 1) |
 321                                 A3XX_VFD_FETCH_INSTR_0_BUFSTRIDE(vbufs[i].stride) |
 322                                 COND(switchnext, A3XX_VFD_FETCH_INSTR_0_SWITCHNEXT) |
 323                                 A3XX_VFD_FETCH_INSTR_0_INDEXCODE(i) |
 324                                 A3XX_VFD_FETCH_INSTR_0_STEPRATE(1));
 325                 OUT_RELOC(ring, rsc->bo, vbufs[i].offset, 0);
 326
 327                 OUT_PKT0(ring, REG_A3XX_VFD_DECODE_INSTR(i), 1);
 328                 OUT_RING(ring, A3XX_VFD_DECODE_INSTR_CONSTFILL |
 329                                 A3XX_VFD_DECODE_INSTR_WRITEMASK(vp->inputs[i].compmask) |
 330                                 A3XX_VFD_DECODE_INSTR_FORMAT(fmt) |
 331                                 A3XX_VFD_DECODE_INSTR_REGID(vp->inputs[i].regid) |
 332                                 A3XX_VFD_DECODE_INSTR_SHIFTCNT(fs) |
 333                                 A3XX_VFD_DECODE_INSTR_LASTCOMPVALID |
 334                                 COND(switchnext, A3XX_VFD_DECODE_INSTR_SWITCHNEXT));
 335         }
 336 }
 337
 338 void
 339 fd3_emit_state(struct fd_context *ctx, uint32_t dirty)
 340 {
 341         struct fd_ringbuffer *ring = ctx->ring;
 342
 343         if (dirty & FD_DIRTY_SAMPLE_MASK) {
 344                 OUT_PKT0(ring, REG_A3XX_RB_MSAA_CONTROL, 1);
 345                 OUT_RING(ring, A3XX_RB_MSAA_CONTROL_DISABLE |
 346                                 A3XX_RB_MSAA_CONTROL_SAMPLES(MSAA_ONE) |
 347                                 A3XX_RB_MSAA_CONTROL_SAMPLE_MASK(ctx->sample_mask));
 348         }
 349
 350         if (dirty & (FD_DIRTY_ZSA | FD_DIRTY_STENCIL_REF)) {
 351                 struct fd3_zsa_stateobj *zsa = fd3_zsa_stateobj(ctx->zsa);
 352                 struct pipe_stencil_ref *sr = &ctx->stencil_ref;
 353
 354                 fd3_emit_rbrc_draw_state(ring, zsa->rb_render_control);
 355
 356                 OUT_PKT0(ring, REG_A3XX_RB_DEPTH_CONTROL, 1);
 357                 OUT_RING(ring, zsa->rb_depth_control);
 358
 359                 OUT_PKT0(ring, REG_A3XX_RB_STENCIL_CONTROL, 1);
 360                 OUT_RING(ring, zsa->rb_stencil_control);
 361
 362                 OUT_PKT0(ring, REG_A3XX_RB_STENCILREFMASK, 2);
 363                 OUT_RING(ring, zsa->rb_stencilrefmask |
 364                                 A3XX_RB_STENCILREFMASK_STENCILREF(sr->ref_value[0]));
 365                 OUT_RING(ring, zsa->rb_stencilrefmask_bf |
 366                                 A3XX_RB_STENCILREFMASK_BF_STENCILREF(sr->ref_value[1]));
 367         }
 368
 369         if (dirty & FD_DIRTY_RASTERIZER) {
 370                 struct fd3_rasterizer_stateobj *rasterizer =
 371                                 fd3_rasterizer_stateobj(ctx->rasterizer);
 372
 373                 OUT_PKT0(ring, REG_A3XX_GRAS_SU_MODE_CONTROL, 1);
 374                 OUT_RING(ring, rasterizer->gras_su_mode_control);
 375
 376                 OUT_PKT0(ring, REG_A3XX_GRAS_SU_POINT_MINMAX, 2);
 377                 OUT_RING(ring, rasterizer->gras_su_point_minmax);
 378                 OUT_RING(ring, rasterizer->gras_su_point_size);
 379
 380                 OUT_PKT0(ring, REG_A3XX_GRAS_SU_POLY_OFFSET_SCALE, 2);
 381                 OUT_RING(ring, rasterizer->gras_su_poly_offset_scale);
 382                 OUT_RING(ring, rasterizer->gras_su_poly_offset_offset);
 383
 384                 OUT_PKT0(ring, REG_A3XX_GRAS_CL_CLIP_CNTL, 1);
 385                 OUT_RING(ring, rasterizer->gras_cl_clip_cntl);
 386         }
 387
 388         if (dirty & (FD_DIRTY_RASTERIZER | FD_DIRTY_PROG)) {
 389                 struct fd3_rasterizer_stateobj *rasterizer =
 390                                 fd3_rasterizer_stateobj(ctx->rasterizer);
 391                 struct fd3_shader_stateobj *fp = ctx->prog.fp;
 392                 uint32_t stride_in_vpc;
 393
 394                 stride_in_vpc = align(fp->total_in, 4) / 4;
 395                 if (stride_in_vpc > 0)
 396                         stride_in_vpc = MAX2(stride_in_vpc, 2);
 397
 398                 OUT_PKT0(ring, REG_A3XX_PC_PRIM_VTX_CNTL, 1);
 399                 OUT_RING(ring, rasterizer->pc_prim_vtx_cntl |
 400                                 A3XX_PC_PRIM_VTX_CNTL_STRIDE_IN_VPC(stride_in_vpc));
 401         }
 402
 403         if (dirty & FD_DIRTY_SCISSOR) {
 404                 OUT_PKT0(ring, REG_A3XX_GRAS_SC_WINDOW_SCISSOR_TL, 2);
 405                 OUT_RING(ring, A3XX_GRAS_SC_WINDOW_SCISSOR_TL_X(ctx->scissor.minx) |
 406                                 A3XX_GRAS_SC_WINDOW_SCISSOR_TL_Y(ctx->scissor.miny));
 407                 OUT_RING(ring, A3XX_GRAS_SC_WINDOW_SCISSOR_BR_X(ctx->scissor.maxx - 1) |
 408                                 A3XX_GRAS_SC_WINDOW_SCISSOR_BR_Y(ctx->scissor.maxy - 1));
 409
 410                 ctx->max_scissor.minx = MIN2(ctx->max_scissor.minx, ctx->scissor.minx);
 411                 ctx->max_scissor.miny = MIN2(ctx->max_scissor.miny, ctx->scissor.miny);
 412                 ctx->max_scissor.maxx = MAX2(ctx->max_scissor.maxx, ctx->scissor.maxx);
 413                 ctx->max_scissor.maxy = MAX2(ctx->max_scissor.maxy, ctx->scissor.maxy);
 414         }
 415
 416         if (dirty & FD_DIRTY_VIEWPORT) {
 417                 OUT_PKT0(ring, REG_A3XX_GRAS_CL_VPORT_XOFFSET, 6);
 418                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_XOFFSET(ctx->viewport.translate[0] - 0.5));
 419                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_XSCALE(ctx->viewport.scale[0]));
 420                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_YOFFSET(ctx->viewport.translate[1] - 0.5));
 421                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_YSCALE(ctx->viewport.scale[1]));
 422                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_ZOFFSET(ctx->viewport.translate[2]));
 423                 OUT_RING(ring, A3XX_GRAS_CL_VPORT_ZSCALE(ctx->viewport.scale[2]));
 424         }
 425
 426         if (dirty & FD_DIRTY_PROG)
 427                 fd3_program_emit(ring, &ctx->prog);
 428
 429         if (dirty & (FD_DIRTY_PROG | FD_DIRTY_CONSTBUF)) {
 430                 struct fd_program_stateobj *prog = &ctx->prog;
 431
 432                 emit_constants(ring,  SB_VERT_SHADER,
 433                                 &ctx->constbuf[PIPE_SHADER_VERTEX],
 434                                 (prog->dirty & FD_SHADER_DIRTY_VP) ? prog->vp : NULL);
 435                 emit_constants(ring, SB_FRAG_SHADER,
 436                                 &ctx->constbuf[PIPE_SHADER_FRAGMENT],
 437                                 (prog->dirty & FD_SHADER_DIRTY_FP) ? prog->fp : NULL);
 438         }
 439
 440         if (dirty & FD_DIRTY_BLEND) {
 441                 struct fd3_blend_stateobj *blend = fd3_blend_stateobj(ctx->blend);
 442                 uint32_t i;
 443
 444                 for (i = 0; i < ARRAY_SIZE(blend->rb_mrt); i++) {
 445                         OUT_PKT0(ring, REG_A3XX_RB_MRT_CONTROL(i), 1);
 446                         OUT_RING(ring, blend->rb_mrt[i].control);
 447
 448                         OUT_PKT0(ring, REG_A3XX_RB_MRT_BLEND_CONTROL(i), 1);
 449                         OUT_RING(ring, blend->rb_mrt[i].blend_control);
 450                 }
 451         }
 452
 453         if (dirty & FD_DIRTY_VERTTEX)
 454                 emit_textures(ring, SB_VERT_TEX, &ctx->verttex);
 455
 456         if (dirty & FD_DIRTY_FRAGTEX)
 457                 emit_textures(ring, SB_FRAG_TEX, &ctx->fragtex);
 458
 459         ctx->dirty &= ~dirty;
 460 }
 461
 462 /* emit setup at begin of new cmdstream buffer (don't rely on previous
 463  * state, there could have been a context switch between ioctls):
 464  */
 465 void
 466 fd3_emit_restore(struct fd_context *ctx)
 467 {
 468         struct fd3_context *fd3_ctx = fd3_context(ctx);
 469         struct fd_ringbuffer *ring = ctx->ring;
 470         int i;
 471
 472         OUT_PKT3(ring, CP_REG_RMW, 3);
 473         OUT_RING(ring, REG_A3XX_RBBM_CLOCK_CTL);
 474         OUT_RING(ring, 0xfffcffff);
 475         OUT_RING(ring, 0x00000000);
 476
 477         OUT_PKT3(ring, CP_INVALIDATE_STATE, 1);
 478         OUT_RING(ring, 0x00007fff);
 479
 480         OUT_PKT0(ring, REG_A3XX_SP_VS_PVT_MEM_CTRL_REG, 3);
 481         OUT_RING(ring, 0x08000001);                  /* SP_VS_PVT_MEM_CTRL_REG */
 482         OUT_RELOC(ring, fd3_ctx->vs_pvt_mem, 0, 0);  /* SP_VS_PVT_MEM_ADDR_REG */
 483         OUT_RING(ring, 0x00000000);                  /* SP_VS_PVT_MEM_SIZE_REG */
 484
 485         OUT_PKT0(ring, REG_A3XX_SP_FS_PVT_MEM_CTRL_REG, 3);
 486         OUT_RING(ring, 0x08000001);                  /* SP_FS_PVT_MEM_CTRL_REG */
 487         OUT_RELOC(ring, fd3_ctx->fs_pvt_mem, 0, 0);  /* SP_FS_PVT_MEM_ADDR_REG */
 488         OUT_RING(ring, 0x00000000);                  /* SP_FS_PVT_MEM_SIZE_REG */
 489
 490         OUT_PKT0(ring, REG_A3XX_PC_VERTEX_REUSE_BLOCK_CNTL, 1);
 491         OUT_RING(ring, 0x0000000b);                  /* PC_VERTEX_REUSE_BLOCK_CNTL */
 492
 493         OUT_PKT0(ring, REG_A3XX_GRAS_SC_CONTROL, 1);
 494         OUT_RING(ring, A3XX_GRAS_SC_CONTROL_RENDER_MODE(RB_RENDERING_PASS) |
 495                         A3XX_GRAS_SC_CONTROL_MSAA_SAMPLES(MSAA_ONE) |
 496                         A3XX_GRAS_SC_CONTROL_RASTER_MODE(0));
 497
 498         OUT_PKT0(ring, REG_A3XX_RB_MSAA_CONTROL, 2);
 499         OUT_RING(ring, A3XX_RB_MSAA_CONTROL_DISABLE |
 500                         A3XX_RB_MSAA_CONTROL_SAMPLES(MSAA_ONE) |
 501                         A3XX_RB_MSAA_CONTROL_SAMPLE_MASK(0xffff));
 502         OUT_RING(ring, 0x00000000);        /* UNKNOWN_20C3 */
 503
 504         OUT_PKT0(ring, REG_A3XX_GRAS_CL_GB_CLIP_ADJ, 1);
 505         OUT_RING(ring, A3XX_GRAS_CL_GB_CLIP_ADJ_HORZ(0) |
 506                         A3XX_GRAS_CL_GB_CLIP_ADJ_VERT(0));
 507
 508         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0C81, 1);
 509         OUT_RING(ring, 0x00000001);        /* UNKNOWN_0C81 */
 510
 511         OUT_PKT0(ring, REG_A3XX_TPL1_TP_VS_TEX_OFFSET, 1);
 512         OUT_RING(ring, A3XX_TPL1_TP_VS_TEX_OFFSET_SAMPLEROFFSET(VERT_TEX_OFF) |
 513                         A3XX_TPL1_TP_VS_TEX_OFFSET_MEMOBJOFFSET(VERT_TEX_OFF) |
 514                         A3XX_TPL1_TP_VS_TEX_OFFSET_BASETABLEPTR(BASETABLE_SZ * VERT_TEX_OFF));
 515
 516         OUT_PKT0(ring, REG_A3XX_TPL1_TP_FS_TEX_OFFSET, 1);
 517         OUT_RING(ring, A3XX_TPL1_TP_FS_TEX_OFFSET_SAMPLEROFFSET(FRAG_TEX_OFF) |
 518                         A3XX_TPL1_TP_FS_TEX_OFFSET_MEMOBJOFFSET(FRAG_TEX_OFF) |
 519                         A3XX_TPL1_TP_FS_TEX_OFFSET_BASETABLEPTR(BASETABLE_SZ * FRAG_TEX_OFF));
 520
 521         OUT_PKT0(ring, REG_A3XX_VPC_VARY_CYLWRAP_ENABLE_0, 2);
 522         OUT_RING(ring, 0x00000000);        /* VPC_VARY_CYLWRAP_ENABLE_0 */
 523         OUT_RING(ring, 0x00000000);        /* VPC_VARY_CYLWRAP_ENABLE_1 */
 524
 525         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0E43, 1);
 526         OUT_RING(ring, 0x00000001);        /* UNKNOWN_0E43 */
 527
 528         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0F03, 1);
 529         OUT_RING(ring, 0x00000001);        /* UNKNOWN_0F03 */
 530
 531         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0EE0, 1);
 532         OUT_RING(ring, 0x00000003);        /* UNKNOWN_0EE0 */
 533
 534         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0C3D, 1);
 535         OUT_RING(ring, 0x00000001);        /* UNKNOWN_0C3D */
 536
 537         OUT_PKT0(ring, REG_A3XX_UNKNOWN_0E00, 1);
 538         OUT_RING(ring, 0x00000000);        /* UNKNOWN_0E00 */
 539
 540         OUT_PKT0(ring, REG_A3XX_HLSQ_CONST_VSPRESV_RANGE_REG, 2);
 541         OUT_RING(ring, A3XX_HLSQ_CONST_VSPRESV_RANGE_REG_STARTENTRY(0) |
 542                         A3XX_HLSQ_CONST_VSPRESV_RANGE_REG_ENDENTRY(0));
 543         OUT_RING(ring, A3XX_HLSQ_CONST_FSPRESV_RANGE_REG_STARTENTRY(0) |
 544                         A3XX_HLSQ_CONST_FSPRESV_RANGE_REG_ENDENTRY(0));
 545
 546         OUT_PKT0(ring, REG_A3XX_UCHE_CACHE_MODE_CONTROL_REG, 1);
 547         OUT_RING(ring, 0x00000001);        /* UCHE_CACHE_MODE_CONTROL_REG */
 548
 549         OUT_PKT0(ring, REG_A3XX_VSC_SIZE_ADDRESS, 1);
 550         OUT_RELOC(ring, fd3_ctx->vsc_size_mem, 0, 0); /* VSC_SIZE_ADDRESS */
 551
 552         OUT_PKT0(ring, REG_A3XX_GRAS_CL_CLIP_CNTL, 1);
 553         OUT_RING(ring, 0x00000000);                  /* GRAS_CL_CLIP_CNTL */
 554
 555         OUT_PKT0(ring, REG_A3XX_GRAS_SU_POINT_MINMAX, 2);
 556         OUT_RING(ring, 0xffc00010);        /* GRAS_SU_POINT_MINMAX */
 557         OUT_RING(ring, 0x00000008);        /* GRAS_SU_POINT_SIZE */
 558
 559         OUT_PKT0(ring, REG_A3XX_PC_RESTART_INDEX, 1);
 560         OUT_RING(ring, 0xffffffff);        /* PC_RESTART_INDEX */
 561
 562         OUT_PKT0(ring, REG_A3XX_PA_SC_WINDOW_OFFSET, 1);
 563         OUT_RING(ring, A3XX_PA_SC_WINDOW_OFFSET_X(0) |
 564                         A3XX_PA_SC_WINDOW_OFFSET_Y(0));
 565
 566         OUT_PKT0(ring, REG_A3XX_RB_BLEND_RED, 4);
 567         OUT_RING(ring, 0x00000000);        /* RB_BLEND_RED */
 568         OUT_RING(ring, 0x00000000);        /* RB_BLEND_GREEN */
 569         OUT_RING(ring, 0x00000000);        /* RB_BLEND_BLUE */
 570         OUT_RING(ring, 0x3c0000ff);        /* RB_BLEND_ALPHA */
 571
 572         for (i = 0; i < 6; i++) {
 573                 OUT_PKT0(ring, REG_A3XX_GRAS_CL_USER_PLANE(i), 4);
 574                 OUT_RING(ring, 0x00000000);    /* GRAS_CL_USER_PLANE[i].X */
 575                 OUT_RING(ring, 0x00000000);    /* GRAS_CL_USER_PLANE[i].Y */
 576                 OUT_RING(ring, 0x00000000);    /* GRAS_CL_USER_PLANE[i].Z */
 577                 OUT_RING(ring, 0x00000000);    /* GRAS_CL_USER_PLANE[i].W */
 578         }
 579
 580         emit_cache_flush(ring);
 581 }