src/mesa/drivers/dri/i965/brw_curbe.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a copy of this software and associated documentation files (the
   8  "Software"), to deal in the Software without restriction, including
   9  without limitation the rights to use, copy, modify, merge, publish,
  10  distribute, sublicense, and/or sell copies of the Software, and to
  11  permit persons to whom the Software is furnished to do so, subject to
  12  the following conditions:
  13
  14  The above copyright notice and this permission notice (including the
  15  next paragraph) shall be included in all copies or substantial
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keithw@vmware.com>
  30   */
  31
  32 /** @file brw_curbe.c
  33  *
  34  * Push constant handling for gen4/5.
  35  *
  36  * Push constants are constant values (such as GLSL uniforms) that are
  37  * pre-loaded into a shader stage's register space at thread spawn time.  On
  38  * gen4 and gen5, we create a blob in memory containing all the push constants
  39  * for all the stages in order.  At CMD_CONST_BUFFER time that blob is loaded
  40  * into URB space as a constant URB entry (CURBE) so that it can be accessed
  41  * quickly at thread setup time.  Each individual fixed function unit's state
  42  * (brw_vs_state.c for example) tells the hardware which subset of the CURBE
  43  * it wants in its register space, and we calculate those areas here under the
  44  * BRW_NEW_CURBE_OFFSETS state flag.  The brw_urb.c allocation will control
  45  * how many CURBEs can be loaded into the hardware at once before a pipeline
  46  * stall occurs at CMD_CONST_BUFFER time.
  47  *
  48  * On gen6+, constant handling becomes a much simpler set of per-unit state.
  49  * See gen6_upload_vec4_push_constants() in gen6_vs_state.c for that code.
  50  */
  51
  52
  53 #include "main/context.h"
  54 #include "main/macros.h"
  55 #include "main/enums.h"
  56 #include "program/prog_parameter.h"
  57 #include "program/prog_print.h"
  58 #include "program/prog_statevars.h"
  59 #include "intel_batchbuffer.h"
  60 #include "intel_buffer_objects.h"
  61 #include "brw_context.h"
  62 #include "brw_defines.h"
  63 #include "brw_state.h"
  64 #include "brw_util.h"
  65
  66
  67 /**
  68  * Partition the CURBE between the various users of constant values.
  69  *
  70  * If the users all fit within the previous allocatation, we avoid changing
  71  * the layout because that means reuploading all unit state and uploading new
  72  * constant buffers.
  73  */
  74 static void calculate_curbe_offsets( struct brw_context *brw )
  75 {
  76    struct gl_context *ctx = &brw->ctx;
  77    /* BRW_NEW_FS_PROG_DATA */
  78    const GLuint nr_fp_regs = (brw->wm.prog_data->base.nr_params + 15) / 16;
  79
  80    /* BRW_NEW_VS_PROG_DATA */
  81    const GLuint nr_vp_regs = (brw->vs.prog_data->base.base.nr_params + 15) / 16;
  82    GLuint nr_clip_regs = 0;
  83    GLuint total_regs;
  84
  85    /* _NEW_TRANSFORM */
  86    if (ctx->Transform.ClipPlanesEnabled) {
  87       GLuint nr_planes = 6 + _mesa_bitcount_64(ctx->Transform.ClipPlanesEnabled);
  88       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  89    }
  90
  91
  92    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  93
  94    /* The CURBE allocation size is limited to 32 512-bit units (128 EU
  95     * registers, or 1024 floats).  See CS_URB_STATE in the gen4 or gen5
  96     * (volume 1, part 1) PRMs.
  97     *
  98     * Note that in brw_fs.cpp we're only loading up to 16 EU registers of
  99     * values as push constants before spilling to pull constants, and in
 100     * brw_vec4.cpp we're loading up to 32 registers of push constants.  An EU
 101     * register is 1/2 of one of these URB entry units, so that leaves us 16 EU
 102     * regs for clip.
 103     */
 104    assert(total_regs <= 32);
 105
 106    /* Lazy resize:
 107     */
 108    if (nr_fp_regs > brw->curbe.wm_size ||
 109        nr_vp_regs > brw->curbe.vs_size ||
 110        nr_clip_regs != brw->curbe.clip_size ||
 111        (total_regs < brw->curbe.total_size / 4 &&
 112         brw->curbe.total_size > 16)) {
 113
 114       GLuint reg = 0;
 115
 116       /* Calculate a new layout:
 117        */
 118       reg = 0;
 119       brw->curbe.wm_start = reg;
 120       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
 121       brw->curbe.clip_start = reg;
 122       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
 123       brw->curbe.vs_start = reg;
 124       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
 125       brw->curbe.total_size = reg;
 126
 127       if (0)
 128          fprintf(stderr, "curbe wm %d+%d clip %d+%d vs %d+%d\n",
 129                  brw->curbe.wm_start,
 130                  brw->curbe.wm_size,
 131                  brw->curbe.clip_start,
 132                  brw->curbe.clip_size,
 133                  brw->curbe.vs_start,
 134                  brw->curbe.vs_size );
 135
 136       brw->ctx.NewDriverState |= BRW_NEW_CURBE_OFFSETS;
 137    }
 138 }
 139
 140
 141 const struct brw_tracked_state brw_curbe_offsets = {
 142    .dirty = {
 143       .mesa = _NEW_TRANSFORM,
 144       .brw  = BRW_NEW_CONTEXT |
 145               BRW_NEW_FS_PROG_DATA |
 146               BRW_NEW_VS_PROG_DATA,
 147    },
 148    .emit = calculate_curbe_offsets
 149 };
 150
 151
 152
 153
 154 /** Uploads the CS_URB_STATE packet.
 155  *
 156  * Just like brw_vs_state.c and brw_wm_state.c define a URB entry size and
 157  * number of entries for their stages, constant buffers do so using this state
 158  * packet.  Having multiple CURBEs in the URB at the same time allows the
 159  * hardware to avoid a pipeline stall between primitives using different
 160  * constant buffer contents.
 161  */
 162 void brw_upload_cs_urb_state(struct brw_context *brw)
 163 {
 164    BEGIN_BATCH(2);
 165    OUT_BATCH(CMD_CS_URB_STATE << 16 | (2-2));
 166
 167    /* BRW_NEW_URB_FENCE */
 168    if (brw->urb.csize == 0) {
 169       OUT_BATCH(0);
 170    } else {
 171       /* BRW_NEW_URB_FENCE */
 172       assert(brw->urb.nr_cs_entries);
 173       OUT_BATCH((brw->urb.csize - 1) << 4 | brw->urb.nr_cs_entries);
 174    }
 175    ADVANCE_BATCH();
 176 }
 177
 178 static const GLfloat fixed_plane[6][4] = {
 179    { 0,    0,   -1, 1 },
 180    { 0,    0,    1, 1 },
 181    { 0,   -1,    0, 1 },
 182    { 0,    1,    0, 1 },
 183    {-1,    0,    0, 1 },
 184    { 1,    0,    0, 1 }
 185 };
 186
 187 /**
 188  * Gathers together all the uniform values into a block of memory to be
 189  * uploaded into the CURBE, then emits the state packet telling the hardware
 190  * the new location.
 191  */
 192 static void
 193 brw_upload_constant_buffer(struct brw_context *brw)
 194 {
 195    struct gl_context *ctx = &brw->ctx;
 196    /* BRW_NEW_CURBE_OFFSETS */
 197    const GLuint sz = brw->curbe.total_size;
 198    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 199    gl_constant_value *buf;
 200    GLuint i;
 201    gl_clip_plane *clip_planes;
 202
 203    if (sz == 0) {
 204       goto emit;
 205    }
 206
 207    buf = intel_upload_space(brw, bufsz, 64,
 208                             &brw->curbe.curbe_bo, &brw->curbe.curbe_offset);
 209
 210    STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
 211
 212    /* fragment shader constants */
 213    if (brw->curbe.wm_size) {
 214       _mesa_load_state_parameters(ctx, brw->fragment_program->Base.Parameters);
 215
 216       /* BRW_NEW_CURBE_OFFSETS */
 217       GLuint offset = brw->curbe.wm_start * 16;
 218
 219       /* BRW_NEW_FS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
 220       for (i = 0; i < brw->wm.prog_data->base.nr_params; i++) {
 221          buf[offset + i] = *brw->wm.prog_data->base.param[i];
 222       }
 223    }
 224
 225    /* clipper constants */
 226    if (brw->curbe.clip_size) {
 227       GLuint offset = brw->curbe.clip_start * 16;
 228       GLuint j;
 229
 230       /* If any planes are going this way, send them all this way:
 231        */
 232       for (i = 0; i < 6; i++) {
 233          buf[offset + i * 4 + 0].f = fixed_plane[i][0];
 234          buf[offset + i * 4 + 1].f = fixed_plane[i][1];
 235          buf[offset + i * 4 + 2].f = fixed_plane[i][2];
 236          buf[offset + i * 4 + 3].f = fixed_plane[i][3];
 237       }
 238
 239       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 240        * clip-space:
 241        */
 242       clip_planes = brw_select_clip_planes(ctx);
 243       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 244          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 245             buf[offset + i * 4 + 0].f = clip_planes[j][0];
 246             buf[offset + i * 4 + 1].f = clip_planes[j][1];
 247             buf[offset + i * 4 + 2].f = clip_planes[j][2];
 248             buf[offset + i * 4 + 3].f = clip_planes[j][3];
 249             i++;
 250          }
 251       }
 252    }
 253
 254    /* vertex shader constants */
 255    if (brw->curbe.vs_size) {
 256       _mesa_load_state_parameters(ctx, brw->vertex_program->Base.Parameters);
 257
 258       GLuint offset = brw->curbe.vs_start * 16;
 259
 260       /* BRW_NEW_VS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
 261       for (i = 0; i < brw->vs.prog_data->base.base.nr_params; i++) {
 262          buf[offset + i] = *brw->vs.prog_data->base.base.param[i];
 263       }
 264    }
 265
 266    if (0) {
 267       for (i = 0; i < sz*16; i+=4)
 268          fprintf(stderr, "curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 269                  buf[i+0].f, buf[i+1].f, buf[i+2].f, buf[i+3].f);
 270    }
 271
 272    /* Because this provokes an action (ie copy the constants into the
 273     * URB), it shouldn't be shortcircuited if identical to the
 274     * previous time - because eg. the urb destination may have
 275     * changed, or the urb contents different to last time.
 276     *
 277     * Note that the data referred to is actually copied internally,
 278     * not just used in place according to passed pointer.
 279     *
 280     * It appears that the CS unit takes care of using each available
 281     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 282     * flushes as necessary when doublebuffering of CURBEs isn't
 283     * possible.
 284     */
 285
 286 emit:
 287    /* BRW_NEW_URB_FENCE: From the gen4 PRM, volume 1, section 3.9.8
 288     * (CONSTANT_BUFFER (CURBE Load)):
 289     *
 290     *     "Modifying the CS URB allocation via URB_FENCE invalidates any
 291     *      previous CURBE entries. Therefore software must subsequently
 292     *      [re]issue a CONSTANT_BUFFER command before CURBE data can be used
 293     *      in the pipeline."
 294     */
 295    BEGIN_BATCH(2);
 296    if (brw->curbe.total_size == 0) {
 297       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 298       OUT_BATCH(0);
 299    } else {
 300       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 301       OUT_RELOC(brw->curbe.curbe_bo,
 302                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 303                 (brw->curbe.total_size - 1) + brw->curbe.curbe_offset);
 304    }
 305    ADVANCE_BATCH();
 306
 307    /* Work around a Broadwater/Crestline depth interpolator bug.  The
 308     * following sequence will cause GPU hangs:
 309     *
 310     * 1. Change state so that all depth related fields in CC_STATE are
 311     *    disabled, and in WM_STATE, only "PS Use Source Depth" is enabled.
 312     * 2. Emit a CONSTANT_BUFFER packet.
 313     * 3. Draw via 3DPRIMITIVE.
 314     *
 315     * The recommended workaround is to emit a non-pipelined state change after
 316     * emitting CONSTANT_BUFFER, in order to drain the windowizer pipeline.
 317     *
 318     * We arbitrarily choose 3DSTATE_GLOBAL_DEPTH_CLAMP_OFFSET (as it's small),
 319     * and always emit it when "PS Use Source Depth" is set.  We could be more
 320     * precise, but the additional complexity is probably not worth it.
 321     *
 322     * BRW_NEW_FRAGMENT_PROGRAM
 323     */
 324    if (brw->gen == 4 && !brw->is_g4x &&
 325        (brw->fragment_program->Base.InputsRead & (1 << VARYING_SLOT_POS))) {
 326       BEGIN_BATCH(2);
 327       OUT_BATCH(_3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP << 16 | (2 - 2));
 328       OUT_BATCH(0);
 329       ADVANCE_BATCH();
 330    }
 331 }
 332
 333 const struct brw_tracked_state brw_constant_buffer = {
 334    .dirty = {
 335       .mesa = _NEW_PROGRAM_CONSTANTS,
 336       .brw  = BRW_NEW_BATCH |
 337               BRW_NEW_CURBE_OFFSETS |
 338               BRW_NEW_FRAGMENT_PROGRAM |
 339               BRW_NEW_FS_PROG_DATA |
 340               BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 341               BRW_NEW_URB_FENCE |
 342               BRW_NEW_VS_PROG_DATA,
 343    },
 344    .emit = brw_upload_constant_buffer,
 345 };
 346