src/gallium/drivers/freedreno/a2xx/fd2_draw.c

   1 /*
   2  * Copyright (C) 2012-2013 Rob Clark <robclark@freedesktop.org>
   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 (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "pipe/p_state.h"
  28 #include "util/u_string.h"
  29 #include "util/u_memory.h"
  30 #include "util/u_prim.h"
  31
  32 #include "freedreno_state.h"
  33 #include "freedreno_resource.h"
  34
  35 #include "fd2_draw.h"
  36 #include "fd2_context.h"
  37 #include "fd2_emit.h"
  38 #include "fd2_program.h"
  39 #include "fd2_util.h"
  40 #include "fd2_zsa.h"
  41
  42
  43 static void
  44 emit_cacheflush(struct fd_ringbuffer *ring)
  45 {
  46         unsigned i;
  47
  48         for (i = 0; i < 12; i++) {
  49                 OUT_PKT3(ring, CP_EVENT_WRITE, 1);
  50                 OUT_RING(ring, CACHE_FLUSH);
  51         }
  52 }
  53
  54 static void
  55 emit_vertexbufs(struct fd_context *ctx)
  56 {
  57         struct fd_vertex_stateobj *vtx = ctx->vtx.vtx;
  58         struct fd_vertexbuf_stateobj *vertexbuf = &ctx->vtx.vertexbuf;
  59         struct fd2_vertex_buf bufs[PIPE_MAX_ATTRIBS];
  60         unsigned i;
  61
  62         if (!vtx->num_elements)
  63                 return;
  64
  65         for (i = 0; i < vtx->num_elements; i++) {
  66                 struct pipe_vertex_element *elem = &vtx->pipe[i];
  67                 struct pipe_vertex_buffer *vb =
  68                                 &vertexbuf->vb[elem->vertex_buffer_index];
  69                 bufs[i].offset = vb->buffer_offset;
  70                 bufs[i].size = fd_bo_size(fd_resource(vb->buffer.resource)->bo);
  71                 bufs[i].prsc = vb->buffer.resource;
  72         }
  73
  74         // NOTE I believe the 0x78 (or 0x9c in solid_vp) relates to the
  75         // CONST(20,0) (or CONST(26,0) in soliv_vp)
  76
  77         fd2_emit_vertex_bufs(ctx->batch->draw, 0x78, bufs, vtx->num_elements);
  78         fd2_emit_vertex_bufs(ctx->batch->binning, 0x78, bufs, vtx->num_elements);
  79 }
  80
  81 static void
  82 draw_impl(struct fd_context *ctx, const struct pipe_draw_info *info,
  83                    struct fd_ringbuffer *ring, unsigned index_offset, bool binning)
  84 {
  85         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
  86         OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
  87         OUT_RING(ring, info->index_size ? 0 : info->start);
  88
  89         OUT_PKT0(ring, REG_A2XX_TC_CNTL_STATUS, 1);
  90         OUT_RING(ring, A2XX_TC_CNTL_STATUS_L2_INVALIDATE);
  91
  92         if (is_a20x(ctx->screen)) {
  93                 /* wait for DMA to finish and
  94                  * dummy draw one triangle with indexes 0,0,0.
  95                  * with PRE_FETCH_CULL_ENABLE | GRP_CULL_ENABLE.
  96                  *
  97                  * this workaround is for a HW bug related to DMA alignment:
  98                  * it is necessary for indexed draws and possibly also
  99                  * draws that read binning data
 100                  */
 101                 OUT_PKT3(ring, CP_WAIT_REG_EQ, 4);
 102                 OUT_RING(ring, 0x000005d0); /* RBBM_STATUS */
 103                 OUT_RING(ring, 0x00000000);
 104                 OUT_RING(ring, 0x00001000); /* bit: 12: VGT_BUSY_NO_DMA */
 105                 OUT_RING(ring, 0x00000001);
 106
 107                 OUT_PKT3(ring, CP_DRAW_INDX_BIN, 6);
 108                 OUT_RING(ring, 0x00000000);
 109                 OUT_RING(ring, 0x0003c004);
 110                 OUT_RING(ring, 0x00000000);
 111                 OUT_RING(ring, 0x00000003);
 112                 OUT_RELOC(ring, fd_resource(fd2_context(ctx)->solid_vertexbuf)->bo, 64, 0, 0);
 113                 OUT_RING(ring, 0x00000006);
 114         } else {
 115                 OUT_WFI (ring);
 116
 117                 OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 118                 OUT_RING(ring, CP_REG(REG_A2XX_VGT_MAX_VTX_INDX));
 119                 OUT_RING(ring, info->max_index);        /* VGT_MAX_VTX_INDX */
 120                 OUT_RING(ring, info->min_index);        /* VGT_MIN_VTX_INDX */
 121         }
 122
 123         /* binning shader will take offset from C64 */
 124         if (binning && is_a20x(ctx->screen)) {
 125                 OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 126                 OUT_RING(ring, 0x00000180);
 127                 OUT_RING(ring, fui(ctx->batch->num_vertices));
 128                 OUT_RING(ring, fui(0.0f));
 129                 OUT_RING(ring, fui(0.0f));
 130                 OUT_RING(ring, fui(0.0f));
 131         }
 132
 133         enum pc_di_vis_cull_mode vismode = USE_VISIBILITY;
 134         if (binning || info->mode == PIPE_PRIM_POINTS)
 135                 vismode = IGNORE_VISIBILITY;
 136
 137         fd_draw_emit(ctx->batch, ring, ctx->primtypes[info->mode],
 138                                  vismode, info, index_offset);
 139
 140         if (is_a20x(ctx->screen)) {
 141                 /* not sure why this is required, but it fixes some hangs */
 142                 OUT_WFI(ring);
 143         } else {
 144                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 145                 OUT_RING(ring, CP_REG(REG_A2XX_UNKNOWN_2010));
 146                 OUT_RING(ring, 0x00000000);
 147         }
 148
 149         emit_cacheflush(ring);
 150 }
 151
 152
 153 static bool
 154 fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *pinfo,
 155                          unsigned index_offset)
 156 {
 157         if (!ctx->prog.fp || !ctx->prog.vp)
 158                 return false;
 159
 160         if (ctx->dirty & FD_DIRTY_VTXBUF)
 161                 emit_vertexbufs(ctx);
 162
 163         if (fd_binning_enabled)
 164                 fd2_emit_state_binning(ctx, ctx->dirty);
 165
 166         fd2_emit_state(ctx, ctx->dirty);
 167
 168         /* a2xx can draw only 65535 vertices at once
 169          * on a22x the field in the draw command is 32bits but seems limited too
 170          * using a limit of 32k because it fixes an unexplained hang
 171          * 32766 works for all primitives (multiple of 2 and 3)
 172          */
 173         if (pinfo->count > 32766) {
 174                 static const uint16_t step_tbl[PIPE_PRIM_MAX] = {
 175                         [0 ... PIPE_PRIM_MAX - 1]  = 32766,
 176                         [PIPE_PRIM_LINE_STRIP]     = 32765,
 177                         [PIPE_PRIM_TRIANGLE_STRIP] = 32764,
 178
 179                         /* needs more work */
 180                         [PIPE_PRIM_TRIANGLE_FAN]   = 0,
 181                         [PIPE_PRIM_LINE_LOOP]      = 0,
 182                 };
 183
 184                 struct pipe_draw_info info = *pinfo;
 185                 unsigned count = info.count;
 186                 unsigned step = step_tbl[info.mode];
 187                 unsigned num_vertices = ctx->batch->num_vertices;
 188
 189                 if (!step)
 190                         return false;
 191
 192                 for (; count + step > 32766; count -= step) {
 193                         info.count = MIN2(count, 32766);
 194                         draw_impl(ctx, &info, ctx->batch->draw, index_offset, false);
 195                         draw_impl(ctx, &info, ctx->batch->binning, index_offset, true);
 196                         info.start += step;
 197                         ctx->batch->num_vertices += step;
 198                 }
 199                 /* changing this value is a hack, restore it */
 200                 ctx->batch->num_vertices = num_vertices;
 201         } else {
 202                 draw_impl(ctx, pinfo, ctx->batch->draw, index_offset, false);
 203                 draw_impl(ctx, pinfo, ctx->batch->binning, index_offset, true);
 204         }
 205
 206         fd_context_all_clean(ctx);
 207
 208         return true;
 209 }
 210
 211 static void
 212 clear_state(struct fd_batch *batch, struct fd_ringbuffer *ring,
 213         unsigned buffers, bool fast_clear)
 214 {
 215         struct fd_context *ctx = batch->ctx;
 216         struct fd2_context *fd2_ctx = fd2_context(ctx);
 217         uint32_t reg;
 218
 219         fd2_emit_vertex_bufs(ring, 0x9c, (struct fd2_vertex_buf[]) {
 220                         { .prsc = fd2_ctx->solid_vertexbuf, .size = 36 },
 221                 }, 1);
 222
 223         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 224         OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
 225         OUT_RING(ring, 0);
 226
 227         fd2_program_emit(ctx, ring, &ctx->solid_prog);
 228
 229         OUT_PKT0(ring, REG_A2XX_TC_CNTL_STATUS, 1);
 230         OUT_RING(ring, A2XX_TC_CNTL_STATUS_L2_INVALIDATE);
 231
 232         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 233                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 234                 OUT_RING(ring, CP_REG(REG_A2XX_RB_DEPTHCONTROL));
 235                 reg = 0;
 236                 if (buffers & PIPE_CLEAR_DEPTH) {
 237                         reg |= A2XX_RB_DEPTHCONTROL_ZFUNC(FUNC_ALWAYS) |
 238                                 A2XX_RB_DEPTHCONTROL_Z_ENABLE |
 239                                 A2XX_RB_DEPTHCONTROL_Z_WRITE_ENABLE |
 240                                 A2XX_RB_DEPTHCONTROL_EARLY_Z_ENABLE;
 241                 }
 242                 if (buffers & PIPE_CLEAR_STENCIL) {
 243                         reg |= A2XX_RB_DEPTHCONTROL_STENCILFUNC(FUNC_ALWAYS) |
 244                                         A2XX_RB_DEPTHCONTROL_STENCIL_ENABLE |
 245                                         A2XX_RB_DEPTHCONTROL_STENCILZPASS(STENCIL_REPLACE);
 246                 }
 247                 OUT_RING(ring, reg);
 248         }
 249
 250         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 251         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLORCONTROL));
 252         OUT_RING(ring, A2XX_RB_COLORCONTROL_ALPHA_FUNC(FUNC_ALWAYS) |
 253                         A2XX_RB_COLORCONTROL_BLEND_DISABLE |
 254                         A2XX_RB_COLORCONTROL_ROP_CODE(12) |
 255                         A2XX_RB_COLORCONTROL_DITHER_MODE(DITHER_DISABLE) |
 256                         A2XX_RB_COLORCONTROL_DITHER_TYPE(DITHER_PIXEL));
 257
 258         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 259         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_CLIP_CNTL));
 260         OUT_RING(ring, 0x00000000);        /* PA_CL_CLIP_CNTL */
 261         OUT_RING(ring, A2XX_PA_SU_SC_MODE_CNTL_PROVOKING_VTX_LAST |  /* PA_SU_SC_MODE_CNTL */
 262                         A2XX_PA_SU_SC_MODE_CNTL_FRONT_PTYPE(PC_DRAW_TRIANGLES) |
 263                         A2XX_PA_SU_SC_MODE_CNTL_BACK_PTYPE(PC_DRAW_TRIANGLES) |
 264                         (fast_clear ? A2XX_PA_SU_SC_MODE_CNTL_MSAA_ENABLE : 0));
 265
 266         if (fast_clear) {
 267                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 268                 OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_AA_CONFIG));
 269                 OUT_RING(ring, A2XX_PA_SC_AA_CONFIG_MSAA_NUM_SAMPLES(3));
 270         }
 271
 272         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 273         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_AA_MASK));
 274         OUT_RING(ring, 0x0000ffff);
 275
 276         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 277         OUT_RING(ring, CP_REG(REG_A2XX_RB_COLOR_MASK));
 278         if (buffers & PIPE_CLEAR_COLOR) {
 279                 OUT_RING(ring, A2XX_RB_COLOR_MASK_WRITE_RED |
 280                                 A2XX_RB_COLOR_MASK_WRITE_GREEN |
 281                                 A2XX_RB_COLOR_MASK_WRITE_BLUE |
 282                                 A2XX_RB_COLOR_MASK_WRITE_ALPHA);
 283         } else {
 284                 OUT_RING(ring, 0x0);
 285         }
 286
 287         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 288         OUT_RING(ring, CP_REG(REG_A2XX_RB_BLEND_CONTROL));
 289         OUT_RING(ring, 0);
 290
 291         if (is_a20x(batch->ctx->screen))
 292                 return;
 293
 294         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 295         OUT_RING(ring, CP_REG(REG_A2XX_VGT_MAX_VTX_INDX));
 296         OUT_RING(ring, 3);                 /* VGT_MAX_VTX_INDX */
 297         OUT_RING(ring, 0);                 /* VGT_MIN_VTX_INDX */
 298
 299         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 300         OUT_RING(ring, CP_REG(REG_A2XX_RB_STENCILREFMASK_BF));
 301         OUT_RING(ring, 0xff000000 | A2XX_RB_STENCILREFMASK_BF_STENCILWRITEMASK(0xff));
 302         OUT_RING(ring, 0xff000000 | A2XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
 303
 304         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 305         OUT_RING(ring, CP_REG(REG_A2XX_A220_RB_LRZ_VSC_CONTROL));
 306         OUT_RING(ring, 0x00000084);
 307
 308         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 309         OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
 310         OUT_RING(ring, 0x0000028f);
 311 }
 312
 313 static void
 314 clear_state_restore(struct fd_context *ctx, struct fd_ringbuffer *ring)
 315 {
 316         if (is_a20x(ctx->screen))
 317                 return;
 318
 319         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 320         OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
 321         OUT_RING(ring, 0x00000000);
 322
 323         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 324         OUT_RING(ring, CP_REG(REG_A2XX_A220_RB_LRZ_VSC_CONTROL));
 325         OUT_RING(ring, 0x00000000);
 326
 327         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 328         OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
 329         OUT_RING(ring, 0x0000003b);
 330 }
 331
 332 static void
 333 clear_fast(struct fd_batch *batch, struct fd_ringbuffer *ring,
 334         uint32_t color_clear, uint32_t depth_clear, unsigned patch_type)
 335 {
 336         BEGIN_RING(ring, 8); /* preallocate next 2 packets (for patching) */
 337
 338         /* zero values are patched in */
 339         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 340         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_SCREEN_SCISSOR_BR));
 341         OUT_RINGP(ring, patch_type, &batch->gmem_patches);
 342
 343         OUT_PKT3(ring, CP_SET_CONSTANT, 4);
 344         OUT_RING(ring, CP_REG(REG_A2XX_RB_SURFACE_INFO));
 345         OUT_RING(ring, 0x8000 | 32);
 346         OUT_RING(ring, 0);
 347         OUT_RING(ring, 0);
 348
 349         /* set fill values */
 350         if (!is_a20x(batch->ctx->screen)) {
 351                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 352                 OUT_RING(ring, CP_REG(REG_A2XX_CLEAR_COLOR));
 353                 OUT_RING(ring, color_clear);
 354
 355                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 356                 OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
 357                 OUT_RING(ring, A2XX_RB_COPY_CONTROL_DEPTH_CLEAR_ENABLE |
 358                         A2XX_RB_COPY_CONTROL_CLEAR_MASK(0xf));
 359
 360                 OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 361                 OUT_RING(ring, CP_REG(REG_A2XX_RB_DEPTH_CLEAR));
 362                 OUT_RING(ring, depth_clear);
 363         } else {
 364                 const float sc = 1.0f / 255.0f;
 365
 366                 OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 367                 OUT_RING(ring, 0x00000480);
 368                 OUT_RING(ring, fui((float) (color_clear >>  0 & 0xff) * sc));
 369                 OUT_RING(ring, fui((float) (color_clear >>  8 & 0xff) * sc));
 370                 OUT_RING(ring, fui((float) (color_clear >> 16 & 0xff) * sc));
 371                 OUT_RING(ring, fui((float) (color_clear >> 24 & 0xff) * sc));
 372
 373                 // XXX if using float the rounding error breaks it..
 374                 float depth = ((double) (depth_clear >> 8)) * (1.0/(double) 0xffffff);
 375                 assert((unsigned) (((double) depth * (double) 0xffffff)) ==
 376                         (depth_clear >> 8));
 377
 378                 OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 379                 OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VPORT_ZSCALE));
 380                 OUT_RING(ring, fui(0.0f));
 381                 OUT_RING(ring, fui(depth));
 382
 383                 OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 384                 OUT_RING(ring, CP_REG(REG_A2XX_RB_STENCILREFMASK_BF));
 385                 OUT_RING(ring, 0xff000000 |
 386                         A2XX_RB_STENCILREFMASK_BF_STENCILREF(depth_clear & 0xff) |
 387                         A2XX_RB_STENCILREFMASK_BF_STENCILWRITEMASK(0xff));
 388                 OUT_RING(ring, 0xff000000 |
 389                         A2XX_RB_STENCILREFMASK_STENCILREF(depth_clear & 0xff) |
 390                         A2XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
 391         }
 392
 393         fd_draw(batch, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
 394                         DI_SRC_SEL_AUTO_INDEX, 3, 0, INDEX_SIZE_IGN, 0, 0, NULL);
 395 }
 396
 397 static bool
 398 fd2_clear_fast(struct fd_context *ctx, unsigned buffers,
 399                 const union pipe_color_union *color, double depth, unsigned stencil)
 400 {
 401         /* using 4x MSAA allows clearing ~2x faster
 402          * then we can use higher bpp clearing to clear lower bpp
 403          * 1 "pixel" can clear 64 bits (rgba8+depth24+stencil8)
 404          * note: its possible to clear with 32_32_32_32 format but its not faster
 405          * note: fast clear doesn't work with sysmem rendering
 406          * (sysmem rendering is disabled when clear is used)
 407          *
 408          * we only have 16-bit / 32-bit color formats
 409          * and 16-bit / 32-bit depth formats
 410          * so there are only a few possible combinations
 411          *
 412          * if the bpp of the color/depth doesn't match
 413          * we clear with depth/color individually
 414          */
 415         struct fd2_context *fd2_ctx = fd2_context(ctx);
 416         struct fd_batch *batch = ctx->batch;
 417         struct fd_ringbuffer *ring = batch->draw;
 418         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 419         uint32_t color_clear = 0, depth_clear = 0;
 420         enum pipe_format format = pipe_surface_format(pfb->cbufs[0]);
 421         int depth_size = -1; /* -1: no clear, 0: clear 16-bit, 1: clear 32-bit */
 422         int color_size = -1;
 423
 424         /* TODO: need to test performance on a22x */
 425         if (!is_a20x(ctx->screen))
 426                 return false;
 427
 428         if (buffers & PIPE_CLEAR_COLOR)
 429                 color_size = util_format_get_blocksizebits(format) == 32;
 430
 431         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL))
 432                 depth_size = fd_pipe2depth(pfb->zsbuf->format) == DEPTHX_24_8;
 433
 434         assert(color_size >= 0 || depth_size >= 0);
 435
 436         /* when clearing 24_8, depth/stencil must be both cleared
 437          * TODO: if buffer isn't attached we can clear it anyway
 438          */
 439         if (depth_size == 1 && !(buffers & PIPE_CLEAR_STENCIL) != !(buffers & PIPE_CLEAR_DEPTH))
 440                 return false;
 441
 442         if (color_size == 0) {
 443                 color_clear = pack_rgba(format, color->f);
 444                 color_clear = (color_clear << 16) | (color_clear & 0xffff);
 445         } else if (color_size == 1) {
 446                 color_clear = pack_rgba(format, color->f);
 447         }
 448
 449         if (depth_size == 0) {
 450                 depth_clear = (uint32_t)(0xffff * depth);
 451                 depth_clear |= depth_clear << 16;
 452         } else if (depth_size == 1) {
 453                 depth_clear = (((uint32_t)(0xffffff * depth)) << 8);
 454                 depth_clear |= (stencil & 0xff);
 455         }
 456
 457         /* disable "window" scissor.. */
 458         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 459         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_SCISSOR_TL));
 460         OUT_RING(ring, xy2d(0, 0));
 461         OUT_RING(ring, xy2d(0x7fff, 0x7fff));
 462
 463         /* make sure we fill all "pixels" (in SCREEN_SCISSOR) */
 464         OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 465         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VPORT_XSCALE));
 466         OUT_RING(ring, fui(4096.0));
 467         OUT_RING(ring, fui(4096.0));
 468         OUT_RING(ring, fui(4096.0));
 469         OUT_RING(ring, fui(4096.0));
 470
 471         clear_state(batch, ring, ~0u, true);
 472
 473         if (color_size >= 0 && depth_size != color_size)
 474                 clear_fast(batch, ring, color_clear, color_clear, GMEM_PATCH_FASTCLEAR_COLOR);
 475
 476         if (depth_size >= 0 && depth_size != color_size)
 477                 clear_fast(batch, ring, depth_clear, depth_clear, GMEM_PATCH_FASTCLEAR_DEPTH);
 478
 479         if (depth_size == color_size)
 480                 clear_fast(batch, ring, color_clear, depth_clear, GMEM_PATCH_FASTCLEAR_COLOR_DEPTH);
 481
 482         clear_state_restore(ctx, ring);
 483
 484         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 485         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_AA_CONFIG));
 486         OUT_RING(ring, 0);
 487
 488         /* can't patch in SCREEN_SCISSOR_BR as it can be different for each tile.
 489          * MEM_WRITE the value in tile_renderprep, and use CP_LOAD_CONSTANT_CONTEXT
 490          * the value is read from byte offset 60 in the given bo
 491          */
 492         OUT_PKT3(ring, CP_LOAD_CONSTANT_CONTEXT, 3);
 493         OUT_RELOC(ring, fd_resource(fd2_ctx->solid_vertexbuf)->bo, 0, 0, 0);
 494         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_SCREEN_SCISSOR_BR));
 495         OUT_RING(ring, 1);
 496
 497         OUT_PKT3(ring, CP_SET_CONSTANT, 4);
 498         OUT_RING(ring, CP_REG(REG_A2XX_RB_SURFACE_INFO));
 499         OUT_RINGP(ring, GMEM_PATCH_RESTORE_INFO, &batch->gmem_patches);
 500         OUT_RING(ring, 0);
 501         OUT_RING(ring, 0);
 502         return true;
 503 }
 504
 505 static bool
 506 fd2_clear(struct fd_context *ctx, unsigned buffers,
 507                 const union pipe_color_union *color, double depth, unsigned stencil)
 508 {
 509         struct fd_ringbuffer *ring = ctx->batch->draw;
 510         struct pipe_framebuffer_state *fb = &ctx->batch->framebuffer;
 511
 512         if (fd2_clear_fast(ctx, buffers, color, depth, stencil))
 513                 goto dirty;
 514
 515         /* set clear value */
 516         if (is_a20x(ctx->screen)) {
 517                 if (buffers & PIPE_CLEAR_COLOR) {
 518                         /* C0 used by fragment shader */
 519                         OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 520                         OUT_RING(ring, 0x00000480);
 521                         OUT_RING(ring, color->ui[0]);
 522                         OUT_RING(ring, color->ui[1]);
 523                         OUT_RING(ring, color->ui[2]);
 524                         OUT_RING(ring, color->ui[3]);
 525                 }
 526
 527                 if (buffers & PIPE_CLEAR_DEPTH) {
 528                         /* use viewport to set depth value */
 529                         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 530                         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VPORT_ZSCALE));
 531                         OUT_RING(ring, fui(0.0f));
 532                         OUT_RING(ring, fui(depth));
 533                 }
 534
 535                 if (buffers & PIPE_CLEAR_STENCIL) {
 536                         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 537                         OUT_RING(ring, CP_REG(REG_A2XX_RB_STENCILREFMASK_BF));
 538                         OUT_RING(ring, 0xff000000 |
 539                                 A2XX_RB_STENCILREFMASK_BF_STENCILREF(stencil) |
 540                                 A2XX_RB_STENCILREFMASK_BF_STENCILWRITEMASK(0xff));
 541                         OUT_RING(ring, 0xff000000 |
 542                                 A2XX_RB_STENCILREFMASK_STENCILREF(stencil) |
 543                                 A2XX_RB_STENCILREFMASK_STENCILWRITEMASK(0xff));
 544                 }
 545         } else {
 546                 if (buffers & PIPE_CLEAR_COLOR) {
 547                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 548                         OUT_RING(ring, CP_REG(REG_A2XX_CLEAR_COLOR));
 549                         OUT_RING(ring, pack_rgba(PIPE_FORMAT_R8G8B8A8_UNORM, color->f));
 550                 }
 551
 552                 if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 553                         uint32_t clear_mask, depth_clear;
 554                         switch (fd_pipe2depth(fb->zsbuf->format)) {
 555                         case DEPTHX_24_8:
 556                                 clear_mask = ((buffers & PIPE_CLEAR_DEPTH) ? 0xe : 0) |
 557                                         ((buffers & PIPE_CLEAR_STENCIL) ? 0x1 : 0);
 558                                 depth_clear = (((uint32_t)(0xffffff * depth)) << 8) |
 559                                         (stencil & 0xff);
 560                                 break;
 561                         case DEPTHX_16:
 562                                 clear_mask = 0xf;
 563                                 depth_clear = (uint32_t)(0xffffffff * depth);
 564                                 break;
 565                         default:
 566                                 unreachable("invalid depth");
 567                                 break;
 568                         }
 569
 570                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 571                         OUT_RING(ring, CP_REG(REG_A2XX_RB_COPY_CONTROL));
 572                         OUT_RING(ring, A2XX_RB_COPY_CONTROL_DEPTH_CLEAR_ENABLE |
 573                                 A2XX_RB_COPY_CONTROL_CLEAR_MASK(clear_mask));
 574
 575                         OUT_PKT3(ring, CP_SET_CONSTANT, 2);
 576                         OUT_RING(ring, CP_REG(REG_A2XX_RB_DEPTH_CLEAR));
 577                         OUT_RING(ring, depth_clear);
 578                 }
 579         }
 580
 581         /* scissor state */
 582         OUT_PKT3(ring, CP_SET_CONSTANT, 3);
 583         OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_SCISSOR_TL));
 584         OUT_RING(ring, xy2d(0, 0));
 585         OUT_RING(ring, xy2d(fb->width, fb->height));
 586
 587         /* viewport state */
 588         OUT_PKT3(ring, CP_SET_CONSTANT, 5);
 589         OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_VPORT_XSCALE));
 590         OUT_RING(ring, fui((float) fb->width / 2.0));
 591         OUT_RING(ring, fui((float) fb->width / 2.0));
 592         OUT_RING(ring, fui((float) fb->height / 2.0));
 593         OUT_RING(ring, fui((float) fb->height / 2.0));
 594
 595         /* common state */
 596         clear_state(ctx->batch, ring, buffers, false);
 597
 598         fd_draw(ctx->batch, ring, DI_PT_RECTLIST, IGNORE_VISIBILITY,
 599                         DI_SRC_SEL_AUTO_INDEX, 3, 0, INDEX_SIZE_IGN, 0, 0, NULL);
 600
 601         clear_state_restore(ctx, ring);
 602
 603 dirty:
 604         ctx->dirty |= FD_DIRTY_ZSA |
 605                         FD_DIRTY_VIEWPORT |
 606                         FD_DIRTY_RASTERIZER |
 607                         FD_DIRTY_SAMPLE_MASK |
 608                         FD_DIRTY_PROG |
 609                         FD_DIRTY_CONST |
 610                         FD_DIRTY_BLEND |
 611                         FD_DIRTY_FRAMEBUFFER |
 612                         FD_DIRTY_SCISSOR;
 613
 614         ctx->dirty_shader[PIPE_SHADER_VERTEX]   |= FD_DIRTY_SHADER_PROG;
 615         ctx->dirty_shader[PIPE_SHADER_FRAGMENT] |= FD_DIRTY_SHADER_PROG | FD_DIRTY_SHADER_CONST;
 616
 617         return true;
 618 }
 619
 620 void
 621 fd2_draw_init(struct pipe_context *pctx)
 622 {
 623         struct fd_context *ctx = fd_context(pctx);
 624         ctx->draw_vbo = fd2_draw_vbo;
 625         ctx->clear = fd2_clear;
 626 }