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    /* Allocate the references to the uniforms that will end up in the
  49     * prog_data associated with the compiled program, and which will be freed
  50     * by the state cache.
  51     *
  52     * Note: param_count needs to be num_uniform_components * 4, since we add
  53     * padding around uniform values below vec4 size, so the worst case is that
  54     * every uniform is a float which gets padded to the size of a vec4.
  55     */
  56    struct gl_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
  57    int param_count = gs->num_uniform_components * 4;
  58
  59    /* We also upload clip plane data as uniforms */
  60    param_count += MAX_CLIP_PLANES * 4;
  61
  62    c.prog_data.base.param = rzalloc_array(NULL, const float *, param_count);
  63    c.prog_data.base.pull_param = rzalloc_array(NULL, const float *, param_count);
  64
  65    brw_compute_vue_map(brw, &c.prog_data.base.vue_map,
  66                        gp->program.Base.OutputsWritten,
  67                        c.key.base.userclip_active);
  68
  69    /* Compute the output vertex size.
  70     *
  71     * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
  72     * Size (p168):
  73     *
  74     *     [0,62] indicating [1,63] 16B units
  75     *
  76     *     Specifies the size of each vertex stored in the GS output entry
  77     *     (following any Control Header data) as a number of 128-bit units
  78     *     (minus one).
  79     *
  80     *     Programming Restrictions: The vertex size must be programmed as a
  81     *     multiple of 32B units with the following exception: Rendering is
  82     *     disabled (as per SOL stage state) and the vertex size output by the
  83     *     GS thread is 16B.
  84     *
  85     *     If rendering is enabled (as per SOL state) the vertex size must be
  86     *     programmed as a multiple of 32B units. In other words, the only time
  87     *     software can program a vertex size with an odd number of 16B units
  88     *     is when rendering is disabled.
  89     *
  90     * Note: B=bytes in the above text.
  91     *
  92     * It doesn't seem worth the extra trouble to optimize the case where the
  93     * vertex size is 16B (especially since this would require special-casing
  94     * the GEN assembly that writes to the URB).  So we just set the vertex
  95     * size to a multiple of 32B (2 vec4's) in all cases.
  96     *
  97     * The maximum output vertex size is 62*16 = 992 bytes (31 hwords).  We
  98     * budget that as follows:
  99     *
 100     *   512 bytes for varyings (a varying component is 4 bytes and
 101     *             gl_MaxGeometryOutputComponents = 128)
 102     *    16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 103     *             bytes)
 104     *    16 bytes overhead for gl_Position (we allocate it a slot in the VUE
 105     *             even if it's not used)
 106     *    32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 107     *             whenever clip planes are enabled, even if the shader doesn't
 108     *             write to gl_ClipDistance)
 109     *    16 bytes overhead since the VUE size must be a multiple of 32 bytes
 110     *             (see below)--this causes up to 1 VUE slot to be wasted
 111     *   400 bytes available for varying packing overhead
 112     *
 113     * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
 114     * per interpolation type, so this is plenty.
 115     *
 116     */
 117    unsigned output_vertex_size_bytes = c.prog_data.base.vue_map.num_slots * 16;
 118    assert(output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
 119    c.prog_data.output_vertex_size_hwords =
 120       ALIGN(output_vertex_size_bytes, 32) / 32;
 121
 122    /* Compute URB entry size.  The maximum allowed URB entry size is 32k.
 123     * That divides up as follows:
 124     *
 125     *     64 bytes for the control data header (cut indices or StreamID bits)
 126     *   4096 bytes for varyings (a varying component is 4 bytes and
 127     *              gl_MaxGeometryTotalOutputComponents = 1024)
 128     *   4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 129     *              bytes/vertex and gl_MaxGeometryOutputVertices is 256)
 130     *   4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
 131     *              even if it's not used)
 132     *   8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 133     *              whenever clip planes are enabled, even if the shader doesn't
 134     *              write to gl_ClipDistance)
 135     *   4096 bytes overhead since the VUE size must be a multiple of 32
 136     *              bytes (see above)--this causes up to 1 VUE slot to be wasted
 137     *   8128 bytes available for varying packing overhead
 138     *
 139     * Worst-case varying packing overhead is 3/4 of a varying slot per
 140     * interpolation type, which works out to 3072 bytes, so this would allow
 141     * us to accommodate 2 interpolation types without any danger of running
 142     * out of URB space.
 143     *
 144     * In practice, the risk of running out of URB space is very small, since
 145     * the above figures are all worst-case, and most of them scale with the
 146     * number of output vertices.  So we'll just calculate the amount of space
 147     * we need, and if it's too large, fail to compile.
 148     */
 149    unsigned output_size_bytes =
 150       c.prog_data.output_vertex_size_hwords * 32 * gp->program.VerticesOut;
 151
 152    assert(output_size_bytes >= 1);
 153    if (output_size_bytes > GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES)
 154       return false;
 155
 156    /* URB entry sizes are stored as a multiple of 64 bytes. */
 157    c.prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
 158
 159    c.prog_data.output_topology = prim_to_hw_prim[gp->program.OutputType];
 160
 161    /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
 162     * need to program a URB read length of ceiling(num_slots / 2).
 163     */
 164    c.prog_data.base.urb_read_length = (c.key.input_vue_map.num_slots + 1) / 2;
 165
 166    void *mem_ctx = ralloc_context(NULL);
 167    unsigned program_size;
 168    const unsigned *program =
 169       brw_gs_emit(brw, prog, &c, mem_ctx, &program_size);
 170    if (program == NULL) {
 171       ralloc_free(mem_ctx);
 172       return false;
 173    }
 174
 175    /* Scratch space is used for register spilling */
 176    if (c.base.last_scratch) {
 177       perf_debug("Geometry shader triggered register spilling.  "
 178                  "Try reducing the number of live vec4 values to "
 179                  "improve performance.\n");
 180
 181       c.prog_data.base.total_scratch
 182          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 183
 184       brw_get_scratch_bo(brw, &stage_state->scratch_bo,
 185                          c.prog_data.base.total_scratch * brw->max_gs_threads);
 186    }
 187
 188    brw_upload_cache(&brw->cache, BRW_GS_PROG,
 189                     &c.key, sizeof(c.key),
 190                     program, program_size,
 191                     &c.prog_data, sizeof(c.prog_data),
 192                     &stage_state->prog_offset, &brw->gs.prog_data);
 193    ralloc_free(mem_ctx);
 194
 195    return true;
 196 }
 197
 198
 199 static void
 200 brw_upload_gs_prog(struct brw_context *brw)
 201 {
 202    struct gl_context *ctx = &brw->ctx;
 203    struct brw_stage_state *stage_state = &brw->gs.base;
 204    struct brw_gs_prog_key key;
 205    /* BRW_NEW_GEOMETRY_PROGRAM */
 206    struct brw_geometry_program *gp =
 207       (struct brw_geometry_program *) brw->geometry_program;
 208
 209    if (gp == NULL) {
 210       /* No geometry shader.  Vertex data just passes straight through. */
 211       if (brw->state.dirty.brw & BRW_NEW_VUE_MAP_VS) {
 212          brw->vue_map_geom_out = brw->vue_map_vs;
 213          brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 214       }
 215       return;
 216    }
 217
 218    struct gl_program *prog = &gp->program.Base;
 219
 220    memset(&key, 0, sizeof(key));
 221
 222    key.base.program_string_id = gp->id;
 223    brw_setup_vec4_key_clip_info(brw, &key.base, gp->program.UsesClipDistance);
 224
 225    /* _NEW_LIGHT | _NEW_BUFFERS */
 226    key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
 227
 228    /* _NEW_TEXTURE */
 229    brw_populate_sampler_prog_key_data(ctx, prog, stage_state->sampler_count,
 230                                       &key.base.tex);
 231
 232    /* BRW_NEW_VUE_MAP_VS */
 233    key.input_vue_map = brw->vue_map_vs;
 234
 235    if (!brw_search_cache(&brw->cache, BRW_GS_PROG,
 236                          &key, sizeof(key),
 237                          &stage_state->prog_offset, &brw->gs.prog_data)) {
 238       bool success = do_gs_prog(brw, ctx->Shader.CurrentGeometryProgram,
 239                                 gp, &key);
 240       assert(success);
 241    }
 242    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 243               sizeof(brw->vue_map_geom_out)) != 0) {
 244       brw->vue_map_geom_out = brw->gs.prog_data->base.vue_map;
 245       brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 246    }
 247 }
 248
 249
 250 const struct brw_tracked_state brw_gs_prog = {
 251    .dirty = {
 252       .mesa  = (_NEW_LIGHT | _NEW_BUFFERS | _NEW_TEXTURE),
 253       .brw   = BRW_NEW_GEOMETRY_PROGRAM | BRW_NEW_VUE_MAP_VS,
 254    },
 255    .emit = brw_upload_gs_prog
 256 };
 257
 258
 259 bool
 260 brw_gs_prog_data_compare(const void *in_a, const void *in_b,
 261                          int aux_size, const void *in_key)
 262 {
 263    const struct brw_gs_prog_data *a = in_a;
 264    const struct brw_gs_prog_data *b = in_b;
 265
 266    /* Compare the base vec4 structure. */
 267    if (!brw_vec4_prog_data_compare(&a->base, &b->base))
 268       return false;
 269
 270    /* Compare the rest of the struct. */
 271    const unsigned offset = sizeof(struct brw_vec4_prog_data);
 272    if (memcmp(((char *) &a) + offset, ((char *) &b) + offset,
 273               sizeof(struct brw_gs_prog_data) - offset)) {
 274       return false;
 275    }
 276
 277    return true;
 278 }
 279
 280
 281 void
 282 brw_gs_prog_data_free(const void *in_prog_data)
 283 {
 284    const struct brw_gs_prog_data *prog_data = in_prog_data;
 285
 286    brw_vec4_prog_data_free(&prog_data->base);
 287 }