Subversion Repositories Kolibri OS

Rev

Go to most recent revision | Blame | Last modification | View Log | RSS feed

  1. /*
  2.  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
  3.  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
  4.  *
  5.  * Permission is hereby granted, free of charge, to any person obtaining a
  6.  * copy of this software and associated documentation files (the "Software"),
  7.  * to deal in the Software without restriction, including without limitation
  8.  * on the rights to use, copy, modify, merge, publish, distribute, sub
  9.  * license, and/or sell copies of the Software, and to permit persons to whom
  10.  * the Software is furnished to do so, subject to the following conditions:
  11.  *
  12.  * The above copyright notice and this permission notice (including the next
  13.  * paragraph) shall be included in all copies or substantial portions of the
  14.  * Software.
  15.  *
  16.  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17.  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18.  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19.  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20.  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21.  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22.  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  23.  
  24. #include "draw/draw_context.h"
  25.  
  26. #include "util/u_framebuffer.h"
  27. #include "util/u_half.h"
  28. #include "util/u_helpers.h"
  29. #include "util/u_math.h"
  30. #include "util/u_mm.h"
  31. #include "util/u_memory.h"
  32. #include "util/u_pack_color.h"
  33. #include "util/u_transfer.h"
  34.  
  35. #include "tgsi/tgsi_parse.h"
  36.  
  37. #include "pipe/p_config.h"
  38.  
  39. #include "r300_cb.h"
  40. #include "r300_context.h"
  41. #include "r300_emit.h"
  42. #include "r300_reg.h"
  43. #include "r300_screen.h"
  44. #include "r300_screen_buffer.h"
  45. #include "r300_state_inlines.h"
  46. #include "r300_fs.h"
  47. #include "r300_texture.h"
  48. #include "r300_vs.h"
  49.  
  50. /* r300_state: Functions used to intialize state context by translating
  51.  * Gallium state objects into semi-native r300 state objects. */
  52.  
  53. #define UPDATE_STATE(cso, atom) \
  54.     if (cso != atom.state) { \
  55.         atom.state = cso;    \
  56.         r300_mark_atom_dirty(r300, &(atom));   \
  57.     }
  58.  
  59. static boolean blend_discard_if_src_alpha_0(unsigned srcRGB, unsigned srcA,
  60.                                             unsigned dstRGB, unsigned dstA)
  61. {
  62.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  63.      * SRC_ALPHA == 0, and the following state is set, the colorbuffer
  64.      * will not be changed.
  65.      * Notice that the dst factors are the src factors inverted. */
  66.     return (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
  67.             srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
  68.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  69.            (srcA == PIPE_BLENDFACTOR_SRC_COLOR ||
  70.             srcA == PIPE_BLENDFACTOR_SRC_ALPHA ||
  71.             srcA == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
  72.             srcA == PIPE_BLENDFACTOR_ZERO) &&
  73.            (dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  74.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  75.            (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  76.             dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  77.             dstA == PIPE_BLENDFACTOR_ONE);
  78. }
  79.  
  80. static boolean blend_discard_if_src_alpha_1(unsigned srcRGB, unsigned srcA,
  81.                                             unsigned dstRGB, unsigned dstA)
  82. {
  83.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  84.      * SRC_ALPHA == 1, and the following state is set, the colorbuffer
  85.      * will not be changed.
  86.      * Notice that the dst factors are the src factors inverted. */
  87.     return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  88.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  89.            (srcA == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  90.             srcA == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  91.             srcA == PIPE_BLENDFACTOR_ZERO) &&
  92.            (dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
  93.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  94.            (dstA == PIPE_BLENDFACTOR_SRC_COLOR ||
  95.             dstA == PIPE_BLENDFACTOR_SRC_ALPHA ||
  96.             dstA == PIPE_BLENDFACTOR_ONE);
  97. }
  98.  
  99. static boolean blend_discard_if_src_color_0(unsigned srcRGB, unsigned srcA,
  100.                                             unsigned dstRGB, unsigned dstA)
  101. {
  102.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  103.      * SRC_COLOR == (0,0,0), and the following state is set, the colorbuffer
  104.      * will not be changed.
  105.      * Notice that the dst factors are the src factors inverted. */
  106.     return (srcRGB == PIPE_BLENDFACTOR_SRC_COLOR ||
  107.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  108.            (srcA == PIPE_BLENDFACTOR_ZERO) &&
  109.            (dstRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  110.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  111.            (dstA == PIPE_BLENDFACTOR_ONE);
  112. }
  113.  
  114. static boolean blend_discard_if_src_color_1(unsigned srcRGB, unsigned srcA,
  115.                                             unsigned dstRGB, unsigned dstA)
  116. {
  117.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  118.      * SRC_COLOR == (1,1,1), and the following state is set, the colorbuffer
  119.      * will not be changed.
  120.      * Notice that the dst factors are the src factors inverted. */
  121.     return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  122.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  123.            (srcA == PIPE_BLENDFACTOR_ZERO) &&
  124.            (dstRGB == PIPE_BLENDFACTOR_SRC_COLOR ||
  125.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  126.            (dstA == PIPE_BLENDFACTOR_ONE);
  127. }
  128.  
  129. static boolean blend_discard_if_src_alpha_color_0(unsigned srcRGB, unsigned srcA,
  130.                                                   unsigned dstRGB, unsigned dstA)
  131. {
  132.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  133.      * SRC_ALPHA_COLOR == (0,0,0,0), and the following state is set,
  134.      * the colorbuffer will not be changed.
  135.      * Notice that the dst factors are the src factors inverted. */
  136.     return (srcRGB == PIPE_BLENDFACTOR_SRC_COLOR ||
  137.             srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
  138.             srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
  139.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  140.            (srcA == PIPE_BLENDFACTOR_SRC_COLOR ||
  141.             srcA == PIPE_BLENDFACTOR_SRC_ALPHA ||
  142.             srcA == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
  143.             srcA == PIPE_BLENDFACTOR_ZERO) &&
  144.            (dstRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  145.             dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  146.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  147.            (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  148.             dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  149.             dstA == PIPE_BLENDFACTOR_ONE);
  150. }
  151.  
  152. static boolean blend_discard_if_src_alpha_color_1(unsigned srcRGB, unsigned srcA,
  153.                                                   unsigned dstRGB, unsigned dstA)
  154. {
  155.     /* If the blend equation is ADD or REVERSE_SUBTRACT,
  156.      * SRC_ALPHA_COLOR == (1,1,1,1), and the following state is set,
  157.      * the colorbuffer will not be changed.
  158.      * Notice that the dst factors are the src factors inverted. */
  159.     return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  160.             srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  161.             srcRGB == PIPE_BLENDFACTOR_ZERO) &&
  162.            (srcA == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  163.             srcA == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  164.             srcA == PIPE_BLENDFACTOR_ZERO) &&
  165.            (dstRGB == PIPE_BLENDFACTOR_SRC_COLOR ||
  166.             dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
  167.             dstRGB == PIPE_BLENDFACTOR_ONE) &&
  168.            (dstA == PIPE_BLENDFACTOR_SRC_COLOR ||
  169.             dstA == PIPE_BLENDFACTOR_SRC_ALPHA ||
  170.             dstA == PIPE_BLENDFACTOR_ONE);
  171. }
  172.  
  173. static unsigned blend_discard_conditionally(unsigned eqRGB, unsigned eqA,
  174.                                             unsigned dstRGB, unsigned dstA,
  175.                                             unsigned srcRGB, unsigned srcA)
  176. {
  177.     unsigned blend_control = 0;
  178.  
  179.     /* Optimization: discard pixels which don't change the colorbuffer.
  180.      *
  181.      * The code below is non-trivial and some math is involved.
  182.      *
  183.      * Discarding pixels must be disabled when FP16 AA is enabled.
  184.      * This is a hardware bug. Also, this implementation wouldn't work
  185.      * with FP blending enabled and equation clamping disabled.
  186.      *
  187.      * Equations other than ADD are rarely used and therefore won't be
  188.      * optimized. */
  189.     if ((eqRGB == PIPE_BLEND_ADD || eqRGB == PIPE_BLEND_REVERSE_SUBTRACT) &&
  190.         (eqA == PIPE_BLEND_ADD || eqA == PIPE_BLEND_REVERSE_SUBTRACT)) {
  191.         /* ADD: X+Y
  192.          * REVERSE_SUBTRACT: Y-X
  193.          *
  194.          * The idea is:
  195.          * If X = src*srcFactor = 0 and Y = dst*dstFactor = 1,
  196.          * then CB will not be changed.
  197.          *
  198.          * Given the srcFactor and dstFactor variables, we can derive
  199.          * what src and dst should be equal to and discard appropriate
  200.          * pixels.
  201.          */
  202.         if (blend_discard_if_src_alpha_0(srcRGB, srcA, dstRGB, dstA)) {
  203.             blend_control |= R300_DISCARD_SRC_PIXELS_SRC_ALPHA_0;
  204.         } else if (blend_discard_if_src_alpha_1(srcRGB, srcA,
  205.                                                 dstRGB, dstA)) {
  206.             blend_control |= R300_DISCARD_SRC_PIXELS_SRC_ALPHA_1;
  207.         } else if (blend_discard_if_src_color_0(srcRGB, srcA,
  208.                                                 dstRGB, dstA)) {
  209.             blend_control |= R300_DISCARD_SRC_PIXELS_SRC_COLOR_0;
  210.         } else if (blend_discard_if_src_color_1(srcRGB, srcA,
  211.                                                 dstRGB, dstA)) {
  212.             blend_control |= R300_DISCARD_SRC_PIXELS_SRC_COLOR_1;
  213.         } else if (blend_discard_if_src_alpha_color_0(srcRGB, srcA,
  214.                                                       dstRGB, dstA)) {
  215.             blend_control |=
  216.                 R300_DISCARD_SRC_PIXELS_SRC_ALPHA_COLOR_0;
  217.         } else if (blend_discard_if_src_alpha_color_1(srcRGB, srcA,
  218.                                                       dstRGB, dstA)) {
  219.             blend_control |=
  220.                 R300_DISCARD_SRC_PIXELS_SRC_ALPHA_COLOR_1;
  221.         }
  222.     }
  223.     return blend_control;
  224. }
  225.  
  226. /* The hardware colormask is clunky a must be swizzled depending on the format.
  227.  * This was figured out by trial-and-error. */
  228. static unsigned bgra_cmask(unsigned mask)
  229. {
  230.     return ((mask & PIPE_MASK_R) << 2) |
  231.            ((mask & PIPE_MASK_B) >> 2) |
  232.            (mask & (PIPE_MASK_G | PIPE_MASK_A));
  233. }
  234.  
  235. static unsigned rgba_cmask(unsigned mask)
  236. {
  237.     return mask & PIPE_MASK_RGBA;
  238. }
  239.  
  240. static unsigned rrrr_cmask(unsigned mask)
  241. {
  242.     return (mask & PIPE_MASK_R) |
  243.            ((mask & PIPE_MASK_R) << 1) |
  244.            ((mask & PIPE_MASK_R) << 2) |
  245.            ((mask & PIPE_MASK_R) << 3);
  246. }
  247.  
  248. static unsigned aaaa_cmask(unsigned mask)
  249. {
  250.     return ((mask & PIPE_MASK_A) >> 3) |
  251.            ((mask & PIPE_MASK_A) >> 2) |
  252.            ((mask & PIPE_MASK_A) >> 1) |
  253.            (mask & PIPE_MASK_A);
  254. }
  255.  
  256. static unsigned grrg_cmask(unsigned mask)
  257. {
  258.     return ((mask & PIPE_MASK_R) << 1) |
  259.            ((mask & PIPE_MASK_R) << 2) |
  260.            ((mask & PIPE_MASK_G) >> 1) |
  261.            ((mask & PIPE_MASK_G) << 2);
  262. }
  263.  
  264. static unsigned arra_cmask(unsigned mask)
  265. {
  266.     return ((mask & PIPE_MASK_R) << 1) |
  267.            ((mask & PIPE_MASK_R) << 2) |
  268.            ((mask & PIPE_MASK_A) >> 3) |
  269.            (mask & PIPE_MASK_A);
  270. }
  271.  
  272. static unsigned blend_read_enable(unsigned eqRGB, unsigned eqA,
  273.                                   unsigned dstRGB, unsigned dstA,
  274.                                   unsigned srcRGB, unsigned srcA,
  275.                                   boolean src_alpha_optz)
  276. {
  277.     unsigned blend_control = 0;
  278.  
  279.     /* Optimization: some operations do not require the destination color.
  280.      *
  281.      * When SRC_ALPHA_SATURATE is used, colorbuffer reads must be enabled,
  282.      * otherwise blending gives incorrect results. It seems to be
  283.      * a hardware bug. */
  284.     if (eqRGB == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MIN ||
  285.         eqRGB == PIPE_BLEND_MAX || eqA == PIPE_BLEND_MAX ||
  286.         dstRGB != PIPE_BLENDFACTOR_ZERO ||
  287.         dstA != PIPE_BLENDFACTOR_ZERO ||
  288.         srcRGB == PIPE_BLENDFACTOR_DST_COLOR ||
  289.         srcRGB == PIPE_BLENDFACTOR_DST_ALPHA ||
  290.         srcRGB == PIPE_BLENDFACTOR_INV_DST_COLOR ||
  291.         srcRGB == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
  292.         srcA == PIPE_BLENDFACTOR_DST_COLOR ||
  293.         srcA == PIPE_BLENDFACTOR_DST_ALPHA ||
  294.         srcA == PIPE_BLENDFACTOR_INV_DST_COLOR ||
  295.         srcA == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
  296.         srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) {
  297.         /* Enable reading from the colorbuffer. */
  298.         blend_control |= R300_READ_ENABLE;
  299.  
  300.         if (src_alpha_optz) {
  301.             /* Optimization: Depending on incoming pixels, we can
  302.              * conditionally disable the reading in hardware... */
  303.             if (eqRGB != PIPE_BLEND_MIN && eqA != PIPE_BLEND_MIN &&
  304.                 eqRGB != PIPE_BLEND_MAX && eqA != PIPE_BLEND_MAX) {
  305.                 /* Disable reading if SRC_ALPHA == 0. */
  306.                 if ((dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
  307.                      dstRGB == PIPE_BLENDFACTOR_ZERO) &&
  308.                     (dstA == PIPE_BLENDFACTOR_SRC_COLOR ||
  309.                      dstA == PIPE_BLENDFACTOR_SRC_ALPHA ||
  310.                      dstA == PIPE_BLENDFACTOR_ZERO) &&
  311.                     (srcRGB != PIPE_BLENDFACTOR_DST_COLOR &&
  312.                      srcRGB != PIPE_BLENDFACTOR_DST_ALPHA &&
  313.                      srcRGB != PIPE_BLENDFACTOR_INV_DST_COLOR &&
  314.                      srcRGB != PIPE_BLENDFACTOR_INV_DST_ALPHA)) {
  315.                      blend_control |= R500_SRC_ALPHA_0_NO_READ;
  316.                 }
  317.  
  318.                 /* Disable reading if SRC_ALPHA == 1. */
  319.                 if ((dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  320.                      dstRGB == PIPE_BLENDFACTOR_ZERO) &&
  321.                     (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR ||
  322.                      dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
  323.                      dstA == PIPE_BLENDFACTOR_ZERO) &&
  324.                     (srcRGB != PIPE_BLENDFACTOR_DST_COLOR &&
  325.                      srcRGB != PIPE_BLENDFACTOR_DST_ALPHA &&
  326.                      srcRGB != PIPE_BLENDFACTOR_INV_DST_COLOR &&
  327.                      srcRGB != PIPE_BLENDFACTOR_INV_DST_ALPHA)) {
  328.                      blend_control |= R500_SRC_ALPHA_1_NO_READ;
  329.                 }
  330.             }
  331.         }
  332.     }
  333.     return blend_control;
  334. }
  335.  
  336. /* Create a new blend state based on the CSO blend state.
  337.  *
  338.  * This encompasses alpha blending, logic/raster ops, and blend dithering. */
  339. static void* r300_create_blend_state(struct pipe_context* pipe,
  340.                                      const struct pipe_blend_state* state)
  341. {
  342.     struct r300_screen* r300screen = r300_screen(pipe->screen);
  343.     struct r300_blend_state* blend = CALLOC_STRUCT(r300_blend_state);
  344.     uint32_t blend_control = 0;       /* R300_RB3D_CBLEND: 0x4e04 */
  345.     uint32_t blend_control_noclamp = 0;    /* R300_RB3D_CBLEND: 0x4e04 */
  346.     uint32_t blend_control_noalpha = 0;    /* R300_RB3D_CBLEND: 0x4e04 */
  347.     uint32_t blend_control_noalpha_noclamp = 0;    /* R300_RB3D_CBLEND: 0x4e04 */
  348.     uint32_t alpha_blend_control = 0; /* R300_RB3D_ABLEND: 0x4e08 */
  349.     uint32_t alpha_blend_control_noclamp = 0; /* R300_RB3D_ABLEND: 0x4e08 */
  350.     uint32_t alpha_blend_control_noalpha = 0; /* R300_RB3D_ABLEND: 0x4e08 */
  351.     uint32_t alpha_blend_control_noalpha_noclamp = 0; /* R300_RB3D_ABLEND: 0x4e08 */
  352.     uint32_t rop = 0;                 /* R300_RB3D_ROPCNTL: 0x4e18 */
  353.     uint32_t dither = 0;              /* R300_RB3D_DITHER_CTL: 0x4e50 */
  354.     int i;
  355.  
  356.     const unsigned eqRGB = state->rt[0].rgb_func;
  357.     const unsigned srcRGB = state->rt[0].rgb_src_factor;
  358.     const unsigned dstRGB = state->rt[0].rgb_dst_factor;
  359.  
  360.     const unsigned eqA = state->rt[0].alpha_func;
  361.     const unsigned srcA = state->rt[0].alpha_src_factor;
  362.     const unsigned dstA = state->rt[0].alpha_dst_factor;
  363.  
  364.     unsigned srcRGBX = srcRGB;
  365.     unsigned dstRGBX = dstRGB;
  366.     CB_LOCALS;
  367.  
  368.     blend->state = *state;
  369.  
  370.     /* force DST_ALPHA to ONE where we can */
  371.     switch (srcRGBX) {
  372.     case PIPE_BLENDFACTOR_DST_ALPHA:
  373.         srcRGBX = PIPE_BLENDFACTOR_ONE;
  374.         break;
  375.     case PIPE_BLENDFACTOR_INV_DST_ALPHA:
  376.         srcRGBX = PIPE_BLENDFACTOR_ZERO;
  377.         break;
  378.     }
  379.  
  380.     switch (dstRGBX) {
  381.     case PIPE_BLENDFACTOR_DST_ALPHA:
  382.         dstRGBX = PIPE_BLENDFACTOR_ONE;
  383.         break;
  384.     case PIPE_BLENDFACTOR_INV_DST_ALPHA:
  385.         dstRGBX = PIPE_BLENDFACTOR_ZERO;
  386.         break;
  387.     }
  388.  
  389.     /* Get blending register values. */
  390.     if (state->rt[0].blend_enable) {
  391.         unsigned blend_eq, blend_eq_noclamp;
  392.  
  393.         /* despite the name, ALPHA_BLEND_ENABLE has nothing to do with alpha,
  394.          * this is just the crappy D3D naming */
  395.         blend_control = blend_control_noclamp =
  396.             R300_ALPHA_BLEND_ENABLE |
  397.             ( r300_translate_blend_factor(srcRGB) << R300_SRC_BLEND_SHIFT) |
  398.             ( r300_translate_blend_factor(dstRGB) << R300_DST_BLEND_SHIFT);
  399.  
  400.         blend_control_noalpha = blend_control_noalpha_noclamp =
  401.             R300_ALPHA_BLEND_ENABLE |
  402.             ( r300_translate_blend_factor(srcRGBX) << R300_SRC_BLEND_SHIFT) |
  403.             ( r300_translate_blend_factor(dstRGBX) << R300_DST_BLEND_SHIFT);
  404.  
  405.         blend_eq = r300_translate_blend_function(eqRGB, TRUE);
  406.         blend_eq_noclamp = r300_translate_blend_function(eqRGB, FALSE);
  407.  
  408.         blend_control |= blend_eq;
  409.         blend_control_noalpha |= blend_eq;
  410.         blend_control_noclamp |= blend_eq_noclamp;
  411.         blend_control_noalpha_noclamp |= blend_eq_noclamp;
  412.  
  413.         /* Optimization: some operations do not require the destination color. */
  414.         blend_control |= blend_read_enable(eqRGB, eqA, dstRGB, dstA,
  415.                                            srcRGB, srcA, r300screen->caps.is_r500);
  416.         blend_control_noclamp |= blend_read_enable(eqRGB, eqA, dstRGB, dstA,
  417.                                                    srcRGB, srcA, FALSE);
  418.         blend_control_noalpha |= blend_read_enable(eqRGB, eqA, dstRGBX, dstA,
  419.                                                    srcRGBX, srcA, r300screen->caps.is_r500);
  420.         blend_control_noalpha_noclamp |= blend_read_enable(eqRGB, eqA, dstRGBX, dstA,
  421.                                                            srcRGBX, srcA, FALSE);
  422.  
  423.         /* Optimization: discard pixels which don't change the colorbuffer.
  424.          * It cannot be used with FP16 AA. */
  425.         blend_control |= blend_discard_conditionally(eqRGB, eqA, dstRGB, dstA,
  426.                                                      srcRGB, srcA);
  427.         blend_control_noalpha |= blend_discard_conditionally(eqRGB, eqA, dstRGBX, dstA,
  428.                                                              srcRGBX, srcA);
  429.  
  430.         /* separate alpha */
  431.         if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
  432.             blend_control |= R300_SEPARATE_ALPHA_ENABLE;
  433.             blend_control_noclamp |= R300_SEPARATE_ALPHA_ENABLE;
  434.  
  435.             alpha_blend_control = alpha_blend_control_noclamp =
  436.                 (r300_translate_blend_factor(srcA) << R300_SRC_BLEND_SHIFT) |
  437.                 (r300_translate_blend_factor(dstA) << R300_DST_BLEND_SHIFT);
  438.             alpha_blend_control |= r300_translate_blend_function(eqA, TRUE);
  439.             alpha_blend_control_noclamp |= r300_translate_blend_function(eqA, FALSE);
  440.         }
  441.         if (srcA != srcRGBX || dstA != dstRGBX || eqA != eqRGB) {
  442.             blend_control_noalpha |= R300_SEPARATE_ALPHA_ENABLE;
  443.             blend_control_noalpha_noclamp |= R300_SEPARATE_ALPHA_ENABLE;
  444.  
  445.             alpha_blend_control_noalpha = alpha_blend_control_noalpha_noclamp =
  446.                 (r300_translate_blend_factor(srcA) << R300_SRC_BLEND_SHIFT) |
  447.                 (r300_translate_blend_factor(dstA) << R300_DST_BLEND_SHIFT);
  448.             alpha_blend_control_noalpha |= r300_translate_blend_function(eqA, TRUE);
  449.             alpha_blend_control_noalpha_noclamp |= r300_translate_blend_function(eqA, FALSE);
  450.         }
  451.     }
  452.  
  453.     /* PIPE_LOGICOP_* don't need to be translated, fortunately. */
  454.     if (state->logicop_enable) {
  455.         rop = R300_RB3D_ROPCNTL_ROP_ENABLE |
  456.                 (state->logicop_func) << R300_RB3D_ROPCNTL_ROP_SHIFT;
  457.     }
  458.  
  459.     /* Neither fglrx nor classic r300 ever set this, regardless of dithering
  460.      * state. Since it's an optional implementation detail, we can leave it
  461.      * out and never dither.
  462.      *
  463.      * This could be revisited if we ever get quality or conformance hints.
  464.      *
  465.     if (state->dither) {
  466.         dither = R300_RB3D_DITHER_CTL_DITHER_MODE_LUT |
  467.                         R300_RB3D_DITHER_CTL_ALPHA_DITHER_MODE_LUT;
  468.     }
  469.     */
  470.  
  471.     /* Build a command buffer. */
  472.     {
  473.         unsigned (*func[COLORMASK_NUM_SWIZZLES])(unsigned) = {
  474.             bgra_cmask,
  475.             rgba_cmask,
  476.             rrrr_cmask,
  477.             aaaa_cmask,
  478.             grrg_cmask,
  479.             arra_cmask,
  480.             bgra_cmask,
  481.             rgba_cmask
  482.         };
  483.  
  484.         for (i = 0; i < COLORMASK_NUM_SWIZZLES; i++) {
  485.             boolean has_alpha = i != COLORMASK_RGBX && i != COLORMASK_BGRX;
  486.  
  487.             BEGIN_CB(blend->cb_clamp[i], 8);
  488.             OUT_CB_REG(R300_RB3D_ROPCNTL, rop);
  489.             OUT_CB_REG_SEQ(R300_RB3D_CBLEND, 3);
  490.             OUT_CB(has_alpha ? blend_control : blend_control_noalpha);
  491.             OUT_CB(has_alpha ? alpha_blend_control : alpha_blend_control_noalpha);
  492.             OUT_CB(func[i](state->rt[0].colormask));
  493.             OUT_CB_REG(R300_RB3D_DITHER_CTL, dither);
  494.             END_CB;
  495.         }
  496.     }
  497.  
  498.     /* Build a command buffer (for RGBA16F). */
  499.     BEGIN_CB(blend->cb_noclamp, 8);
  500.     OUT_CB_REG(R300_RB3D_ROPCNTL, rop);
  501.     OUT_CB_REG_SEQ(R300_RB3D_CBLEND, 3);
  502.     OUT_CB(blend_control_noclamp);
  503.     OUT_CB(alpha_blend_control_noclamp);
  504.     OUT_CB(rgba_cmask(state->rt[0].colormask));
  505.     OUT_CB_REG(R300_RB3D_DITHER_CTL, dither);
  506.     END_CB;
  507.  
  508.     /* Build a command buffer (for RGB16F). */
  509.     BEGIN_CB(blend->cb_noclamp_noalpha, 8);
  510.     OUT_CB_REG(R300_RB3D_ROPCNTL, rop);
  511.     OUT_CB_REG_SEQ(R300_RB3D_CBLEND, 3);
  512.     OUT_CB(blend_control_noalpha_noclamp);
  513.     OUT_CB(alpha_blend_control_noalpha_noclamp);
  514.     OUT_CB(rgba_cmask(state->rt[0].colormask));
  515.     OUT_CB_REG(R300_RB3D_DITHER_CTL, dither);
  516.     END_CB;
  517.  
  518.     /* The same as above, but with no colorbuffer reads and writes. */
  519.     BEGIN_CB(blend->cb_no_readwrite, 8);
  520.     OUT_CB_REG(R300_RB3D_ROPCNTL, rop);
  521.     OUT_CB_REG_SEQ(R300_RB3D_CBLEND, 3);
  522.     OUT_CB(0);
  523.     OUT_CB(0);
  524.     OUT_CB(0);
  525.     OUT_CB_REG(R300_RB3D_DITHER_CTL, dither);
  526.     END_CB;
  527.  
  528.     return (void*)blend;
  529. }
  530.  
  531. /* Bind blend state. */
  532. static void r300_bind_blend_state(struct pipe_context* pipe,
  533.                                   void* state)
  534. {
  535.     struct r300_context* r300 = r300_context(pipe);
  536.     struct r300_blend_state *blend  = (struct r300_blend_state*)state;
  537.     boolean last_alpha_to_one = r300->alpha_to_one;
  538.     boolean last_alpha_to_coverage = r300->alpha_to_coverage;
  539.  
  540.     UPDATE_STATE(state, r300->blend_state);
  541.  
  542.     if (!blend)
  543.         return;
  544.  
  545.     r300->alpha_to_one = blend->state.alpha_to_one;
  546.     r300->alpha_to_coverage = blend->state.alpha_to_coverage;
  547.  
  548.     if (r300->alpha_to_one != last_alpha_to_one && r300->msaa_enable &&
  549.         r300->fs_status == FRAGMENT_SHADER_VALID) {
  550.         r300->fs_status = FRAGMENT_SHADER_MAYBE_DIRTY;
  551.     }
  552.  
  553.     if (r300->alpha_to_coverage != last_alpha_to_coverage &&
  554.         r300->msaa_enable) {
  555.         r300_mark_atom_dirty(r300, &r300->dsa_state);
  556.     }
  557. }
  558.  
  559. /* Free blend state. */
  560. static void r300_delete_blend_state(struct pipe_context* pipe,
  561.                                     void* state)
  562. {
  563.     FREE(state);
  564. }
  565.  
  566. /* Convert float to 10bit integer */
  567. static unsigned float_to_fixed10(float f)
  568. {
  569.     return CLAMP((unsigned)(f * 1023.9f), 0, 1023);
  570. }
  571.  
  572. /* Set blend color.
  573.  * Setup both R300 and R500 registers, figure out later which one to write. */
  574. static void r300_set_blend_color(struct pipe_context* pipe,
  575.                                  const struct pipe_blend_color* color)
  576. {
  577.     struct r300_context* r300 = r300_context(pipe);
  578.     struct pipe_framebuffer_state *fb = r300->fb_state.state;
  579.     struct r300_blend_color_state *state =
  580.         (struct r300_blend_color_state*)r300->blend_color_state.state;
  581.     struct pipe_blend_color c;
  582.     enum pipe_format format = fb->nr_cbufs ? fb->cbufs[0]->format : 0;
  583.     float tmp;
  584.     CB_LOCALS;
  585.  
  586.     state->state = *color; /* Save it, so that we can reuse it in set_fb_state */
  587.     c = *color;
  588.  
  589.     /* The blend color is dependent on the colorbuffer format. */
  590.     if (fb->nr_cbufs) {
  591.         switch (format) {
  592.         case PIPE_FORMAT_R8_UNORM:
  593.         case PIPE_FORMAT_L8_UNORM:
  594.         case PIPE_FORMAT_I8_UNORM:
  595.             c.color[1] = c.color[0];
  596.             break;
  597.  
  598.         case PIPE_FORMAT_A8_UNORM:
  599.             c.color[1] = c.color[3];
  600.             break;
  601.  
  602.         case PIPE_FORMAT_R8G8_UNORM:
  603.             c.color[2] = c.color[1];
  604.             break;
  605.  
  606.         case PIPE_FORMAT_L8A8_UNORM:
  607.         case PIPE_FORMAT_R8A8_UNORM:
  608.             c.color[2] = c.color[3];
  609.             break;
  610.  
  611.         case PIPE_FORMAT_R8G8B8A8_UNORM:
  612.         case PIPE_FORMAT_R8G8B8X8_UNORM:
  613.             tmp = c.color[0];
  614.             c.color[0] = c.color[2];
  615.             c.color[2] = tmp;
  616.             break;
  617.  
  618.         default:;
  619.         }
  620.     }
  621.  
  622.     if (r300->screen->caps.is_r500) {
  623.         BEGIN_CB(state->cb, 3);
  624.         OUT_CB_REG_SEQ(R500_RB3D_CONSTANT_COLOR_AR, 2);
  625.  
  626.         switch (format) {
  627.         case PIPE_FORMAT_R16G16B16A16_FLOAT:
  628.         case PIPE_FORMAT_R16G16B16X16_FLOAT:
  629.             OUT_CB(util_float_to_half(c.color[2]) |
  630.                    (util_float_to_half(c.color[3]) << 16));
  631.             OUT_CB(util_float_to_half(c.color[0]) |
  632.                    (util_float_to_half(c.color[1]) << 16));
  633.             break;
  634.  
  635.         default:
  636.             OUT_CB(float_to_fixed10(c.color[0]) |
  637.                    (float_to_fixed10(c.color[3]) << 16));
  638.             OUT_CB(float_to_fixed10(c.color[2]) |
  639.                    (float_to_fixed10(c.color[1]) << 16));
  640.         }
  641.  
  642.         END_CB;
  643.     } else {
  644.         union util_color uc;
  645.         util_pack_color(c.color, PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
  646.  
  647.         BEGIN_CB(state->cb, 2);
  648.         OUT_CB_REG(R300_RB3D_BLEND_COLOR, uc.ui);
  649.         END_CB;
  650.     }
  651.  
  652.     r300_mark_atom_dirty(r300, &r300->blend_color_state);
  653. }
  654.  
  655. static void r300_set_clip_state(struct pipe_context* pipe,
  656.                                 const struct pipe_clip_state* state)
  657. {
  658.     struct r300_context* r300 = r300_context(pipe);
  659.     struct r300_clip_state *clip =
  660.             (struct r300_clip_state*)r300->clip_state.state;
  661.     CB_LOCALS;
  662.  
  663.     if (r300->screen->caps.has_tcl) {
  664.         BEGIN_CB(clip->cb, r300->clip_state.size);
  665.         OUT_CB_REG(R300_VAP_PVS_VECTOR_INDX_REG,
  666.                    (r300->screen->caps.is_r500 ?
  667.                     R500_PVS_UCP_START : R300_PVS_UCP_START));
  668.         OUT_CB_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, 6 * 4);
  669.         OUT_CB_TABLE(state->ucp, 6 * 4);
  670.         END_CB;
  671.  
  672.         r300_mark_atom_dirty(r300, &r300->clip_state);
  673.     } else {
  674.         draw_set_clip_state(r300->draw, state);
  675.     }
  676. }
  677.  
  678. /* Create a new depth, stencil, and alpha state based on the CSO dsa state.
  679.  *
  680.  * This contains the depth buffer, stencil buffer, alpha test, and such.
  681.  * On the Radeon, depth and stencil buffer setup are intertwined, which is
  682.  * the reason for some of the strange-looking assignments across registers. */
  683. static void* r300_create_dsa_state(struct pipe_context* pipe,
  684.                           const struct pipe_depth_stencil_alpha_state* state)
  685. {
  686.     boolean is_r500 = r300_screen(pipe->screen)->caps.is_r500;
  687.     struct r300_dsa_state* dsa = CALLOC_STRUCT(r300_dsa_state);
  688.     CB_LOCALS;
  689.     uint32_t alpha_value_fp16 = 0;
  690.     uint32_t z_buffer_control = 0;
  691.     uint32_t z_stencil_control = 0;
  692.     uint32_t stencil_ref_mask = 0;
  693.     uint32_t stencil_ref_bf = 0;
  694.  
  695.     dsa->dsa = *state;
  696.  
  697.     /* Depth test setup. - separate write mask depth for decomp flush */
  698.     if (state->depth.writemask) {
  699.         z_buffer_control |= R300_Z_WRITE_ENABLE;
  700.     }
  701.  
  702.     if (state->depth.enabled) {
  703.         z_buffer_control |= R300_Z_ENABLE;
  704.  
  705.         z_stencil_control |=
  706.             (r300_translate_depth_stencil_function(state->depth.func) <<
  707.                 R300_Z_FUNC_SHIFT);
  708.     }
  709.  
  710.     /* Stencil buffer setup. */
  711.     if (state->stencil[0].enabled) {
  712.         z_buffer_control |= R300_STENCIL_ENABLE;
  713.         z_stencil_control |=
  714.             (r300_translate_depth_stencil_function(state->stencil[0].func) <<
  715.                 R300_S_FRONT_FUNC_SHIFT) |
  716.             (r300_translate_stencil_op(state->stencil[0].fail_op) <<
  717.                 R300_S_FRONT_SFAIL_OP_SHIFT) |
  718.             (r300_translate_stencil_op(state->stencil[0].zpass_op) <<
  719.                 R300_S_FRONT_ZPASS_OP_SHIFT) |
  720.             (r300_translate_stencil_op(state->stencil[0].zfail_op) <<
  721.                 R300_S_FRONT_ZFAIL_OP_SHIFT);
  722.  
  723.         stencil_ref_mask =
  724.                 (state->stencil[0].valuemask << R300_STENCILMASK_SHIFT) |
  725.                 (state->stencil[0].writemask << R300_STENCILWRITEMASK_SHIFT);
  726.  
  727.         if (state->stencil[1].enabled) {
  728.             dsa->two_sided = TRUE;
  729.  
  730.             z_buffer_control |= R300_STENCIL_FRONT_BACK;
  731.             z_stencil_control |=
  732.             (r300_translate_depth_stencil_function(state->stencil[1].func) <<
  733.                 R300_S_BACK_FUNC_SHIFT) |
  734.             (r300_translate_stencil_op(state->stencil[1].fail_op) <<
  735.                 R300_S_BACK_SFAIL_OP_SHIFT) |
  736.             (r300_translate_stencil_op(state->stencil[1].zpass_op) <<
  737.                 R300_S_BACK_ZPASS_OP_SHIFT) |
  738.             (r300_translate_stencil_op(state->stencil[1].zfail_op) <<
  739.                 R300_S_BACK_ZFAIL_OP_SHIFT);
  740.  
  741.             stencil_ref_bf =
  742.                 (state->stencil[1].valuemask << R300_STENCILMASK_SHIFT) |
  743.                 (state->stencil[1].writemask << R300_STENCILWRITEMASK_SHIFT);
  744.  
  745.             if (is_r500) {
  746.                 z_buffer_control |= R500_STENCIL_REFMASK_FRONT_BACK;
  747.             } else {
  748.                 dsa->two_sided_stencil_ref =
  749.                   (state->stencil[0].valuemask != state->stencil[1].valuemask ||
  750.                    state->stencil[0].writemask != state->stencil[1].writemask);
  751.             }
  752.         }
  753.     }
  754.  
  755.     /* Alpha test setup. */
  756.     if (state->alpha.enabled) {
  757.         dsa->alpha_function =
  758.             r300_translate_alpha_function(state->alpha.func) |
  759.             R300_FG_ALPHA_FUNC_ENABLE;
  760.  
  761.         dsa->alpha_function |= float_to_ubyte(state->alpha.ref_value);
  762.         alpha_value_fp16 = util_float_to_half(state->alpha.ref_value);
  763.     }
  764.  
  765.     BEGIN_CB(&dsa->cb_begin, 8);
  766.     OUT_CB_REG_SEQ(R300_ZB_CNTL, 3);
  767.     OUT_CB(z_buffer_control);
  768.     OUT_CB(z_stencil_control);
  769.     OUT_CB(stencil_ref_mask);
  770.     OUT_CB_REG(R500_ZB_STENCILREFMASK_BF, stencil_ref_bf);
  771.     OUT_CB_REG(R500_FG_ALPHA_VALUE, alpha_value_fp16);
  772.     END_CB;
  773.  
  774.     BEGIN_CB(dsa->cb_zb_no_readwrite, 8);
  775.     OUT_CB_REG_SEQ(R300_ZB_CNTL, 3);
  776.     OUT_CB(0);
  777.     OUT_CB(0);
  778.     OUT_CB(0);
  779.     OUT_CB_REG(R500_ZB_STENCILREFMASK_BF, 0);
  780.     OUT_CB_REG(R500_FG_ALPHA_VALUE, alpha_value_fp16);
  781.     END_CB;
  782.  
  783.     return (void*)dsa;
  784. }
  785.  
  786. static void r300_dsa_inject_stencilref(struct r300_context *r300)
  787. {
  788.     struct r300_dsa_state *dsa =
  789.             (struct r300_dsa_state*)r300->dsa_state.state;
  790.  
  791.     if (!dsa)
  792.         return;
  793.  
  794.     dsa->stencil_ref_mask =
  795.         (dsa->stencil_ref_mask & ~R300_STENCILREF_MASK) |
  796.         r300->stencil_ref.ref_value[0];
  797.     dsa->stencil_ref_bf =
  798.         (dsa->stencil_ref_bf & ~R300_STENCILREF_MASK) |
  799.         r300->stencil_ref.ref_value[1];
  800. }
  801.  
  802. /* Bind DSA state. */
  803. static void r300_bind_dsa_state(struct pipe_context* pipe,
  804.                                 void* state)
  805. {
  806.     struct r300_context* r300 = r300_context(pipe);
  807.  
  808.     if (!state) {
  809.         return;
  810.     }
  811.  
  812.     UPDATE_STATE(state, r300->dsa_state);
  813.  
  814.     r300_mark_atom_dirty(r300, &r300->hyperz_state); /* Will be updated before the emission. */
  815.     r300_dsa_inject_stencilref(r300);
  816. }
  817.  
  818. /* Free DSA state. */
  819. static void r300_delete_dsa_state(struct pipe_context* pipe,
  820.                                   void* state)
  821. {
  822.     FREE(state);
  823. }
  824.  
  825. static void r300_set_stencil_ref(struct pipe_context* pipe,
  826.                                  const struct pipe_stencil_ref* sr)
  827. {
  828.     struct r300_context* r300 = r300_context(pipe);
  829.  
  830.     r300->stencil_ref = *sr;
  831.  
  832.     r300_dsa_inject_stencilref(r300);
  833.     r300_mark_atom_dirty(r300, &r300->dsa_state);
  834. }
  835.  
  836. static void r300_tex_set_tiling_flags(struct r300_context *r300,
  837.                                       struct r300_resource *tex,
  838.                                       unsigned level)
  839. {
  840.     /* Check if the macrotile flag needs to be changed.
  841.      * Skip changing the flags otherwise. */
  842.     if (tex->tex.macrotile[tex->surface_level] !=
  843.         tex->tex.macrotile[level]) {
  844.         r300->rws->buffer_set_tiling(tex->buf, r300->cs,
  845.                 tex->tex.microtile, tex->tex.macrotile[level],
  846.                 0, 0, 0, 0, 0,
  847.                 tex->tex.stride_in_bytes[0]);
  848.  
  849.         tex->surface_level = level;
  850.     }
  851. }
  852.  
  853. /* This switcheroo is needed just because of goddamned MACRO_SWITCH. */
  854. static void r300_fb_set_tiling_flags(struct r300_context *r300,
  855.                                const struct pipe_framebuffer_state *state)
  856. {
  857.     unsigned i;
  858.  
  859.     /* Set tiling flags for new surfaces. */
  860.     for (i = 0; i < state->nr_cbufs; i++) {
  861.         r300_tex_set_tiling_flags(r300,
  862.                                   r300_resource(state->cbufs[i]->texture),
  863.                                   state->cbufs[i]->u.tex.level);
  864.     }
  865.     if (state->zsbuf) {
  866.         r300_tex_set_tiling_flags(r300,
  867.                                   r300_resource(state->zsbuf->texture),
  868.                                   state->zsbuf->u.tex.level);
  869.     }
  870. }
  871.  
  872. static void r300_print_fb_surf_info(struct pipe_surface *surf, unsigned index,
  873.                                     const char *binding)
  874. {
  875.     struct pipe_resource *tex = surf->texture;
  876.     struct r300_resource *rtex = r300_resource(tex);
  877.  
  878.     fprintf(stderr,
  879.             "r300:   %s[%i] Dim: %ix%i, Firstlayer: %i, "
  880.             "Lastlayer: %i, Level: %i, Format: %s\n"
  881.  
  882.             "r300:     TEX: Macro: %s, Micro: %s, "
  883.             "Dim: %ix%ix%i, LastLevel: %i, Format: %s\n",
  884.  
  885.             binding, index, surf->width, surf->height,
  886.             surf->u.tex.first_layer, surf->u.tex.last_layer, surf->u.tex.level,
  887.             util_format_short_name(surf->format),
  888.  
  889.             rtex->tex.macrotile[0] ? "YES" : " NO",
  890.             rtex->tex.microtile ? "YES" : " NO",
  891.             tex->width0, tex->height0, tex->depth0,
  892.             tex->last_level, util_format_short_name(surf->format));
  893. }
  894.  
  895. void r300_mark_fb_state_dirty(struct r300_context *r300,
  896.                               enum r300_fb_state_change change)
  897. {
  898.     struct pipe_framebuffer_state *state = r300->fb_state.state;
  899.  
  900.     r300_mark_atom_dirty(r300, &r300->gpu_flush);
  901.     r300_mark_atom_dirty(r300, &r300->fb_state);
  902.  
  903.     /* What is marked as dirty depends on the enum r300_fb_state_change. */
  904.     if (change == R300_CHANGED_FB_STATE) {
  905.         r300_mark_atom_dirty(r300, &r300->aa_state);
  906.         r300_mark_atom_dirty(r300, &r300->dsa_state); /* for AlphaRef */
  907.         r300_set_blend_color(&r300->context, r300->blend_color_state.state);
  908.     }
  909.  
  910.     if (change == R300_CHANGED_FB_STATE ||
  911.         change == R300_CHANGED_HYPERZ_FLAG) {
  912.         r300_mark_atom_dirty(r300, &r300->hyperz_state);
  913.     }
  914.  
  915.     if (change == R300_CHANGED_FB_STATE ||
  916.         change == R300_CHANGED_MULTIWRITE) {
  917.         r300_mark_atom_dirty(r300, &r300->fb_state_pipelined);
  918.     }
  919.  
  920.     /* Now compute the fb_state atom size. */
  921.     r300->fb_state.size = 2 + (8 * state->nr_cbufs);
  922.  
  923.     if (r300->cbzb_clear)
  924.         r300->fb_state.size += 10;
  925.     else if (state->zsbuf) {
  926.         r300->fb_state.size += 10;
  927.         if (r300->hyperz_enabled)
  928.             r300->fb_state.size += 8;
  929.     }
  930.  
  931.     if (r300->cmask_in_use) {
  932.         r300->fb_state.size += 6;
  933.         if (r300->screen->caps.is_r500 && r300->screen->info.drm_minor >= 29) {
  934.             r300->fb_state.size += 3;
  935.         }
  936.     }
  937.  
  938.     /* The size of the rest of atoms stays the same. */
  939. }
  940.  
  941. static unsigned r300_get_num_samples(struct r300_context *r300)
  942. {
  943.     struct pipe_framebuffer_state* fb =
  944.             (struct pipe_framebuffer_state*)r300->fb_state.state;
  945.     unsigned i, num_samples;
  946.  
  947.     if (!fb->nr_cbufs && !fb->zsbuf)
  948.         return 1;
  949.  
  950.     num_samples = 6;
  951.  
  952.     for (i = 0; i < fb->nr_cbufs; i++)
  953.         num_samples = MIN2(num_samples, fb->cbufs[i]->texture->nr_samples);
  954.  
  955.     if (fb->zsbuf)
  956.         num_samples = MIN2(num_samples, fb->zsbuf->texture->nr_samples);
  957.  
  958.     if (!num_samples)
  959.         num_samples = 1;
  960.  
  961.     return num_samples;
  962. }
  963.  
  964. static void
  965. r300_set_framebuffer_state(struct pipe_context* pipe,
  966.                            const struct pipe_framebuffer_state* state)
  967. {
  968.     struct r300_context* r300 = r300_context(pipe);
  969.     struct r300_aa_state *aa = (struct r300_aa_state*)r300->aa_state.state;
  970.     struct pipe_framebuffer_state *old_state = r300->fb_state.state;
  971.     unsigned max_width, max_height, i;
  972.     uint32_t zbuffer_bpp = 0;
  973.     boolean unlock_zbuffer = FALSE;
  974.  
  975.     if (r300->screen->caps.is_r500) {
  976.         max_width = max_height = 4096;
  977.     } else if (r300->screen->caps.is_r400) {
  978.         max_width = max_height = 4021;
  979.     } else {
  980.         max_width = max_height = 2560;
  981.     }
  982.  
  983.     if (state->width > max_width || state->height > max_height) {
  984.         fprintf(stderr, "r300: Implementation error: Render targets are too "
  985.         "big in %s, refusing to bind framebuffer state!\n", __FUNCTION__);
  986.         return;
  987.     }
  988.  
  989.     if (old_state->zsbuf && r300->zmask_in_use && !r300->locked_zbuffer) {
  990.         /* There is a zmask in use, what are we gonna do? */
  991.         if (state->zsbuf) {
  992.             if (!pipe_surface_equal(old_state->zsbuf, state->zsbuf)) {
  993.                 /* Decompress the currently bound zbuffer before we bind another one. */
  994.                 r300_decompress_zmask(r300);
  995.                 r300->hiz_in_use = FALSE;
  996.             }
  997.         } else {
  998.             /* We don't bind another zbuffer, so lock the current one. */
  999.             pipe_surface_reference(&r300->locked_zbuffer, old_state->zsbuf);
  1000.         }
  1001.     } else if (r300->locked_zbuffer) {
  1002.         /* We have a locked zbuffer now, what are we gonna do? */
  1003.         if (state->zsbuf) {
  1004.             if (!pipe_surface_equal(r300->locked_zbuffer, state->zsbuf)) {
  1005.                 /* We are binding some other zbuffer, so decompress the locked one,
  1006.                  * it gets unlocked automatically. */
  1007.                 r300_decompress_zmask_locked_unsafe(r300);
  1008.                 r300->hiz_in_use = FALSE;
  1009.             } else {
  1010.                 /* We are binding the locked zbuffer again, so unlock it. */
  1011.                 unlock_zbuffer = TRUE;
  1012.             }
  1013.         }
  1014.     }
  1015.     assert(state->zsbuf || (r300->locked_zbuffer && !unlock_zbuffer) || !r300->zmask_in_use);
  1016.  
  1017.     /* Set whether CMASK can be used. */
  1018.     r300->cmask_in_use =
  1019.         state->nr_cbufs == 1 &&
  1020.         r300->screen->cmask_resource == state->cbufs[0]->texture;
  1021.  
  1022.     /* Need to reset clamping or colormask. */
  1023.     r300_mark_atom_dirty(r300, &r300->blend_state);
  1024.  
  1025.     /* Re-swizzle the blend color. */
  1026.     r300_set_blend_color(pipe, &((struct r300_blend_color_state*)r300->blend_color_state.state)->state);
  1027.  
  1028.     /* If zsbuf is set from NULL to non-NULL or vice versa.. */
  1029.     if (!!old_state->zsbuf != !!state->zsbuf) {
  1030.         r300_mark_atom_dirty(r300, &r300->dsa_state);
  1031.     }
  1032.  
  1033.     if (r300->screen->info.drm_minor < 12) {
  1034.        /* The tiling flags are dependent on the surface miplevel, unfortunately.
  1035.         * This workarounds a bad design decision in old kernels which were
  1036.         * rewriting tile fields in registers. */
  1037.         r300_fb_set_tiling_flags(r300, state);
  1038.     }
  1039.  
  1040.     util_copy_framebuffer_state(r300->fb_state.state, state);
  1041.  
  1042.     if (unlock_zbuffer) {
  1043.         pipe_surface_reference(&r300->locked_zbuffer, NULL);
  1044.     }
  1045.  
  1046.     r300_mark_fb_state_dirty(r300, R300_CHANGED_FB_STATE);
  1047.  
  1048.     if (state->zsbuf) {
  1049.         switch (util_format_get_blocksize(state->zsbuf->format)) {
  1050.         case 2:
  1051.             zbuffer_bpp = 16;
  1052.             break;
  1053.         case 4:
  1054.             zbuffer_bpp = 24;
  1055.             break;
  1056.         }
  1057.  
  1058.         /* Polygon offset depends on the zbuffer bit depth. */
  1059.         if (r300->zbuffer_bpp != zbuffer_bpp) {
  1060.             r300->zbuffer_bpp = zbuffer_bpp;
  1061.  
  1062.             if (r300->polygon_offset_enabled)
  1063.                 r300_mark_atom_dirty(r300, &r300->rs_state);
  1064.         }
  1065.     }
  1066.  
  1067.     r300->num_samples = r300_get_num_samples(r300);
  1068.  
  1069.     /* Set up AA config. */
  1070.     if (r300->num_samples > 1) {
  1071.         switch (r300->num_samples) {
  1072.         case 2:
  1073.             aa->aa_config = R300_GB_AA_CONFIG_AA_ENABLE |
  1074.                             R300_GB_AA_CONFIG_NUM_AA_SUBSAMPLES_2;
  1075.             break;
  1076.         case 4:
  1077.             aa->aa_config = R300_GB_AA_CONFIG_AA_ENABLE |
  1078.                             R300_GB_AA_CONFIG_NUM_AA_SUBSAMPLES_4;
  1079.             break;
  1080.         case 6:
  1081.             aa->aa_config = R300_GB_AA_CONFIG_AA_ENABLE |
  1082.                             R300_GB_AA_CONFIG_NUM_AA_SUBSAMPLES_6;
  1083.             break;
  1084.         }
  1085.     } else {
  1086.         aa->aa_config = 0;
  1087.     }
  1088.  
  1089.     if (DBG_ON(r300, DBG_FB)) {
  1090.         fprintf(stderr, "r300: set_framebuffer_state:\n");
  1091.         for (i = 0; i < state->nr_cbufs; i++) {
  1092.             r300_print_fb_surf_info(state->cbufs[i], i, "CB");
  1093.         }
  1094.         if (state->zsbuf) {
  1095.             r300_print_fb_surf_info(state->zsbuf, 0, "ZB");
  1096.         }
  1097.     }
  1098. }
  1099.  
  1100. /* Create fragment shader state. */
  1101. static void* r300_create_fs_state(struct pipe_context* pipe,
  1102.                                   const struct pipe_shader_state* shader)
  1103. {
  1104.     struct r300_fragment_shader* fs = NULL;
  1105.  
  1106.     fs = (struct r300_fragment_shader*)CALLOC_STRUCT(r300_fragment_shader);
  1107.  
  1108.     /* Copy state directly into shader. */
  1109.     fs->state = *shader;
  1110.     fs->state.tokens = tgsi_dup_tokens(shader->tokens);
  1111.  
  1112.     return (void*)fs;
  1113. }
  1114.  
  1115. void r300_mark_fs_code_dirty(struct r300_context *r300)
  1116. {
  1117.     struct r300_fragment_shader* fs = r300_fs(r300);
  1118.  
  1119.     r300_mark_atom_dirty(r300, &r300->fs);
  1120.     r300_mark_atom_dirty(r300, &r300->fs_rc_constant_state);
  1121.     r300_mark_atom_dirty(r300, &r300->fs_constants);
  1122.     r300->fs.size = fs->shader->cb_code_size;
  1123.  
  1124.     if (r300->screen->caps.is_r500) {
  1125.         r300->fs_rc_constant_state.size = fs->shader->rc_state_count * 7;
  1126.         r300->fs_constants.size = fs->shader->externals_count * 4 + 3;
  1127.     } else {
  1128.         r300->fs_rc_constant_state.size = fs->shader->rc_state_count * 5;
  1129.         r300->fs_constants.size = fs->shader->externals_count * 4 + 1;
  1130.     }
  1131.  
  1132.     ((struct r300_constant_buffer*)r300->fs_constants.state)->remap_table =
  1133.             fs->shader->code.constants_remap_table;
  1134. }
  1135.  
  1136. /* Bind fragment shader state. */
  1137. static void r300_bind_fs_state(struct pipe_context* pipe, void* shader)
  1138. {
  1139.     struct r300_context* r300 = r300_context(pipe);
  1140.     struct r300_fragment_shader* fs = (struct r300_fragment_shader*)shader;
  1141.  
  1142.     if (fs == NULL) {
  1143.         r300->fs.state = NULL;
  1144.         return;
  1145.     }
  1146.  
  1147.     r300->fs.state = fs;
  1148.     r300->fs_status = FRAGMENT_SHADER_DIRTY;
  1149.  
  1150.     r300_mark_atom_dirty(r300, &r300->rs_block_state); /* Will be updated before the emission. */
  1151. }
  1152.  
  1153. /* Delete fragment shader state. */
  1154. static void r300_delete_fs_state(struct pipe_context* pipe, void* shader)
  1155. {
  1156.     struct r300_fragment_shader* fs = (struct r300_fragment_shader*)shader;
  1157.     struct r300_fragment_shader_code *tmp, *ptr = fs->first;
  1158.  
  1159.     while (ptr) {
  1160.         tmp = ptr;
  1161.         ptr = ptr->next;
  1162.         rc_constants_destroy(&tmp->code.constants);
  1163.         FREE(tmp->cb_code);
  1164.         FREE(tmp);
  1165.     }
  1166.     FREE((void*)fs->state.tokens);
  1167.     FREE(shader);
  1168. }
  1169.  
  1170. static void r300_set_polygon_stipple(struct pipe_context* pipe,
  1171.                                      const struct pipe_poly_stipple* state)
  1172. {
  1173.     /* XXX no idea how to set this up, but not terribly important */
  1174. }
  1175.  
  1176. /* Create a new rasterizer state based on the CSO rasterizer state.
  1177.  *
  1178.  * This is a very large chunk of state, and covers most of the graphics
  1179.  * backend (GB), geometry assembly (GA), and setup unit (SU) blocks.
  1180.  *
  1181.  * In a not entirely unironic sidenote, this state has nearly nothing to do
  1182.  * with the actual block on the Radeon called the rasterizer (RS). */
  1183. static void* r300_create_rs_state(struct pipe_context* pipe,
  1184.                                   const struct pipe_rasterizer_state* state)
  1185. {
  1186.     struct r300_rs_state* rs = CALLOC_STRUCT(r300_rs_state);
  1187.     uint32_t vap_control_status;    /* R300_VAP_CNTL_STATUS: 0x2140 */
  1188.     uint32_t vap_clip_cntl;         /* R300_VAP_CLIP_CNTL: 0x221C */
  1189.     uint32_t point_size;            /* R300_GA_POINT_SIZE: 0x421c */
  1190.     uint32_t point_minmax;          /* R300_GA_POINT_MINMAX: 0x4230 */
  1191.     uint32_t line_control;          /* R300_GA_LINE_CNTL: 0x4234 */
  1192.     uint32_t polygon_offset_enable; /* R300_SU_POLY_OFFSET_ENABLE: 0x42b4 */
  1193.     uint32_t cull_mode;             /* R300_SU_CULL_MODE: 0x42b8 */
  1194.     uint32_t line_stipple_config;   /* R300_GA_LINE_STIPPLE_CONFIG: 0x4328 */
  1195.     uint32_t line_stipple_value;    /* R300_GA_LINE_STIPPLE_VALUE: 0x4260 */
  1196.     uint32_t polygon_mode;          /* R300_GA_POLY_MODE: 0x4288 */
  1197.     uint32_t clip_rule;             /* R300_SC_CLIP_RULE: 0x43D0 */
  1198.     uint32_t round_mode;            /* R300_GA_ROUND_MODE: 0x428c */
  1199.  
  1200.     /* Point sprites texture coordinates, 0: lower left, 1: upper right */
  1201.     float point_texcoord_left = 0;  /* R300_GA_POINT_S0: 0x4200 */
  1202.     float point_texcoord_bottom = 0;/* R300_GA_POINT_T0: 0x4204 */
  1203.     float point_texcoord_right = 1; /* R300_GA_POINT_S1: 0x4208 */
  1204.     float point_texcoord_top = 0;   /* R300_GA_POINT_T1: 0x420c */
  1205.     boolean vclamp = !r300_context(pipe)->screen->caps.is_r500;
  1206.     CB_LOCALS;
  1207.  
  1208.     /* Copy rasterizer state. */
  1209.     rs->rs = *state;
  1210.     rs->rs_draw = *state;
  1211.  
  1212.     rs->rs.sprite_coord_enable = state->point_quad_rasterization *
  1213.                                  state->sprite_coord_enable;
  1214.  
  1215.     /* Override some states for Draw. */
  1216.     rs->rs_draw.sprite_coord_enable = 0; /* We can do this in HW. */
  1217.     rs->rs_draw.offset_point = 0;
  1218.     rs->rs_draw.offset_line = 0;
  1219.     rs->rs_draw.offset_tri = 0;
  1220.     rs->rs_draw.offset_clamp = 0;
  1221.  
  1222. #ifdef PIPE_ARCH_LITTLE_ENDIAN
  1223.     vap_control_status = R300_VC_NO_SWAP;
  1224. #else
  1225.     vap_control_status = R300_VC_32BIT_SWAP;
  1226. #endif
  1227.  
  1228.     /* If no TCL engine is present, turn off the HW TCL. */
  1229.     if (!r300_screen(pipe->screen)->caps.has_tcl) {
  1230.         vap_control_status |= R300_VAP_TCL_BYPASS;
  1231.     }
  1232.  
  1233.     /* Point size width and height. */
  1234.     point_size =
  1235.         pack_float_16_6x(state->point_size) |
  1236.         (pack_float_16_6x(state->point_size) << R300_POINTSIZE_X_SHIFT);
  1237.  
  1238.     /* Point size clamping. */
  1239.     if (state->point_size_per_vertex) {
  1240.         /* Per-vertex point size.
  1241.          * Clamp to [0, max FB size] */
  1242.         float min_psiz = util_get_min_point_size(state);
  1243.         float max_psiz = pipe->screen->get_paramf(pipe->screen,
  1244.                                         PIPE_CAPF_MAX_POINT_WIDTH);
  1245.         point_minmax =
  1246.             (pack_float_16_6x(min_psiz) << R300_GA_POINT_MINMAX_MIN_SHIFT) |
  1247.             (pack_float_16_6x(max_psiz) << R300_GA_POINT_MINMAX_MAX_SHIFT);
  1248.     } else {
  1249.         /* We cannot disable the point-size vertex output,
  1250.          * so clamp it. */
  1251.         float psiz = state->point_size;
  1252.         point_minmax =
  1253.             (pack_float_16_6x(psiz) << R300_GA_POINT_MINMAX_MIN_SHIFT) |
  1254.             (pack_float_16_6x(psiz) << R300_GA_POINT_MINMAX_MAX_SHIFT);
  1255.     }
  1256.  
  1257.     /* Line control. */
  1258.     line_control = pack_float_16_6x(state->line_width) |
  1259.         R300_GA_LINE_CNTL_END_TYPE_COMP;
  1260.  
  1261.     /* Enable polygon mode */
  1262.     polygon_mode = 0;
  1263.     if (state->fill_front != PIPE_POLYGON_MODE_FILL ||
  1264.         state->fill_back != PIPE_POLYGON_MODE_FILL) {
  1265.         polygon_mode = R300_GA_POLY_MODE_DUAL;
  1266.     }
  1267.  
  1268.     /* Front face */
  1269.     if (state->front_ccw)
  1270.         cull_mode = R300_FRONT_FACE_CCW;
  1271.     else
  1272.         cull_mode = R300_FRONT_FACE_CW;
  1273.  
  1274.     /* Polygon offset */
  1275.     polygon_offset_enable = 0;
  1276.     if (util_get_offset(state, state->fill_front)) {
  1277.        polygon_offset_enable |= R300_FRONT_ENABLE;
  1278.     }
  1279.     if (util_get_offset(state, state->fill_back)) {
  1280.        polygon_offset_enable |= R300_BACK_ENABLE;
  1281.     }
  1282.  
  1283.     rs->polygon_offset_enable = polygon_offset_enable != 0;
  1284.  
  1285.     /* Polygon mode */
  1286.     if (polygon_mode) {
  1287.        polygon_mode |=
  1288.           r300_translate_polygon_mode_front(state->fill_front);
  1289.        polygon_mode |=
  1290.           r300_translate_polygon_mode_back(state->fill_back);
  1291.     }
  1292.  
  1293.     if (state->cull_face & PIPE_FACE_FRONT) {
  1294.         cull_mode |= R300_CULL_FRONT;
  1295.     }
  1296.     if (state->cull_face & PIPE_FACE_BACK) {
  1297.         cull_mode |= R300_CULL_BACK;
  1298.     }
  1299.  
  1300.     if (state->line_stipple_enable) {
  1301.         line_stipple_config =
  1302.             R300_GA_LINE_STIPPLE_CONFIG_LINE_RESET_LINE |
  1303.             (fui((float)state->line_stipple_factor) &
  1304.                 R300_GA_LINE_STIPPLE_CONFIG_STIPPLE_SCALE_MASK);
  1305.         /* XXX this might need to be scaled up */
  1306.         line_stipple_value = state->line_stipple_pattern;
  1307.     } else {
  1308.         line_stipple_config = 0;
  1309.         line_stipple_value = 0;
  1310.     }
  1311.  
  1312.     if (state->flatshade) {
  1313.         rs->color_control = R300_SHADE_MODEL_FLAT;
  1314.     } else {
  1315.         rs->color_control = R300_SHADE_MODEL_SMOOTH;
  1316.     }
  1317.  
  1318.     clip_rule = state->scissor ? 0xAAAA : 0xFFFF;
  1319.  
  1320.     /* Point sprites coord mode */
  1321.     if (rs->rs.sprite_coord_enable) {
  1322.         switch (state->sprite_coord_mode) {
  1323.             case PIPE_SPRITE_COORD_UPPER_LEFT:
  1324.                 point_texcoord_top = 0.0f;
  1325.                 point_texcoord_bottom = 1.0f;
  1326.                 break;
  1327.             case PIPE_SPRITE_COORD_LOWER_LEFT:
  1328.                 point_texcoord_top = 1.0f;
  1329.                 point_texcoord_bottom = 0.0f;
  1330.                 break;
  1331.         }
  1332.     }
  1333.  
  1334.     if (r300_screen(pipe->screen)->caps.has_tcl) {
  1335.        vap_clip_cntl = (state->clip_plane_enable & 63) |
  1336.                        R300_PS_UCP_MODE_CLIP_AS_TRIFAN;
  1337.     } else {
  1338.        vap_clip_cntl = R300_CLIP_DISABLE;
  1339.     }
  1340.  
  1341.     /* Vertex color clamping. FP20 means no clamping. */
  1342.     round_mode =
  1343.       R300_GA_ROUND_MODE_GEOMETRY_ROUND_NEAREST |
  1344.       (!vclamp ? (R300_GA_ROUND_MODE_RGB_CLAMP_FP20 |
  1345.                   R300_GA_ROUND_MODE_ALPHA_CLAMP_FP20) : 0);
  1346.  
  1347.     /* Build the main command buffer. */
  1348.     BEGIN_CB(rs->cb_main, RS_STATE_MAIN_SIZE);
  1349.     OUT_CB_REG(R300_VAP_CNTL_STATUS, vap_control_status);
  1350.     OUT_CB_REG(R300_VAP_CLIP_CNTL, vap_clip_cntl);
  1351.     OUT_CB_REG(R300_GA_POINT_SIZE, point_size);
  1352.     OUT_CB_REG_SEQ(R300_GA_POINT_MINMAX, 2);
  1353.     OUT_CB(point_minmax);
  1354.     OUT_CB(line_control);
  1355.     OUT_CB_REG_SEQ(R300_SU_POLY_OFFSET_ENABLE, 2);
  1356.     OUT_CB(polygon_offset_enable);
  1357.     rs->cull_mode_index = 11;
  1358.     OUT_CB(cull_mode);
  1359.     OUT_CB_REG(R300_GA_LINE_STIPPLE_CONFIG, line_stipple_config);
  1360.     OUT_CB_REG(R300_GA_LINE_STIPPLE_VALUE, line_stipple_value);
  1361.     OUT_CB_REG(R300_GA_POLY_MODE, polygon_mode);
  1362.     OUT_CB_REG(R300_GA_ROUND_MODE, round_mode);
  1363.     OUT_CB_REG(R300_SC_CLIP_RULE, clip_rule);
  1364.     OUT_CB_REG_SEQ(R300_GA_POINT_S0, 4);
  1365.     OUT_CB_32F(point_texcoord_left);
  1366.     OUT_CB_32F(point_texcoord_bottom);
  1367.     OUT_CB_32F(point_texcoord_right);
  1368.     OUT_CB_32F(point_texcoord_top);
  1369.     END_CB;
  1370.  
  1371.     /* Build the two command buffers for polygon offset setup. */
  1372.     if (polygon_offset_enable) {
  1373.         float scale = state->offset_scale * 12;
  1374.         float offset = state->offset_units * 4;
  1375.  
  1376.         BEGIN_CB(rs->cb_poly_offset_zb16, 5);
  1377.         OUT_CB_REG_SEQ(R300_SU_POLY_OFFSET_FRONT_SCALE, 4);
  1378.         OUT_CB_32F(scale);
  1379.         OUT_CB_32F(offset);
  1380.         OUT_CB_32F(scale);
  1381.         OUT_CB_32F(offset);
  1382.         END_CB;
  1383.  
  1384.         offset = state->offset_units * 2;
  1385.  
  1386.         BEGIN_CB(rs->cb_poly_offset_zb24, 5);
  1387.         OUT_CB_REG_SEQ(R300_SU_POLY_OFFSET_FRONT_SCALE, 4);
  1388.         OUT_CB_32F(scale);
  1389.         OUT_CB_32F(offset);
  1390.         OUT_CB_32F(scale);
  1391.         OUT_CB_32F(offset);
  1392.         END_CB;
  1393.     }
  1394.  
  1395.     return (void*)rs;
  1396. }
  1397.  
  1398. /* Bind rasterizer state. */
  1399. static void r300_bind_rs_state(struct pipe_context* pipe, void* state)
  1400. {
  1401.     struct r300_context* r300 = r300_context(pipe);
  1402.     struct r300_rs_state* rs = (struct r300_rs_state*)state;
  1403.     int last_sprite_coord_enable = r300->sprite_coord_enable;
  1404.     boolean last_two_sided_color = r300->two_sided_color;
  1405.     boolean last_msaa_enable = r300->msaa_enable;
  1406.     boolean last_flatshade = r300->flatshade;
  1407.  
  1408.     if (r300->draw && rs) {
  1409.         draw_set_rasterizer_state(r300->draw, &rs->rs_draw, state);
  1410.     }
  1411.  
  1412.     if (rs) {
  1413.         r300->polygon_offset_enabled = rs->polygon_offset_enable;
  1414.         r300->sprite_coord_enable = rs->rs.sprite_coord_enable;
  1415.         r300->two_sided_color = rs->rs.light_twoside;
  1416.         r300->msaa_enable = rs->rs.multisample;
  1417.         r300->flatshade = rs->rs.flatshade;
  1418.     } else {
  1419.         r300->polygon_offset_enabled = FALSE;
  1420.         r300->sprite_coord_enable = 0;
  1421.         r300->two_sided_color = FALSE;
  1422.         r300->msaa_enable = FALSE;
  1423.         r300->flatshade = FALSE;
  1424.     }
  1425.  
  1426.     UPDATE_STATE(state, r300->rs_state);
  1427.     r300->rs_state.size = RS_STATE_MAIN_SIZE + (r300->polygon_offset_enabled ? 5 : 0);
  1428.  
  1429.     if (last_sprite_coord_enable != r300->sprite_coord_enable ||
  1430.         last_two_sided_color != r300->two_sided_color ||
  1431.         last_flatshade != r300->flatshade) {
  1432.         r300_mark_atom_dirty(r300, &r300->rs_block_state);
  1433.     }
  1434.  
  1435.     if (last_msaa_enable != r300->msaa_enable) {
  1436.         if (r300->alpha_to_coverage) {
  1437.             r300_mark_atom_dirty(r300, &r300->dsa_state);
  1438.         }
  1439.  
  1440.         if (r300->alpha_to_one &&
  1441.             r300->fs_status == FRAGMENT_SHADER_VALID) {
  1442.             r300->fs_status = FRAGMENT_SHADER_MAYBE_DIRTY;
  1443.         }
  1444.     }
  1445. }
  1446.  
  1447. /* Free rasterizer state. */
  1448. static void r300_delete_rs_state(struct pipe_context* pipe, void* state)
  1449. {
  1450.     FREE(state);
  1451. }
  1452.  
  1453. static void*
  1454.         r300_create_sampler_state(struct pipe_context* pipe,
  1455.                                   const struct pipe_sampler_state* state)
  1456. {
  1457.     struct r300_context* r300 = r300_context(pipe);
  1458.     struct r300_sampler_state* sampler = CALLOC_STRUCT(r300_sampler_state);
  1459.     boolean is_r500 = r300->screen->caps.is_r500;
  1460.     int lod_bias;
  1461.  
  1462.     sampler->state = *state;
  1463.  
  1464.     /* r300 doesn't handle CLAMP and MIRROR_CLAMP correctly when either MAG
  1465.      * or MIN filter is NEAREST. Since texwrap produces same results
  1466.      * for CLAMP and CLAMP_TO_EDGE, we use them instead. */
  1467.     if (sampler->state.min_img_filter == PIPE_TEX_FILTER_NEAREST ||
  1468.         sampler->state.mag_img_filter == PIPE_TEX_FILTER_NEAREST) {
  1469.         /* Wrap S. */
  1470.         if (sampler->state.wrap_s == PIPE_TEX_WRAP_CLAMP)
  1471.             sampler->state.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
  1472.         else if (sampler->state.wrap_s == PIPE_TEX_WRAP_MIRROR_CLAMP)
  1473.             sampler->state.wrap_s = PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE;
  1474.  
  1475.         /* Wrap T. */
  1476.         if (sampler->state.wrap_t == PIPE_TEX_WRAP_CLAMP)
  1477.             sampler->state.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
  1478.         else if (sampler->state.wrap_t == PIPE_TEX_WRAP_MIRROR_CLAMP)
  1479.             sampler->state.wrap_t = PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE;
  1480.  
  1481.         /* Wrap R. */
  1482.         if (sampler->state.wrap_r == PIPE_TEX_WRAP_CLAMP)
  1483.             sampler->state.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
  1484.         else if (sampler->state.wrap_r == PIPE_TEX_WRAP_MIRROR_CLAMP)
  1485.             sampler->state.wrap_r = PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE;
  1486.     }
  1487.  
  1488.     sampler->filter0 |=
  1489.         (r300_translate_wrap(sampler->state.wrap_s) << R300_TX_WRAP_S_SHIFT) |
  1490.         (r300_translate_wrap(sampler->state.wrap_t) << R300_TX_WRAP_T_SHIFT) |
  1491.         (r300_translate_wrap(sampler->state.wrap_r) << R300_TX_WRAP_R_SHIFT);
  1492.  
  1493.     sampler->filter0 |= r300_translate_tex_filters(state->min_img_filter,
  1494.                                                    state->mag_img_filter,
  1495.                                                    state->min_mip_filter,
  1496.                                                    state->max_anisotropy > 1);
  1497.  
  1498.     sampler->filter0 |= r300_anisotropy(state->max_anisotropy);
  1499.  
  1500.     /* Unfortunately, r300-r500 don't support floating-point mipmap lods. */
  1501.     /* We must pass these to the merge function to clamp them properly. */
  1502.     sampler->min_lod = (unsigned)MAX2(state->min_lod, 0);
  1503.     sampler->max_lod = (unsigned)MAX2(ceilf(state->max_lod), 0);
  1504.  
  1505.     lod_bias = CLAMP((int)(state->lod_bias * 32 + 1), -(1 << 9), (1 << 9) - 1);
  1506.  
  1507.     sampler->filter1 |= (lod_bias << R300_LOD_BIAS_SHIFT) & R300_LOD_BIAS_MASK;
  1508.  
  1509.     /* This is very high quality anisotropic filtering for R5xx.
  1510.      * It's good for benchmarking the performance of texturing but
  1511.      * in practice we don't want to slow down the driver because it's
  1512.      * a pretty good performance killer. Feel free to play with it. */
  1513.     if (DBG_ON(r300, DBG_ANISOHQ) && is_r500) {
  1514.         sampler->filter1 |= r500_anisotropy(state->max_anisotropy);
  1515.     }
  1516.  
  1517.     /* R500-specific fixups and optimizations */
  1518.     if (r300->screen->caps.is_r500) {
  1519.         sampler->filter1 |= R500_BORDER_FIX;
  1520.     }
  1521.  
  1522.     return (void*)sampler;
  1523. }
  1524.  
  1525. static void r300_bind_sampler_states(struct pipe_context* pipe,
  1526.                                      unsigned count,
  1527.                                      void** states)
  1528. {
  1529.     struct r300_context* r300 = r300_context(pipe);
  1530.     struct r300_textures_state* state =
  1531.         (struct r300_textures_state*)r300->textures_state.state;
  1532.     unsigned tex_units = r300->screen->caps.num_tex_units;
  1533.  
  1534.     if (count > tex_units) {
  1535.         return;
  1536.     }
  1537.  
  1538.     memcpy(state->sampler_states, states, sizeof(void*) * count);
  1539.     state->sampler_state_count = count;
  1540.  
  1541.     r300_mark_atom_dirty(r300, &r300->textures_state);
  1542. }
  1543.  
  1544. static void r300_lacks_vertex_textures(struct pipe_context* pipe,
  1545.                                        unsigned count,
  1546.                                        void** states)
  1547. {
  1548. }
  1549.  
  1550. static void r300_delete_sampler_state(struct pipe_context* pipe, void* state)
  1551. {
  1552.     FREE(state);
  1553. }
  1554.  
  1555. static uint32_t r300_assign_texture_cache_region(unsigned index, unsigned num)
  1556. {
  1557.     /* This looks like a hack, but I believe it's suppose to work like
  1558.      * that. To illustrate how this works, let's assume you have 5 textures.
  1559.      * From docs, 5 and the successive numbers are:
  1560.      *
  1561.      * FOURTH_1     = 5
  1562.      * FOURTH_2     = 6
  1563.      * FOURTH_3     = 7
  1564.      * EIGHTH_0     = 8
  1565.      * EIGHTH_1     = 9
  1566.      *
  1567.      * First 3 textures will get 3/4 of size of the cache, divived evenly
  1568.      * between them. The last 1/4 of the cache must be divided between
  1569.      * the last 2 textures, each will therefore get 1/8 of the cache.
  1570.      * Why not just to use "5 + texture_index" ?
  1571.      *
  1572.      * This simple trick works for all "num" <= 16.
  1573.      */
  1574.     if (num <= 1)
  1575.         return R300_TX_CACHE(R300_TX_CACHE_WHOLE);
  1576.     else
  1577.         return R300_TX_CACHE(num + index);
  1578. }
  1579.  
  1580. static void r300_set_fragment_sampler_views(struct pipe_context* pipe,
  1581.                                             unsigned count,
  1582.                                             struct pipe_sampler_view** views)
  1583. {
  1584.     struct r300_context* r300 = r300_context(pipe);
  1585.     struct r300_textures_state* state =
  1586.         (struct r300_textures_state*)r300->textures_state.state;
  1587.     struct r300_resource *texture;
  1588.     unsigned i, real_num_views = 0, view_index = 0;
  1589.     unsigned tex_units = r300->screen->caps.num_tex_units;
  1590.     boolean dirty_tex = FALSE;
  1591.  
  1592.     if (count > tex_units) {
  1593.         return;
  1594.     }
  1595.  
  1596.     /* Calculate the real number of views. */
  1597.     for (i = 0; i < count; i++) {
  1598.         if (views[i])
  1599.             real_num_views++;
  1600.     }
  1601.  
  1602.     for (i = 0; i < count; i++) {
  1603.         pipe_sampler_view_reference(
  1604.                 (struct pipe_sampler_view**)&state->sampler_views[i],
  1605.                 views[i]);
  1606.  
  1607.         if (!views[i]) {
  1608.             continue;
  1609.         }
  1610.  
  1611.         /* A new sampler view (= texture)... */
  1612.         dirty_tex = TRUE;
  1613.  
  1614.         /* Set the texrect factor in the fragment shader.
  1615.              * Needed for RECT and NPOT fallback. */
  1616.         texture = r300_resource(views[i]->texture);
  1617.         if (texture->tex.is_npot) {
  1618.             r300_mark_atom_dirty(r300, &r300->fs_rc_constant_state);
  1619.         }
  1620.  
  1621.         state->sampler_views[i]->texcache_region =
  1622.                 r300_assign_texture_cache_region(view_index, real_num_views);
  1623.         view_index++;
  1624.     }
  1625.  
  1626.     for (i = count; i < tex_units; i++) {
  1627.         if (state->sampler_views[i]) {
  1628.             pipe_sampler_view_reference(
  1629.                     (struct pipe_sampler_view**)&state->sampler_views[i],
  1630.                     NULL);
  1631.         }
  1632.     }
  1633.  
  1634.     state->sampler_view_count = count;
  1635.  
  1636.     r300_mark_atom_dirty(r300, &r300->textures_state);
  1637.  
  1638.     if (dirty_tex) {
  1639.         r300_mark_atom_dirty(r300, &r300->texture_cache_inval);
  1640.     }
  1641. }
  1642.  
  1643. struct pipe_sampler_view *
  1644. r300_create_sampler_view_custom(struct pipe_context *pipe,
  1645.                          struct pipe_resource *texture,
  1646.                          const struct pipe_sampler_view *templ,
  1647.                          unsigned width0_override,
  1648.                          unsigned height0_override)
  1649. {
  1650.     struct r300_sampler_view *view = CALLOC_STRUCT(r300_sampler_view);
  1651.     struct r300_resource *tex = r300_resource(texture);
  1652.     boolean is_r500 = r300_screen(pipe->screen)->caps.is_r500;
  1653.     boolean dxtc_swizzle = r300_screen(pipe->screen)->caps.dxtc_swizzle;
  1654.  
  1655.     if (view) {
  1656.         unsigned hwformat;
  1657.  
  1658.         view->base = *templ;
  1659.         view->base.reference.count = 1;
  1660.         view->base.context = pipe;
  1661.         view->base.texture = NULL;
  1662.         pipe_resource_reference(&view->base.texture, texture);
  1663.  
  1664.         view->width0_override = width0_override;
  1665.         view->height0_override = height0_override;
  1666.         view->swizzle[0] = templ->swizzle_r;
  1667.         view->swizzle[1] = templ->swizzle_g;
  1668.         view->swizzle[2] = templ->swizzle_b;
  1669.         view->swizzle[3] = templ->swizzle_a;
  1670.  
  1671.         hwformat = r300_translate_texformat(templ->format,
  1672.                                             view->swizzle,
  1673.                                             is_r500,
  1674.                                             dxtc_swizzle);
  1675.  
  1676.         if (hwformat == ~0) {
  1677.             fprintf(stderr, "r300: Ooops. Got unsupported format %s in %s.\n",
  1678.                     util_format_short_name(templ->format), __func__);
  1679.         }
  1680.         assert(hwformat != ~0);
  1681.  
  1682.         r300_texture_setup_format_state(r300_screen(pipe->screen), tex,
  1683.                                         templ->format, 0,
  1684.                                         width0_override, height0_override,
  1685.                                         &view->format);
  1686.         view->format.format1 |= hwformat;
  1687.         if (is_r500) {
  1688.             view->format.format2 |= r500_tx_format_msb_bit(templ->format);
  1689.         }
  1690.     }
  1691.  
  1692.     return (struct pipe_sampler_view*)view;
  1693. }
  1694.  
  1695. static struct pipe_sampler_view *
  1696. r300_create_sampler_view(struct pipe_context *pipe,
  1697.                          struct pipe_resource *texture,
  1698.                          const struct pipe_sampler_view *templ)
  1699. {
  1700.     return r300_create_sampler_view_custom(pipe, texture, templ,
  1701.                                            r300_resource(texture)->tex.width0,
  1702.                                            r300_resource(texture)->tex.height0);
  1703. }
  1704.  
  1705.  
  1706. static void
  1707. r300_sampler_view_destroy(struct pipe_context *pipe,
  1708.                           struct pipe_sampler_view *view)
  1709. {
  1710.    pipe_resource_reference(&view->texture, NULL);
  1711.    FREE(view);
  1712. }
  1713.  
  1714. static void r300_set_sample_mask(struct pipe_context *pipe,
  1715.                                  unsigned mask)
  1716. {
  1717.     struct r300_context* r300 = r300_context(pipe);
  1718.  
  1719.     *((unsigned*)r300->sample_mask.state) = mask;
  1720.  
  1721.     r300_mark_atom_dirty(r300, &r300->sample_mask);
  1722. }
  1723.  
  1724. static void r300_set_scissor_states(struct pipe_context* pipe,
  1725.                                     unsigned start_slot,
  1726.                                     unsigned num_scissors,
  1727.                                     const struct pipe_scissor_state* state)
  1728. {
  1729.     struct r300_context* r300 = r300_context(pipe);
  1730.  
  1731.     memcpy(r300->scissor_state.state, state,
  1732.         sizeof(struct pipe_scissor_state));
  1733.  
  1734.     r300_mark_atom_dirty(r300, &r300->scissor_state);
  1735. }
  1736.  
  1737. static void r300_set_viewport_states(struct pipe_context* pipe,
  1738.                                      unsigned start_slot,
  1739.                                      unsigned num_viewports,
  1740.                                      const struct pipe_viewport_state* state)
  1741. {
  1742.     struct r300_context* r300 = r300_context(pipe);
  1743.     struct r300_viewport_state* viewport =
  1744.         (struct r300_viewport_state*)r300->viewport_state.state;
  1745.  
  1746.     r300->viewport = *state;
  1747.  
  1748.     if (r300->draw) {
  1749.         draw_set_viewport_states(r300->draw, start_slot, num_viewports, state);
  1750.         viewport->vte_control = R300_VTX_XY_FMT | R300_VTX_Z_FMT;
  1751.         return;
  1752.     }
  1753.  
  1754.     /* Do the transform in HW. */
  1755.     viewport->vte_control = R300_VTX_W0_FMT;
  1756.  
  1757.     if (state->scale[0] != 1.0f) {
  1758.         viewport->xscale = state->scale[0];
  1759.         viewport->vte_control |= R300_VPORT_X_SCALE_ENA;
  1760.     }
  1761.     if (state->scale[1] != 1.0f) {
  1762.         viewport->yscale = state->scale[1];
  1763.         viewport->vte_control |= R300_VPORT_Y_SCALE_ENA;
  1764.     }
  1765.     if (state->scale[2] != 1.0f) {
  1766.         viewport->zscale = state->scale[2];
  1767.         viewport->vte_control |= R300_VPORT_Z_SCALE_ENA;
  1768.     }
  1769.     if (state->translate[0] != 0.0f) {
  1770.         viewport->xoffset = state->translate[0];
  1771.         viewport->vte_control |= R300_VPORT_X_OFFSET_ENA;
  1772.     }
  1773.     if (state->translate[1] != 0.0f) {
  1774.         viewport->yoffset = state->translate[1];
  1775.         viewport->vte_control |= R300_VPORT_Y_OFFSET_ENA;
  1776.     }
  1777.     if (state->translate[2] != 0.0f) {
  1778.         viewport->zoffset = state->translate[2];
  1779.         viewport->vte_control |= R300_VPORT_Z_OFFSET_ENA;
  1780.     }
  1781.  
  1782.     r300_mark_atom_dirty(r300, &r300->viewport_state);
  1783.     if (r300->fs.state && r300_fs(r300)->shader &&
  1784.         r300_fs(r300)->shader->inputs.wpos != ATTR_UNUSED) {
  1785.         r300_mark_atom_dirty(r300, &r300->fs_rc_constant_state);
  1786.     }
  1787. }
  1788.  
  1789. static void r300_set_vertex_buffers_hwtcl(struct pipe_context* pipe,
  1790.                                     unsigned start_slot, unsigned count,
  1791.                                     const struct pipe_vertex_buffer* buffers)
  1792. {
  1793.     struct r300_context* r300 = r300_context(pipe);
  1794.  
  1795.     util_set_vertex_buffers_count(r300->vertex_buffer,
  1796.                                   &r300->nr_vertex_buffers,
  1797.                                   buffers, start_slot, count);
  1798.  
  1799.     /* There must be at least one vertex buffer set, otherwise it locks up. */
  1800.     if (!r300->nr_vertex_buffers) {
  1801.         util_set_vertex_buffers_count(r300->vertex_buffer,
  1802.                                       &r300->nr_vertex_buffers,
  1803.                                       &r300->dummy_vb, 0, 1);
  1804.     }
  1805.  
  1806.     r300->vertex_arrays_dirty = TRUE;
  1807. }
  1808.  
  1809. static void r300_set_vertex_buffers_swtcl(struct pipe_context* pipe,
  1810.                                     unsigned start_slot, unsigned count,
  1811.                                     const struct pipe_vertex_buffer* buffers)
  1812. {
  1813.     struct r300_context* r300 = r300_context(pipe);
  1814.     unsigned i;
  1815.  
  1816.     util_set_vertex_buffers_count(r300->vertex_buffer,
  1817.                                   &r300->nr_vertex_buffers,
  1818.                                   buffers, start_slot, count);
  1819.     draw_set_vertex_buffers(r300->draw, start_slot, count, buffers);
  1820.  
  1821.     if (!buffers)
  1822.         return;
  1823.  
  1824.     for (i = 0; i < count; i++) {
  1825.         if (buffers[i].user_buffer) {
  1826.             draw_set_mapped_vertex_buffer(r300->draw, start_slot + i,
  1827.                                           buffers[i].user_buffer, ~0);
  1828.         } else if (buffers[i].buffer) {
  1829.             draw_set_mapped_vertex_buffer(r300->draw, start_slot + i,
  1830.                                           r300_resource(buffers[i].buffer)->malloced_buffer, ~0);
  1831.         }
  1832.     }
  1833. }
  1834.  
  1835. static void r300_set_index_buffer_hwtcl(struct pipe_context* pipe,
  1836.                                         const struct pipe_index_buffer *ib)
  1837. {
  1838.     struct r300_context* r300 = r300_context(pipe);
  1839.  
  1840.     if (ib) {
  1841.         pipe_resource_reference(&r300->index_buffer.buffer, ib->buffer);
  1842.         memcpy(&r300->index_buffer, ib, sizeof(*ib));
  1843.     } else {
  1844.         pipe_resource_reference(&r300->index_buffer.buffer, NULL);
  1845.     }
  1846. }
  1847.  
  1848. static void r300_set_index_buffer_swtcl(struct pipe_context* pipe,
  1849.                                         const struct pipe_index_buffer *ib)
  1850. {
  1851.     struct r300_context* r300 = r300_context(pipe);
  1852.  
  1853.     if (ib) {
  1854.         const void *buf = NULL;
  1855.         if (ib->user_buffer) {
  1856.             buf = ib->user_buffer;
  1857.         } else if (ib->buffer) {
  1858.             buf = r300_resource(ib->buffer)->malloced_buffer;
  1859.         }
  1860.         draw_set_indexes(r300->draw,
  1861.                          (const ubyte *) buf + ib->offset,
  1862.                          ib->index_size, ~0);
  1863.     }
  1864. }
  1865.  
  1866. /* Initialize the PSC tables. */
  1867. static void r300_vertex_psc(struct r300_vertex_element_state *velems)
  1868. {
  1869.     struct r300_vertex_stream_state *vstream = &velems->vertex_stream;
  1870.     uint16_t type, swizzle;
  1871.     enum pipe_format format;
  1872.     unsigned i;
  1873.  
  1874.     /* Vertex shaders have no semantics on their inputs,
  1875.      * so PSC should just route stuff based on the vertex elements,
  1876.      * and not on attrib information. */
  1877.     for (i = 0; i < velems->count; i++) {
  1878.         format = velems->velem[i].src_format;
  1879.  
  1880.         type = r300_translate_vertex_data_type(format);
  1881.         if (type == R300_INVALID_FORMAT) {
  1882.             fprintf(stderr, "r300: Bad vertex format %s.\n",
  1883.                     util_format_short_name(format));
  1884.             assert(0);
  1885.             abort();
  1886.         }
  1887.  
  1888.         type |= i << R300_DST_VEC_LOC_SHIFT;
  1889.         swizzle = r300_translate_vertex_data_swizzle(format);
  1890.  
  1891.         if (i & 1) {
  1892.             vstream->vap_prog_stream_cntl[i >> 1] |= type << 16;
  1893.             vstream->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
  1894.         } else {
  1895.             vstream->vap_prog_stream_cntl[i >> 1] |= type;
  1896.             vstream->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
  1897.         }
  1898.     }
  1899.  
  1900.     /* Set the last vector in the PSC. */
  1901.     if (i) {
  1902.         i -= 1;
  1903.     }
  1904.     vstream->vap_prog_stream_cntl[i >> 1] |=
  1905.         (R300_LAST_VEC << (i & 1 ? 16 : 0));
  1906.  
  1907.     vstream->count = (i >> 1) + 1;
  1908. }
  1909.  
  1910. static void* r300_create_vertex_elements_state(struct pipe_context* pipe,
  1911.                                                unsigned count,
  1912.                                                const struct pipe_vertex_element* attribs)
  1913. {
  1914.     struct r300_vertex_element_state *velems;
  1915.     unsigned i;
  1916.     struct pipe_vertex_element dummy_attrib = {0};
  1917.  
  1918.     /* R300 Programmable Stream Control (PSC) doesn't support 0 vertex elements. */
  1919.     if (!count) {
  1920.         dummy_attrib.src_format = PIPE_FORMAT_R8G8B8A8_UNORM;
  1921.         attribs = &dummy_attrib;
  1922.         count = 1;
  1923.     } else if (count > 16) {
  1924.         fprintf(stderr, "r300: More than 16 vertex elements are not supported,"
  1925.                 " requested %i, using 16.\n", count);
  1926.         count = 16;
  1927.     }
  1928.  
  1929.     velems = CALLOC_STRUCT(r300_vertex_element_state);
  1930.     if (!velems)
  1931.         return NULL;
  1932.  
  1933.     velems->count = count;
  1934.     memcpy(velems->velem, attribs, sizeof(struct pipe_vertex_element) * count);
  1935.  
  1936.     if (r300_screen(pipe->screen)->caps.has_tcl) {
  1937.         /* Setup PSC.
  1938.          * The unused components will be replaced by (..., 0, 1). */
  1939.         r300_vertex_psc(velems);
  1940.  
  1941.         for (i = 0; i < count; i++) {
  1942.             velems->format_size[i] =
  1943.                 align(util_format_get_blocksize(velems->velem[i].src_format), 4);
  1944.             velems->vertex_size_dwords += velems->format_size[i] / 4;
  1945.         }
  1946.     }
  1947.  
  1948.     return velems;
  1949. }
  1950.  
  1951. static void r300_bind_vertex_elements_state(struct pipe_context *pipe,
  1952.                                             void *state)
  1953. {
  1954.     struct r300_context *r300 = r300_context(pipe);
  1955.     struct r300_vertex_element_state *velems = state;
  1956.  
  1957.     if (velems == NULL) {
  1958.         return;
  1959.     }
  1960.  
  1961.     r300->velems = velems;
  1962.  
  1963.     if (r300->draw) {
  1964.         draw_set_vertex_elements(r300->draw, velems->count, velems->velem);
  1965.         return;
  1966.     }
  1967.  
  1968.     UPDATE_STATE(&velems->vertex_stream, r300->vertex_stream_state);
  1969.     r300->vertex_stream_state.size = (1 + velems->vertex_stream.count) * 2;
  1970.     r300->vertex_arrays_dirty = TRUE;
  1971. }
  1972.  
  1973. static void r300_delete_vertex_elements_state(struct pipe_context *pipe, void *state)
  1974. {
  1975.     FREE(state);
  1976. }
  1977.  
  1978. static void* r300_create_vs_state(struct pipe_context* pipe,
  1979.                                   const struct pipe_shader_state* shader)
  1980. {
  1981.     struct r300_context* r300 = r300_context(pipe);
  1982.     struct r300_vertex_shader* vs = CALLOC_STRUCT(r300_vertex_shader);
  1983.  
  1984.     /* Copy state directly into shader. */
  1985.     vs->state = *shader;
  1986.     vs->state.tokens = tgsi_dup_tokens(shader->tokens);
  1987.  
  1988.     if (r300->screen->caps.has_tcl) {
  1989.         r300_init_vs_outputs(r300, vs);
  1990.         r300_translate_vertex_shader(r300, vs);
  1991.     } else {
  1992.         r300_draw_init_vertex_shader(r300, vs);
  1993.     }
  1994.  
  1995.     return vs;
  1996. }
  1997.  
  1998. static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
  1999. {
  2000.     struct r300_context* r300 = r300_context(pipe);
  2001.     struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
  2002.  
  2003.     if (vs == NULL) {
  2004.         r300->vs_state.state = NULL;
  2005.         return;
  2006.     }
  2007.     if (vs == r300->vs_state.state) {
  2008.         return;
  2009.     }
  2010.     r300->vs_state.state = vs;
  2011.  
  2012.     /* The majority of the RS block bits is dependent on the vertex shader. */
  2013.     r300_mark_atom_dirty(r300, &r300->rs_block_state); /* Will be updated before the emission. */
  2014.  
  2015.     if (r300->screen->caps.has_tcl) {
  2016.         unsigned fc_op_dwords = r300->screen->caps.is_r500 ? 3 : 2;
  2017.         r300_mark_atom_dirty(r300, &r300->vs_state);
  2018.         r300->vs_state.size = vs->code.length + 9 +
  2019.                         (R300_VS_MAX_FC_OPS * fc_op_dwords + 4);
  2020.  
  2021.         r300_mark_atom_dirty(r300, &r300->vs_constants);
  2022.         r300->vs_constants.size =
  2023.                 2 +
  2024.                 (vs->externals_count ? vs->externals_count * 4 + 3 : 0) +
  2025.                 (vs->immediates_count ? vs->immediates_count * 4 + 3 : 0);
  2026.  
  2027.         ((struct r300_constant_buffer*)r300->vs_constants.state)->remap_table =
  2028.                 vs->code.constants_remap_table;
  2029.  
  2030.         r300_mark_atom_dirty(r300, &r300->pvs_flush);
  2031.     } else {
  2032.         draw_bind_vertex_shader(r300->draw,
  2033.                 (struct draw_vertex_shader*)vs->draw_vs);
  2034.     }
  2035. }
  2036.  
  2037. static void r300_delete_vs_state(struct pipe_context* pipe, void* shader)
  2038. {
  2039.     struct r300_context* r300 = r300_context(pipe);
  2040.     struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
  2041.  
  2042.     if (r300->screen->caps.has_tcl) {
  2043.         rc_constants_destroy(&vs->code.constants);
  2044.         FREE(vs->code.constants_remap_table);
  2045.     } else {
  2046.         draw_delete_vertex_shader(r300->draw,
  2047.                 (struct draw_vertex_shader*)vs->draw_vs);
  2048.     }
  2049.  
  2050.     FREE((void*)vs->state.tokens);
  2051.     FREE(shader);
  2052. }
  2053.  
  2054. static void r300_set_constant_buffer(struct pipe_context *pipe,
  2055.                                      uint shader, uint index,
  2056.                                      struct pipe_constant_buffer *cb)
  2057. {
  2058.     struct r300_context* r300 = r300_context(pipe);
  2059.     struct r300_constant_buffer *cbuf;
  2060.     uint32_t *mapped;
  2061.  
  2062.     if (!cb || (!cb->buffer && !cb->user_buffer))
  2063.         return;
  2064.  
  2065.     switch (shader) {
  2066.         case PIPE_SHADER_VERTEX:
  2067.             cbuf = (struct r300_constant_buffer*)r300->vs_constants.state;
  2068.             break;
  2069.         case PIPE_SHADER_FRAGMENT:
  2070.             cbuf = (struct r300_constant_buffer*)r300->fs_constants.state;
  2071.             break;
  2072.         default:
  2073.             return;
  2074.     }
  2075.  
  2076.  
  2077.     if (cb->user_buffer)
  2078.         mapped = (uint32_t*)cb->user_buffer;
  2079.     else {
  2080.         struct r300_resource *rbuf = r300_resource(cb->buffer);
  2081.  
  2082.         if (rbuf && rbuf->malloced_buffer)
  2083.             mapped = (uint32_t*)rbuf->malloced_buffer;
  2084.         else
  2085.             return;
  2086.     }
  2087.  
  2088.     if (shader == PIPE_SHADER_FRAGMENT ||
  2089.         (shader == PIPE_SHADER_VERTEX && r300->screen->caps.has_tcl)) {
  2090.         cbuf->ptr = mapped;
  2091.     }
  2092.  
  2093.     if (shader == PIPE_SHADER_VERTEX) {
  2094.         if (r300->screen->caps.has_tcl) {
  2095.             struct r300_vertex_shader *vs =
  2096.                     (struct r300_vertex_shader*)r300->vs_state.state;
  2097.  
  2098.             if (!vs) {
  2099.                 cbuf->buffer_base = 0;
  2100.                 return;
  2101.             }
  2102.  
  2103.             cbuf->buffer_base = r300->vs_const_base;
  2104.             r300->vs_const_base += vs->code.constants.Count;
  2105.             if (r300->vs_const_base > R500_MAX_PVS_CONST_VECS) {
  2106.                 r300->vs_const_base = vs->code.constants.Count;
  2107.                 cbuf->buffer_base = 0;
  2108.                 r300_mark_atom_dirty(r300, &r300->pvs_flush);
  2109.             }
  2110.             r300_mark_atom_dirty(r300, &r300->vs_constants);
  2111.         } else if (r300->draw) {
  2112.             draw_set_mapped_constant_buffer(r300->draw, PIPE_SHADER_VERTEX,
  2113.                 0, mapped, cb->buffer_size);
  2114.         }
  2115.     } else if (shader == PIPE_SHADER_FRAGMENT) {
  2116.         r300_mark_atom_dirty(r300, &r300->fs_constants);
  2117.     }
  2118. }
  2119.  
  2120. static void r300_texture_barrier(struct pipe_context *pipe)
  2121. {
  2122.     struct r300_context *r300 = r300_context(pipe);
  2123.  
  2124.     r300_mark_atom_dirty(r300, &r300->gpu_flush);
  2125.     r300_mark_atom_dirty(r300, &r300->texture_cache_inval);
  2126. }
  2127.  
  2128. void r300_init_state_functions(struct r300_context* r300)
  2129. {
  2130.     r300->context.create_blend_state = r300_create_blend_state;
  2131.     r300->context.bind_blend_state = r300_bind_blend_state;
  2132.     r300->context.delete_blend_state = r300_delete_blend_state;
  2133.  
  2134.     r300->context.set_blend_color = r300_set_blend_color;
  2135.  
  2136.     r300->context.set_clip_state = r300_set_clip_state;
  2137.     r300->context.set_sample_mask = r300_set_sample_mask;
  2138.  
  2139.     r300->context.set_constant_buffer = r300_set_constant_buffer;
  2140.  
  2141.     r300->context.create_depth_stencil_alpha_state = r300_create_dsa_state;
  2142.     r300->context.bind_depth_stencil_alpha_state = r300_bind_dsa_state;
  2143.     r300->context.delete_depth_stencil_alpha_state = r300_delete_dsa_state;
  2144.  
  2145.     r300->context.set_stencil_ref = r300_set_stencil_ref;
  2146.  
  2147.     r300->context.set_framebuffer_state = r300_set_framebuffer_state;
  2148.  
  2149.     r300->context.create_fs_state = r300_create_fs_state;
  2150.     r300->context.bind_fs_state = r300_bind_fs_state;
  2151.     r300->context.delete_fs_state = r300_delete_fs_state;
  2152.  
  2153.     r300->context.set_polygon_stipple = r300_set_polygon_stipple;
  2154.  
  2155.     r300->context.create_rasterizer_state = r300_create_rs_state;
  2156.     r300->context.bind_rasterizer_state = r300_bind_rs_state;
  2157.     r300->context.delete_rasterizer_state = r300_delete_rs_state;
  2158.  
  2159.     r300->context.create_sampler_state = r300_create_sampler_state;
  2160.     r300->context.bind_fragment_sampler_states = r300_bind_sampler_states;
  2161.     r300->context.bind_vertex_sampler_states = r300_lacks_vertex_textures;
  2162.     r300->context.delete_sampler_state = r300_delete_sampler_state;
  2163.  
  2164.     r300->context.set_fragment_sampler_views = r300_set_fragment_sampler_views;
  2165.     r300->context.create_sampler_view = r300_create_sampler_view;
  2166.     r300->context.sampler_view_destroy = r300_sampler_view_destroy;
  2167.  
  2168.     r300->context.set_scissor_states = r300_set_scissor_states;
  2169.  
  2170.     r300->context.set_viewport_states = r300_set_viewport_states;
  2171.  
  2172.     if (r300->screen->caps.has_tcl) {
  2173.         r300->context.set_vertex_buffers = r300_set_vertex_buffers_hwtcl;
  2174.         r300->context.set_index_buffer = r300_set_index_buffer_hwtcl;
  2175.     } else {
  2176.         r300->context.set_vertex_buffers = r300_set_vertex_buffers_swtcl;
  2177.         r300->context.set_index_buffer = r300_set_index_buffer_swtcl;
  2178.     }
  2179.  
  2180.     r300->context.create_vertex_elements_state = r300_create_vertex_elements_state;
  2181.     r300->context.bind_vertex_elements_state = r300_bind_vertex_elements_state;
  2182.     r300->context.delete_vertex_elements_state = r300_delete_vertex_elements_state;
  2183.  
  2184.     r300->context.create_vs_state = r300_create_vs_state;
  2185.     r300->context.bind_vs_state = r300_bind_vs_state;
  2186.     r300->context.delete_vs_state = r300_delete_vs_state;
  2187.  
  2188.     r300->context.texture_barrier = r300_texture_barrier;
  2189. }
  2190.