src/mesa/drivers/dri/i965/brw_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_gs.h"
  31 #include "brw_context.h"
  32 #include "brw_vec4_gs_visitor.h"
  33 #include "brw_state.h"
  34 #include "brw_ff_gs.h"
  35 #include "brw_nir.h"
  36
  37 static void
  38 assign_gs_binding_table_offsets(const struct brw_device_info *devinfo,
  39                                 const struct gl_shader_program *shader_prog,
  40                                 const struct gl_program *prog,
  41                                 struct brw_gs_prog_data *prog_data)
  42 {
  43    /* In gen6 we reserve the first BRW_MAX_SOL_BINDINGS entries for transform
  44     * feedback surfaces.
  45     */
  46    uint32_t reserved = devinfo->gen == 6 ? BRW_MAX_SOL_BINDINGS : 0;
  47
  48    brw_assign_common_binding_table_offsets(MESA_SHADER_GEOMETRY, devinfo,
  49                                            shader_prog, prog,
  50                                            &prog_data->base.base,
  51                                            reserved);
  52 }
  53
  54 bool
  55 brw_codegen_gs_prog(struct brw_context *brw,
  56                     struct gl_shader_program *prog,
  57                     struct brw_geometry_program *gp,
  58                     struct brw_gs_prog_key *key)
  59 {
  60    struct gl_shader *shader = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
  61    nir_shader *nir = gp->program.Base.nir;
  62    struct brw_stage_state *stage_state = &brw->gs.base;
  63    struct brw_gs_prog_data prog_data;
  64    struct brw_gs_compile c;
  65    memset(&prog_data, 0, sizeof(prog_data));
  66    memset(&c, 0, sizeof(c));
  67    c.key = *key;
  68
  69    prog_data.include_primitive_id =
  70       (nir->info.inputs_read & VARYING_BIT_PRIMITIVE_ID) != 0;
  71
  72    prog_data.invocations = nir->info.gs.invocations;
  73
  74    assign_gs_binding_table_offsets(brw->intelScreen->devinfo, prog,
  75                                    &gp->program.Base, &prog_data);
  76
  77    /* Allocate the references to the uniforms that will end up in the
  78     * prog_data associated with the compiled program, and which will be freed
  79     * by the state cache.
  80     *
  81     * Note: param_count needs to be num_uniform_components * 4, since we add
  82     * padding around uniform values below vec4 size, so the worst case is that
  83     * every uniform is a float which gets padded to the size of a vec4.
  84     */
  85    struct gl_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
  86    int param_count = gp->program.Base.nir->num_uniforms * 4;
  87
  88    prog_data.base.base.param =
  89       rzalloc_array(NULL, const gl_constant_value *, param_count);
  90    prog_data.base.base.pull_param =
  91       rzalloc_array(NULL, const gl_constant_value *, param_count);
  92    prog_data.base.base.image_param =
  93       rzalloc_array(NULL, struct brw_image_param, gs->NumImages);
  94    prog_data.base.base.nr_params = param_count;
  95    prog_data.base.base.nr_image_params = gs->NumImages;
  96
  97    brw_nir_setup_glsl_uniforms(gp->program.Base.nir, prog, &gp->program.Base,
  98                                &prog_data.base.base, false);
  99
 100    if (brw->gen >= 8) {
 101       prog_data.static_vertex_count =
 102          nir_gs_count_vertices(gp->program.Base.nir);
 103    }
 104
 105    if (brw->gen >= 7) {
 106       if (nir->info.gs.output_primitive == GL_POINTS) {
 107          /* When the output type is points, the geometry shader may output data
 108           * to multiple streams, and EndPrimitive() has no effect.  So we
 109           * configure the hardware to interpret the control data as stream ID.
 110           */
 111          prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID;
 112
 113          /* We only have to emit control bits if we are using streams */
 114          if (nir->info.gs.uses_streams)
 115             c.control_data_bits_per_vertex = 2;
 116          else
 117             c.control_data_bits_per_vertex = 0;
 118       } else {
 119          /* When the output type is triangle_strip or line_strip, EndPrimitive()
 120           * may be used to terminate the current strip and start a new one
 121           * (similar to primitive restart), and outputting data to multiple
 122           * streams is not supported.  So we configure the hardware to interpret
 123           * the control data as EndPrimitive information (a.k.a. "cut bits").
 124           */
 125          prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT;
 126
 127          /* We only need to output control data if the shader actually calls
 128           * EndPrimitive().
 129           */
 130          c.control_data_bits_per_vertex =
 131             nir->info.gs.uses_end_primitive ? 1 : 0;
 132       }
 133    } else {
 134       /* There are no control data bits in gen6. */
 135       c.control_data_bits_per_vertex = 0;
 136
 137       /* If it is using transform feedback, enable it */
 138       if (nir->info.has_transform_feedback_varyings)
 139          prog_data.gen6_xfb_enabled = true;
 140       else
 141          prog_data.gen6_xfb_enabled = false;
 142    }
 143    c.control_data_header_size_bits =
 144       nir->info.gs.vertices_out * c.control_data_bits_per_vertex;
 145
 146    /* 1 HWORD = 32 bytes = 256 bits */
 147    prog_data.control_data_header_size_hwords =
 148       ALIGN(c.control_data_header_size_bits, 256) / 256;
 149
 150    GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
 151
 152    brw_compute_vue_map(brw->intelScreen->devinfo,
 153                        &prog_data.base.vue_map, outputs_written,
 154                        prog ? prog->SeparateShader : false);
 155
 156    /* Compute the output vertex size.
 157     *
 158     * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
 159     * Size (p168):
 160     *
 161     *     [0,62] indicating [1,63] 16B units
 162     *
 163     *     Specifies the size of each vertex stored in the GS output entry
 164     *     (following any Control Header data) as a number of 128-bit units
 165     *     (minus one).
 166     *
 167     *     Programming Restrictions: The vertex size must be programmed as a
 168     *     multiple of 32B units with the following exception: Rendering is
 169     *     disabled (as per SOL stage state) and the vertex size output by the
 170     *     GS thread is 16B.
 171     *
 172     *     If rendering is enabled (as per SOL state) the vertex size must be
 173     *     programmed as a multiple of 32B units. In other words, the only time
 174     *     software can program a vertex size with an odd number of 16B units
 175     *     is when rendering is disabled.
 176     *
 177     * Note: B=bytes in the above text.
 178     *
 179     * It doesn't seem worth the extra trouble to optimize the case where the
 180     * vertex size is 16B (especially since this would require special-casing
 181     * the GEN assembly that writes to the URB).  So we just set the vertex
 182     * size to a multiple of 32B (2 vec4's) in all cases.
 183     *
 184     * The maximum output vertex size is 62*16 = 992 bytes (31 hwords).  We
 185     * budget that as follows:
 186     *
 187     *   512 bytes for varyings (a varying component is 4 bytes and
 188     *             gl_MaxGeometryOutputComponents = 128)
 189     *    16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 190     *             bytes)
 191     *    16 bytes overhead for gl_Position (we allocate it a slot in the VUE
 192     *             even if it's not used)
 193     *    32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 194     *             whenever clip planes are enabled, even if the shader doesn't
 195     *             write to gl_ClipDistance)
 196     *    16 bytes overhead since the VUE size must be a multiple of 32 bytes
 197     *             (see below)--this causes up to 1 VUE slot to be wasted
 198     *   400 bytes available for varying packing overhead
 199     *
 200     * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
 201     * per interpolation type, so this is plenty.
 202     *
 203     */
 204    unsigned output_vertex_size_bytes = prog_data.base.vue_map.num_slots * 16;
 205    assert(brw->gen == 6 ||
 206           output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
 207    prog_data.output_vertex_size_hwords =
 208       ALIGN(output_vertex_size_bytes, 32) / 32;
 209
 210    /* Compute URB entry size.  The maximum allowed URB entry size is 32k.
 211     * That divides up as follows:
 212     *
 213     *     64 bytes for the control data header (cut indices or StreamID bits)
 214     *   4096 bytes for varyings (a varying component is 4 bytes and
 215     *              gl_MaxGeometryTotalOutputComponents = 1024)
 216     *   4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 217     *              bytes/vertex and gl_MaxGeometryOutputVertices is 256)
 218     *   4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
 219     *              even if it's not used)
 220     *   8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 221     *              whenever clip planes are enabled, even if the shader doesn't
 222     *              write to gl_ClipDistance)
 223     *   4096 bytes overhead since the VUE size must be a multiple of 32
 224     *              bytes (see above)--this causes up to 1 VUE slot to be wasted
 225     *   8128 bytes available for varying packing overhead
 226     *
 227     * Worst-case varying packing overhead is 3/4 of a varying slot per
 228     * interpolation type, which works out to 3072 bytes, so this would allow
 229     * us to accommodate 2 interpolation types without any danger of running
 230     * out of URB space.
 231     *
 232     * In practice, the risk of running out of URB space is very small, since
 233     * the above figures are all worst-case, and most of them scale with the
 234     * number of output vertices.  So we'll just calculate the amount of space
 235     * we need, and if it's too large, fail to compile.
 236     *
 237     * The above is for gen7+ where we have a single URB entry that will hold
 238     * all the output. In gen6, we will have to allocate URB entries for every
 239     * vertex we emit, so our URB entries only need to be large enough to hold
 240     * a single vertex. Also, gen6 does not have a control data header.
 241     */
 242    unsigned output_size_bytes;
 243    if (brw->gen >= 7) {
 244       output_size_bytes =
 245          prog_data.output_vertex_size_hwords * 32 * nir->info.gs.vertices_out;
 246       output_size_bytes += 32 * prog_data.control_data_header_size_hwords;
 247    } else {
 248       output_size_bytes = prog_data.output_vertex_size_hwords * 32;
 249    }
 250
 251    /* Broadwell stores "Vertex Count" as a full 8 DWord (32 byte) URB output,
 252     * which comes before the control header.
 253     */
 254    if (brw->gen >= 8)
 255       output_size_bytes += 32;
 256
 257    assert(output_size_bytes >= 1);
 258    int max_output_size_bytes = GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES;
 259    if (brw->gen == 6)
 260       max_output_size_bytes = GEN6_MAX_GS_URB_ENTRY_SIZE_BYTES;
 261    if (output_size_bytes > max_output_size_bytes)
 262       return false;
 263
 264
 265    /* URB entry sizes are stored as a multiple of 64 bytes in gen7+ and
 266     * a multiple of 128 bytes in gen6.
 267     */
 268    if (brw->gen >= 7)
 269       prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
 270    else
 271       prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 128) / 128;
 272
 273    prog_data.output_topology =
 274       get_hw_prim_for_gl_prim(nir->info.gs.output_primitive);
 275
 276    /* The GLSL linker will have already matched up GS inputs and the outputs
 277     * of prior stages.  The driver does extend VS outputs in some cases, but
 278     * only for legacy OpenGL or Gen4-5 hardware, neither of which offer
 279     * geometry shader support.  So we can safely ignore that.
 280     *
 281     * For SSO pipelines, we use a fixed VUE map layout based on variable
 282     * locations, so we can rely on rendezvous-by-location making this work.
 283     *
 284     * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
 285     * written by previous stages and shows up via payload magic.
 286     */
 287    GLbitfield64 inputs_read =
 288       nir->info.inputs_read & ~VARYING_BIT_PRIMITIVE_ID;
 289    brw_compute_vue_map(brw->intelScreen->devinfo,
 290                        &c.input_vue_map, inputs_read,
 291                        nir->info.separate_shader);
 292
 293    /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
 294     * need to program a URB read length of ceiling(num_slots / 2).
 295     */
 296    prog_data.base.urb_read_length = (c.input_vue_map.num_slots + 1) / 2;
 297
 298    if (unlikely(INTEL_DEBUG & DEBUG_GS))
 299       brw_dump_ir("geometry", prog, gs, NULL);
 300
 301    int st_index = -1;
 302    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
 303       st_index = brw_get_shader_time_index(brw, prog, NULL, ST_GS);
 304
 305    void *mem_ctx = ralloc_context(NULL);
 306    unsigned program_size;
 307    char *error_str;
 308    const unsigned *program =
 309       brw_compile_gs(brw->intelScreen->compiler, brw, mem_ctx, &c,
 310                      &prog_data, shader->Program->nir, prog,
 311                      st_index, &program_size, &error_str);
 312    if (program == NULL) {
 313       ralloc_free(mem_ctx);
 314       return false;
 315    }
 316
 317    /* Scratch space is used for register spilling */
 318    if (prog_data.base.base.total_scratch) {
 319       brw_get_scratch_bo(brw, &stage_state->scratch_bo,
 320                          prog_data.base.base.total_scratch *
 321                          brw->max_gs_threads);
 322    }
 323
 324    brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
 325                     &c.key, sizeof(c.key),
 326                     program, program_size,
 327                     &prog_data, sizeof(prog_data),
 328                     &stage_state->prog_offset, &brw->gs.prog_data);
 329    ralloc_free(mem_ctx);
 330
 331    return true;
 332 }
 333
 334 static bool
 335 brw_gs_state_dirty(struct brw_context *brw)
 336 {
 337    return brw_state_dirty(brw,
 338                           _NEW_TEXTURE,
 339                           BRW_NEW_GEOMETRY_PROGRAM |
 340                           BRW_NEW_TRANSFORM_FEEDBACK);
 341 }
 342
 343 static void
 344 brw_gs_populate_key(struct brw_context *brw,
 345                     struct brw_gs_prog_key *key)
 346 {
 347    struct gl_context *ctx = &brw->ctx;
 348    struct brw_stage_state *stage_state = &brw->gs.base;
 349    struct brw_geometry_program *gp =
 350       (struct brw_geometry_program *) brw->geometry_program;
 351    struct gl_program *prog = &gp->program.Base;
 352
 353    memset(key, 0, sizeof(*key));
 354
 355    key->program_string_id = gp->id;
 356
 357    /* _NEW_TEXTURE */
 358    brw_populate_sampler_prog_key_data(ctx, prog, stage_state->sampler_count,
 359                                       &key->tex);
 360 }
 361
 362 void
 363 brw_upload_gs_prog(struct brw_context *brw)
 364 {
 365    struct gl_context *ctx = &brw->ctx;
 366    struct gl_shader_program **current = ctx->_Shader->CurrentProgram;
 367    struct brw_stage_state *stage_state = &brw->gs.base;
 368    struct brw_gs_prog_key key;
 369    /* BRW_NEW_GEOMETRY_PROGRAM */
 370    struct brw_geometry_program *gp =
 371       (struct brw_geometry_program *) brw->geometry_program;
 372
 373    if (!brw_gs_state_dirty(brw))
 374       return;
 375
 376    if (gp == NULL) {
 377       /* No geometry shader.  Vertex data just passes straight through. */
 378       if (brw->gen == 6 &&
 379           (brw->ctx.NewDriverState & BRW_NEW_TRANSFORM_FEEDBACK)) {
 380          gen6_brw_upload_ff_gs_prog(brw);
 381          return;
 382       }
 383
 384       /* Other state atoms had better not try to access prog_data, since
 385        * there's no GS program.
 386        */
 387       brw->gs.prog_data = NULL;
 388       brw->gs.base.prog_data = NULL;
 389
 390       return;
 391    }
 392
 393    brw_gs_populate_key(brw, &key);
 394
 395    if (!brw_search_cache(&brw->cache, BRW_CACHE_GS_PROG,
 396                          &key, sizeof(key),
 397                          &stage_state->prog_offset, &brw->gs.prog_data)) {
 398       bool success = brw_codegen_gs_prog(brw, current[MESA_SHADER_GEOMETRY],
 399                                          gp, &key);
 400       assert(success);
 401       (void)success;
 402    }
 403    brw->gs.base.prog_data = &brw->gs.prog_data->base.base;
 404 }
 405
 406 bool
 407 brw_gs_precompile(struct gl_context *ctx,
 408                   struct gl_shader_program *shader_prog,
 409                   struct gl_program *prog)
 410 {
 411    struct brw_context *brw = brw_context(ctx);
 412    struct brw_gs_prog_key key;
 413    uint32_t old_prog_offset = brw->gs.base.prog_offset;
 414    struct brw_gs_prog_data *old_prog_data = brw->gs.prog_data;
 415    bool success;
 416
 417    struct gl_geometry_program *gp = (struct gl_geometry_program *) prog;
 418    struct brw_geometry_program *bgp = brw_geometry_program(gp);
 419
 420    memset(&key, 0, sizeof(key));
 421
 422    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 423    key.program_string_id = bgp->id;
 424
 425    success = brw_codegen_gs_prog(brw, shader_prog, bgp, &key);
 426
 427    brw->gs.base.prog_offset = old_prog_offset;
 428    brw->gs.prog_data = old_prog_data;
 429
 430    return success;
 431 }