src/mesa/drivers/dri/i965/brw_vec4_gs.c

   1 /*
   2  * Copyright © 2013 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 /**
  25  * \file brw_vec4_gs.c
  26  *
  27  * State atom for client-programmable geometry shaders, and support code.
  28  */
  29
  30 #include "brw_vec4_gs.h"
  31 #include "brw_context.h"
  32 #include "brw_vec4_gs_visitor.h"
  33 #include "brw_state.h"
  34
  35
  36 static bool
  37 do_gs_prog(struct brw_context *brw,
  38            struct gl_shader_program *prog,
  39            struct brw_geometry_program *gp,
  40            struct brw_gs_prog_key *key)
  41 {
  42    struct brw_stage_state *stage_state = &brw->gs.base;
  43    struct brw_gs_compile c;
  44    memset(&c, 0, sizeof(c));
  45    c.key = *key;
  46    c.gp = gp;
  47
  48    c.prog_data.include_primitive_id =
  49       (gp->program.Base.InputsRead & VARYING_BIT_PRIMITIVE_ID) != 0;
  50
  51    c.prog_data.invocations = gp->program.Invocations;
  52
  53    /* Allocate the references to the uniforms that will end up in the
  54     * prog_data associated with the compiled program, and which will be freed
  55     * by the state cache.
  56     *
  57     * Note: param_count needs to be num_uniform_components * 4, since we add
  58     * padding around uniform values below vec4 size, so the worst case is that
  59     * every uniform is a float which gets padded to the size of a vec4.
  60     */
  61    struct gl_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
  62    int param_count = gs->num_uniform_components * 4;
  63
  64    /* We also upload clip plane data as uniforms */
  65    param_count += MAX_CLIP_PLANES * 4;
  66
  67    c.prog_data.base.base.param =
  68       rzalloc_array(NULL, const float *, param_count);
  69    c.prog_data.base.base.pull_param =
  70       rzalloc_array(NULL, const float *, param_count);
  71
  72    if (gp->program.OutputType == GL_POINTS) {
  73       /* When the output type is points, the geometry shader may output data
  74        * to multiple streams, and EndPrimitive() has no effect.  So we
  75        * configure the hardware to interpret the control data as stream ID.
  76        */
  77       c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID;
  78
  79       /* However, StreamID is not yet supported, so we output zero bits of
  80        * control data per vertex.
  81        */
  82       c.control_data_bits_per_vertex = 0;
  83    } else {
  84       /* When the output type is triangle_strip or line_strip, EndPrimitive()
  85        * may be used to terminate the current strip and start a new one
  86        * (similar to primitive restart), and outputting data to multiple
  87        * streams is not supported.  So we configure the hardware to interpret
  88        * the control data as EndPrimitive information (a.k.a. "cut bits").
  89        */
  90       c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT;
  91
  92       /* We only need to output control data if the shader actually calls
  93        * EndPrimitive().
  94        */
  95       c.control_data_bits_per_vertex = gp->program.UsesEndPrimitive ? 1 : 0;
  96    }
  97    c.control_data_header_size_bits =
  98       gp->program.VerticesOut * c.control_data_bits_per_vertex;
  99
 100    /* 1 HWORD = 32 bytes = 256 bits */
 101    c.prog_data.control_data_header_size_hwords =
 102       ALIGN(c.control_data_header_size_bits, 256) / 256;
 103
 104    GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
 105
 106    /* In order for legacy clipping to work, we need to populate the clip
 107     * distance varying slots whenever clipping is enabled, even if the vertex
 108     * shader doesn't write to gl_ClipDistance.
 109     */
 110    if (c.key.base.userclip_active) {
 111       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 112       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 113    }
 114
 115    brw_compute_vue_map(brw, &c.prog_data.base.vue_map, outputs_written);
 116
 117    /* Compute the output vertex size.
 118     *
 119     * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
 120     * Size (p168):
 121     *
 122     *     [0,62] indicating [1,63] 16B units
 123     *
 124     *     Specifies the size of each vertex stored in the GS output entry
 125     *     (following any Control Header data) as a number of 128-bit units
 126     *     (minus one).
 127     *
 128     *     Programming Restrictions: The vertex size must be programmed as a
 129     *     multiple of 32B units with the following exception: Rendering is
 130     *     disabled (as per SOL stage state) and the vertex size output by the
 131     *     GS thread is 16B.
 132     *
 133     *     If rendering is enabled (as per SOL state) the vertex size must be
 134     *     programmed as a multiple of 32B units. In other words, the only time
 135     *     software can program a vertex size with an odd number of 16B units
 136     *     is when rendering is disabled.
 137     *
 138     * Note: B=bytes in the above text.
 139     *
 140     * It doesn't seem worth the extra trouble to optimize the case where the
 141     * vertex size is 16B (especially since this would require special-casing
 142     * the GEN assembly that writes to the URB).  So we just set the vertex
 143     * size to a multiple of 32B (2 vec4's) in all cases.
 144     *
 145     * The maximum output vertex size is 62*16 = 992 bytes (31 hwords).  We
 146     * budget that as follows:
 147     *
 148     *   512 bytes for varyings (a varying component is 4 bytes and
 149     *             gl_MaxGeometryOutputComponents = 128)
 150     *    16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 151     *             bytes)
 152     *    16 bytes overhead for gl_Position (we allocate it a slot in the VUE
 153     *             even if it's not used)
 154     *    32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 155     *             whenever clip planes are enabled, even if the shader doesn't
 156     *             write to gl_ClipDistance)
 157     *    16 bytes overhead since the VUE size must be a multiple of 32 bytes
 158     *             (see below)--this causes up to 1 VUE slot to be wasted
 159     *   400 bytes available for varying packing overhead
 160     *
 161     * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
 162     * per interpolation type, so this is plenty.
 163     *
 164     */
 165    unsigned output_vertex_size_bytes = c.prog_data.base.vue_map.num_slots * 16;
 166    assert(output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
 167    c.prog_data.output_vertex_size_hwords =
 168       ALIGN(output_vertex_size_bytes, 32) / 32;
 169
 170    /* Compute URB entry size.  The maximum allowed URB entry size is 32k.
 171     * That divides up as follows:
 172     *
 173     *     64 bytes for the control data header (cut indices or StreamID bits)
 174     *   4096 bytes for varyings (a varying component is 4 bytes and
 175     *              gl_MaxGeometryTotalOutputComponents = 1024)
 176     *   4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 177     *              bytes/vertex and gl_MaxGeometryOutputVertices is 256)
 178     *   4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
 179     *              even if it's not used)
 180     *   8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 181     *              whenever clip planes are enabled, even if the shader doesn't
 182     *              write to gl_ClipDistance)
 183     *   4096 bytes overhead since the VUE size must be a multiple of 32
 184     *              bytes (see above)--this causes up to 1 VUE slot to be wasted
 185     *   8128 bytes available for varying packing overhead
 186     *
 187     * Worst-case varying packing overhead is 3/4 of a varying slot per
 188     * interpolation type, which works out to 3072 bytes, so this would allow
 189     * us to accommodate 2 interpolation types without any danger of running
 190     * out of URB space.
 191     *
 192     * In practice, the risk of running out of URB space is very small, since
 193     * the above figures are all worst-case, and most of them scale with the
 194     * number of output vertices.  So we'll just calculate the amount of space
 195     * we need, and if it's too large, fail to compile.
 196     */
 197    unsigned output_size_bytes =
 198       c.prog_data.output_vertex_size_hwords * 32 * gp->program.VerticesOut;
 199    output_size_bytes += 32 * c.prog_data.control_data_header_size_hwords;
 200
 201    /* Broadwell stores "Vertex Count" as a full 8 DWord (32 byte) URB output,
 202     * which comes before the control header.
 203     */
 204    if (brw->gen >= 8)
 205       output_size_bytes += 32;
 206
 207    assert(output_size_bytes >= 1);
 208    if (output_size_bytes > GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES)
 209       return false;
 210
 211    /* URB entry sizes are stored as a multiple of 64 bytes. */
 212    c.prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
 213
 214    c.prog_data.output_topology = prim_to_hw_prim[gp->program.OutputType];
 215
 216    brw_compute_vue_map(brw, &c.input_vue_map, c.key.input_varyings);
 217
 218    /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
 219     * need to program a URB read length of ceiling(num_slots / 2).
 220     */
 221    c.prog_data.base.urb_read_length = (c.input_vue_map.num_slots + 1) / 2;
 222
 223    void *mem_ctx = ralloc_context(NULL);
 224    unsigned program_size;
 225    const unsigned *program =
 226       brw_gs_emit(brw, prog, &c, mem_ctx, &program_size);
 227    if (program == NULL) {
 228       ralloc_free(mem_ctx);
 229       return false;
 230    }
 231
 232    /* Scratch space is used for register spilling */
 233    if (c.base.last_scratch) {
 234       perf_debug("Geometry shader triggered register spilling.  "
 235                  "Try reducing the number of live vec4 values to "
 236                  "improve performance.\n");
 237
 238       c.prog_data.base.total_scratch
 239          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 240
 241       brw_get_scratch_bo(brw, &stage_state->scratch_bo,
 242                          c.prog_data.base.total_scratch * brw->max_gs_threads);
 243    }
 244
 245    brw_upload_cache(&brw->cache, BRW_GS_PROG,
 246                     &c.key, sizeof(c.key),
 247                     program, program_size,
 248                     &c.prog_data, sizeof(c.prog_data),
 249                     &stage_state->prog_offset, &brw->gs.prog_data);
 250    ralloc_free(mem_ctx);
 251
 252    return true;
 253 }
 254
 255
 256 static void
 257 brw_upload_gs_prog(struct brw_context *brw)
 258 {
 259    struct gl_context *ctx = &brw->ctx;
 260    struct brw_stage_state *stage_state = &brw->gs.base;
 261    struct brw_gs_prog_key key;
 262    /* BRW_NEW_GEOMETRY_PROGRAM */
 263    struct brw_geometry_program *gp =
 264       (struct brw_geometry_program *) brw->geometry_program;
 265
 266    if (gp == NULL) {
 267       /* No geometry shader.  Vertex data just passes straight through. */
 268       if (brw->state.dirty.brw & BRW_NEW_VUE_MAP_VS) {
 269          brw->vue_map_geom_out = brw->vue_map_vs;
 270          brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 271       }
 272
 273       /* Other state atoms had better not try to access prog_data, since
 274        * there's no GS program.
 275        */
 276       brw->gs.prog_data = NULL;
 277       brw->gs.base.prog_data = NULL;
 278
 279       return;
 280    }
 281
 282    struct gl_program *prog = &gp->program.Base;
 283
 284    memset(&key, 0, sizeof(key));
 285
 286    key.base.program_string_id = gp->id;
 287    brw_setup_vec4_key_clip_info(brw, &key.base,
 288                                 gp->program.Base.UsesClipDistanceOut);
 289
 290    /* _NEW_LIGHT | _NEW_BUFFERS */
 291    key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
 292
 293    /* _NEW_TEXTURE */
 294    brw_populate_sampler_prog_key_data(ctx, prog, stage_state->sampler_count,
 295                                       &key.base.tex);
 296
 297    /* BRW_NEW_VUE_MAP_VS */
 298    key.input_varyings = brw->vue_map_vs.slots_valid;
 299
 300    if (!brw_search_cache(&brw->cache, BRW_GS_PROG,
 301                          &key, sizeof(key),
 302                          &stage_state->prog_offset, &brw->gs.prog_data)) {
 303       bool success =
 304          do_gs_prog(brw, ctx->Shader.CurrentProgram[MESA_SHADER_GEOMETRY], gp,
 305                     &key);
 306       assert(success);
 307    }
 308    brw->gs.base.prog_data = &brw->gs.prog_data->base.base;
 309
 310    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 311               sizeof(brw->vue_map_geom_out)) != 0) {
 312       brw->vue_map_geom_out = brw->gs.prog_data->base.vue_map;
 313       brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 314    }
 315 }
 316
 317
 318 const struct brw_tracked_state brw_gs_prog = {
 319    .dirty = {
 320       .mesa  = (_NEW_LIGHT | _NEW_BUFFERS | _NEW_TEXTURE),
 321       .brw   = BRW_NEW_GEOMETRY_PROGRAM | BRW_NEW_VUE_MAP_VS,
 322    },
 323    .emit = brw_upload_gs_prog
 324 };
 325
 326
 327 bool
 328 brw_gs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
 329 {
 330    struct brw_context *brw = brw_context(ctx);
 331    struct brw_gs_prog_key key;
 332    uint32_t old_prog_offset = brw->gs.base.prog_offset;
 333    struct brw_gs_prog_data *old_prog_data = brw->gs.prog_data;
 334    bool success;
 335
 336    if (!prog->_LinkedShaders[MESA_SHADER_GEOMETRY])
 337       return true;
 338
 339    struct gl_geometry_program *gp = (struct gl_geometry_program *)
 340       prog->_LinkedShaders[MESA_SHADER_GEOMETRY]->Program;
 341    struct brw_geometry_program *bgp = brw_geometry_program(gp);
 342
 343    memset(&key, 0, sizeof(key));
 344
 345    brw_vec4_setup_prog_key_for_precompile(ctx, &key.base, bgp->id, &gp->Base);
 346
 347    /* Assume that the set of varyings coming in from the vertex shader exactly
 348     * matches what the geometry shader requires.
 349     */
 350    key.input_varyings = gp->Base.InputsRead;
 351
 352    success = do_gs_prog(brw, prog, bgp, &key);
 353
 354    brw->gs.base.prog_offset = old_prog_offset;
 355    brw->gs.prog_data = old_prog_data;
 356
 357    return success;
 358 }
 359
 360
 361 bool
 362 brw_gs_prog_data_compare(const void *in_a, const void *in_b)
 363 {
 364    const struct brw_gs_prog_data *a = in_a;
 365    const struct brw_gs_prog_data *b = in_b;
 366
 367    /* Compare the base structure. */
 368    if (!brw_stage_prog_data_compare(&a->base.base, &b->base.base))
 369       return false;
 370
 371    /* Compare the rest of the struct. */
 372    const unsigned offset = sizeof(struct brw_stage_prog_data);
 373    if (memcmp(((char *) a) + offset, ((char *) b) + offset,
 374               sizeof(struct brw_gs_prog_data) - offset)) {
 375       return false;
 376    }
 377
 378    return true;
 379 }