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    if (gp->program.OutputType == GL_POINTS) {
  66       /* When the output type is points, the geometry shader may output data
  67        * to multiple streams, and EndPrimitive() has no effect.  So we
  68        * configure the hardware to interpret the control data as stream ID.
  69        */
  70       c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID;
  71
  72       /* However, StreamID is not yet supported, so we output zero bits of
  73        * control data per vertex.
  74        */
  75       c.control_data_bits_per_vertex = 0;
  76    } else {
  77       /* When the output type is triangle_strip or line_strip, EndPrimitive()
  78        * may be used to terminate the current strip and start a new one
  79        * (similar to primitive restart), and outputting data to multiple
  80        * streams is not supported.  So we configure the hardware to interpret
  81        * the control data as EndPrimitive information (a.k.a. "cut bits").
  82        */
  83       c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT;
  84
  85       /* We only need to output control data if the shader actually calls
  86        * EndPrimitive().
  87        */
  88       c.control_data_bits_per_vertex = gp->program.UsesEndPrimitive ? 1 : 0;
  89    }
  90    c.control_data_header_size_bits =
  91       gp->program.VerticesOut * c.control_data_bits_per_vertex;
  92
  93    /* 1 HWORD = 32 bytes = 256 bits */
  94    c.prog_data.control_data_header_size_hwords =
  95       ALIGN(c.control_data_header_size_bits, 256) / 256;
  96
  97    GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
  98
  99    /* In order for legacy clipping to work, we need to populate the clip
 100     * distance varying slots whenever clipping is enabled, even if the vertex
 101     * shader doesn't write to gl_ClipDistance.
 102     */
 103    if (c.key.base.userclip_active) {
 104       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
 105       outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
 106    }
 107
 108    brw_compute_vue_map(brw, &c.prog_data.base.vue_map, outputs_written);
 109
 110    /* Compute the output vertex size.
 111     *
 112     * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
 113     * Size (p168):
 114     *
 115     *     [0,62] indicating [1,63] 16B units
 116     *
 117     *     Specifies the size of each vertex stored in the GS output entry
 118     *     (following any Control Header data) as a number of 128-bit units
 119     *     (minus one).
 120     *
 121     *     Programming Restrictions: The vertex size must be programmed as a
 122     *     multiple of 32B units with the following exception: Rendering is
 123     *     disabled (as per SOL stage state) and the vertex size output by the
 124     *     GS thread is 16B.
 125     *
 126     *     If rendering is enabled (as per SOL state) the vertex size must be
 127     *     programmed as a multiple of 32B units. In other words, the only time
 128     *     software can program a vertex size with an odd number of 16B units
 129     *     is when rendering is disabled.
 130     *
 131     * Note: B=bytes in the above text.
 132     *
 133     * It doesn't seem worth the extra trouble to optimize the case where the
 134     * vertex size is 16B (especially since this would require special-casing
 135     * the GEN assembly that writes to the URB).  So we just set the vertex
 136     * size to a multiple of 32B (2 vec4's) in all cases.
 137     *
 138     * The maximum output vertex size is 62*16 = 992 bytes (31 hwords).  We
 139     * budget that as follows:
 140     *
 141     *   512 bytes for varyings (a varying component is 4 bytes and
 142     *             gl_MaxGeometryOutputComponents = 128)
 143     *    16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 144     *             bytes)
 145     *    16 bytes overhead for gl_Position (we allocate it a slot in the VUE
 146     *             even if it's not used)
 147     *    32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 148     *             whenever clip planes are enabled, even if the shader doesn't
 149     *             write to gl_ClipDistance)
 150     *    16 bytes overhead since the VUE size must be a multiple of 32 bytes
 151     *             (see below)--this causes up to 1 VUE slot to be wasted
 152     *   400 bytes available for varying packing overhead
 153     *
 154     * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
 155     * per interpolation type, so this is plenty.
 156     *
 157     */
 158    unsigned output_vertex_size_bytes = c.prog_data.base.vue_map.num_slots * 16;
 159    assert(output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
 160    c.prog_data.output_vertex_size_hwords =
 161       ALIGN(output_vertex_size_bytes, 32) / 32;
 162
 163    /* Compute URB entry size.  The maximum allowed URB entry size is 32k.
 164     * That divides up as follows:
 165     *
 166     *     64 bytes for the control data header (cut indices or StreamID bits)
 167     *   4096 bytes for varyings (a varying component is 4 bytes and
 168     *              gl_MaxGeometryTotalOutputComponents = 1024)
 169     *   4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
 170     *              bytes/vertex and gl_MaxGeometryOutputVertices is 256)
 171     *   4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
 172     *              even if it's not used)
 173     *   8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
 174     *              whenever clip planes are enabled, even if the shader doesn't
 175     *              write to gl_ClipDistance)
 176     *   4096 bytes overhead since the VUE size must be a multiple of 32
 177     *              bytes (see above)--this causes up to 1 VUE slot to be wasted
 178     *   8128 bytes available for varying packing overhead
 179     *
 180     * Worst-case varying packing overhead is 3/4 of a varying slot per
 181     * interpolation type, which works out to 3072 bytes, so this would allow
 182     * us to accommodate 2 interpolation types without any danger of running
 183     * out of URB space.
 184     *
 185     * In practice, the risk of running out of URB space is very small, since
 186     * the above figures are all worst-case, and most of them scale with the
 187     * number of output vertices.  So we'll just calculate the amount of space
 188     * we need, and if it's too large, fail to compile.
 189     */
 190    unsigned output_size_bytes =
 191       c.prog_data.output_vertex_size_hwords * 32 * gp->program.VerticesOut;
 192    output_size_bytes += 32 * c.prog_data.control_data_header_size_hwords;
 193
 194    assert(output_size_bytes >= 1);
 195    if (output_size_bytes > GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES)
 196       return false;
 197
 198    /* URB entry sizes are stored as a multiple of 64 bytes. */
 199    c.prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
 200
 201    c.prog_data.output_topology = prim_to_hw_prim[gp->program.OutputType];
 202
 203    /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
 204     * need to program a URB read length of ceiling(num_slots / 2).
 205     */
 206    c.prog_data.base.urb_read_length = (c.key.input_vue_map.num_slots + 1) / 2;
 207
 208    void *mem_ctx = ralloc_context(NULL);
 209    unsigned program_size;
 210    const unsigned *program =
 211       brw_gs_emit(brw, prog, &c, mem_ctx, &program_size);
 212    if (program == NULL) {
 213       ralloc_free(mem_ctx);
 214       return false;
 215    }
 216
 217    /* Scratch space is used for register spilling */
 218    if (c.base.last_scratch) {
 219       perf_debug("Geometry shader triggered register spilling.  "
 220                  "Try reducing the number of live vec4 values to "
 221                  "improve performance.\n");
 222
 223       c.prog_data.base.total_scratch
 224          = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
 225
 226       brw_get_scratch_bo(brw, &stage_state->scratch_bo,
 227                          c.prog_data.base.total_scratch * brw->max_gs_threads);
 228    }
 229
 230    brw_upload_cache(&brw->cache, BRW_GS_PROG,
 231                     &c.key, sizeof(c.key),
 232                     program, program_size,
 233                     &c.prog_data, sizeof(c.prog_data),
 234                     &stage_state->prog_offset, &brw->gs.prog_data);
 235    ralloc_free(mem_ctx);
 236
 237    return true;
 238 }
 239
 240
 241 static void
 242 brw_upload_gs_prog(struct brw_context *brw)
 243 {
 244    struct gl_context *ctx = &brw->ctx;
 245    struct brw_stage_state *stage_state = &brw->gs.base;
 246    struct brw_gs_prog_key key;
 247    /* BRW_NEW_GEOMETRY_PROGRAM */
 248    struct brw_geometry_program *gp =
 249       (struct brw_geometry_program *) brw->geometry_program;
 250
 251    if (gp == NULL) {
 252       /* No geometry shader.  Vertex data just passes straight through. */
 253       if (brw->state.dirty.brw & BRW_NEW_VUE_MAP_VS) {
 254          brw->vue_map_geom_out = brw->vue_map_vs;
 255          brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 256       }
 257       return;
 258    }
 259
 260    struct gl_program *prog = &gp->program.Base;
 261
 262    memset(&key, 0, sizeof(key));
 263
 264    key.base.program_string_id = gp->id;
 265    brw_setup_vec4_key_clip_info(brw, &key.base, gp->program.UsesClipDistance);
 266
 267    /* _NEW_LIGHT | _NEW_BUFFERS */
 268    key.base.clamp_vertex_color = ctx->Light._ClampVertexColor;
 269
 270    /* _NEW_TEXTURE */
 271    brw_populate_sampler_prog_key_data(ctx, prog, stage_state->sampler_count,
 272                                       &key.base.tex);
 273
 274    /* BRW_NEW_VUE_MAP_VS */
 275    key.input_vue_map = brw->vue_map_vs;
 276
 277    if (!brw_search_cache(&brw->cache, BRW_GS_PROG,
 278                          &key, sizeof(key),
 279                          &stage_state->prog_offset, &brw->gs.prog_data)) {
 280       bool success = do_gs_prog(brw, ctx->Shader.CurrentGeometryProgram,
 281                                 gp, &key);
 282       assert(success);
 283    }
 284    if (memcmp(&brw->vs.prog_data->base.vue_map, &brw->vue_map_geom_out,
 285               sizeof(brw->vue_map_geom_out)) != 0) {
 286       brw->vue_map_geom_out = brw->gs.prog_data->base.vue_map;
 287       brw->state.dirty.brw |= BRW_NEW_VUE_MAP_GEOM_OUT;
 288    }
 289 }
 290
 291
 292 const struct brw_tracked_state brw_gs_prog = {
 293    .dirty = {
 294       .mesa  = (_NEW_LIGHT | _NEW_BUFFERS | _NEW_TEXTURE),
 295       .brw   = BRW_NEW_GEOMETRY_PROGRAM | BRW_NEW_VUE_MAP_VS,
 296    },
 297    .emit = brw_upload_gs_prog
 298 };
 299
 300
 301 bool
 302 brw_gs_prog_data_compare(const void *in_a, const void *in_b,
 303                          int aux_size, const void *in_key)
 304 {
 305    const struct brw_gs_prog_data *a = in_a;
 306    const struct brw_gs_prog_data *b = in_b;
 307
 308    /* Compare the base vec4 structure. */
 309    if (!brw_vec4_prog_data_compare(&a->base, &b->base))
 310       return false;
 311
 312    /* Compare the rest of the struct. */
 313    const unsigned offset = sizeof(struct brw_vec4_prog_data);
 314    if (memcmp(((char *) &a) + offset, ((char *) &b) + offset,
 315               sizeof(struct brw_gs_prog_data) - offset)) {
 316       return false;
 317    }
 318
 319    return true;
 320 }
 321
 322
 323 void
 324 brw_gs_prog_data_free(const void *in_prog_data)
 325 {
 326    const struct brw_gs_prog_data *prog_data = in_prog_data;
 327
 328    brw_vec4_prog_data_free(&prog_data->base);
 329 }