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

   1 /*
   2  * Copyright (c) 2014 - 2015 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 #include "util/ralloc.h"
  25 #include "brw_context.h"
  26 #include "brw_cs.h"
  27 #include "brw_wm.h"
  28 #include "intel_mipmap_tree.h"
  29 #include "brw_state.h"
  30 #include "intel_batchbuffer.h"
  31 #include "compiler/brw_nir.h"
  32 #include "brw_program.h"
  33 #include "compiler/glsl/ir_uniform.h"
  34
  35 static void
  36 assign_cs_binding_table_offsets(const struct gen_device_info *devinfo,
  37                                 const struct gl_program *prog,
  38                                 struct brw_cs_prog_data *prog_data)
  39 {
  40    uint32_t next_binding_table_offset = 0;
  41
  42    /* May not be used if the gl_NumWorkGroups variable is not accessed. */
  43    prog_data->binding_table.work_groups_start = next_binding_table_offset;
  44    next_binding_table_offset++;
  45
  46    brw_assign_common_binding_table_offsets(devinfo, prog, &prog_data->base,
  47                                            next_binding_table_offset);
  48 }
  49
  50 static bool
  51 brw_codegen_cs_prog(struct brw_context *brw,
  52                     struct brw_program *cp,
  53                     struct brw_cs_prog_key *key)
  54 {
  55    const struct gen_device_info *devinfo = &brw->screen->devinfo;
  56    const GLuint *program;
  57    void *mem_ctx = ralloc_context(NULL);
  58    GLuint program_size;
  59    struct brw_cs_prog_data prog_data;
  60    bool start_busy = false;
  61    double start_time = 0;
  62
  63    memset(&prog_data, 0, sizeof(prog_data));
  64
  65    if (cp->program.info.cs.shared_size > 64 * 1024) {
  66       cp->program.sh.data->LinkStatus = linking_failure;
  67       const char *error_str =
  68          "Compute shader used more than 64KB of shared variables";
  69       ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
  70       _mesa_problem(NULL, "Failed to link compute shader: %s\n", error_str);
  71
  72       ralloc_free(mem_ctx);
  73       return false;
  74    } else {
  75       prog_data.base.total_shared = cp->program.info.cs.shared_size;
  76    }
  77
  78    assign_cs_binding_table_offsets(devinfo, &cp->program, &prog_data);
  79
  80    brw_nir_setup_glsl_uniforms(mem_ctx, cp->program.nir,
  81                                &cp->program, &prog_data.base, true);
  82
  83    if (unlikely(brw->perf_debug)) {
  84       start_busy = (brw->batch.last_bo &&
  85                     brw_bo_busy(brw->batch.last_bo));
  86       start_time = get_time();
  87    }
  88
  89    int st_index = -1;
  90    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
  91       st_index = brw_get_shader_time_index(brw, &cp->program, ST_CS, true);
  92
  93    char *error_str;
  94    program = brw_compile_cs(brw->screen->compiler, brw, mem_ctx, key,
  95                             &prog_data, cp->program.nir, st_index,
  96                             &program_size, &error_str);
  97    if (program == NULL) {
  98       cp->program.sh.data->LinkStatus = linking_failure;
  99       ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
 100       _mesa_problem(NULL, "Failed to compile compute shader: %s\n", error_str);
 101
 102       ralloc_free(mem_ctx);
 103       return false;
 104    }
 105
 106    if (unlikely(brw->perf_debug)) {
 107       if (cp->compiled_once) {
 108          _mesa_problem(&brw->ctx, "CS programs shouldn't need recompiles");
 109       }
 110       cp->compiled_once = true;
 111
 112       if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
 113          perf_debug("CS compile took %.03f ms and stalled the GPU\n",
 114                     (get_time() - start_time) * 1000);
 115       }
 116    }
 117
 118    const unsigned subslices = MAX2(brw->screen->subslice_total, 1);
 119
 120    /* WaCSScratchSize:hsw
 121     *
 122     * Haswell's scratch space address calculation appears to be sparse
 123     * rather than tightly packed.  The Thread ID has bits indicating
 124     * which subslice, EU within a subslice, and thread within an EU
 125     * it is.  There's a maximum of two slices and two subslices, so these
 126     * can be stored with a single bit.  Even though there are only 10 EUs
 127     * per subslice, this is stored in 4 bits, so there's an effective
 128     * maximum value of 16 EUs.  Similarly, although there are only 7
 129     * threads per EU, this is stored in a 3 bit number, giving an effective
 130     * maximum value of 8 threads per EU.
 131     *
 132     * This means that we need to use 16 * 8 instead of 10 * 7 for the
 133     * number of threads per subslice.
 134     */
 135    const unsigned scratch_ids_per_subslice =
 136       devinfo->is_haswell ? 16 * 8 : devinfo->max_cs_threads;
 137
 138    brw_alloc_stage_scratch(brw, &brw->cs.base,
 139                            prog_data.base.total_scratch,
 140                            scratch_ids_per_subslice * subslices);
 141
 142    /* The param and pull_param arrays will be freed by the shader cache. */
 143    ralloc_steal(NULL, prog_data.base.param);
 144    ralloc_steal(NULL, prog_data.base.pull_param);
 145    brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
 146                     key, sizeof(*key),
 147                     program, program_size,
 148                     &prog_data, sizeof(prog_data),
 149                     &brw->cs.base.prog_offset, &brw->cs.base.prog_data);
 150    ralloc_free(mem_ctx);
 151
 152    return true;
 153 }
 154
 155
 156 static void
 157 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
 158 {
 159    struct gl_context *ctx = &brw->ctx;
 160    /* BRW_NEW_COMPUTE_PROGRAM */
 161    const struct brw_program *cp =
 162       (struct brw_program *) brw->programs[MESA_SHADER_COMPUTE];
 163    const struct gl_program *prog = (struct gl_program *) cp;
 164
 165    memset(key, 0, sizeof(*key));
 166
 167    /* _NEW_TEXTURE */
 168    brw_populate_sampler_prog_key_data(ctx, prog, &key->tex);
 169
 170    /* The unique compute program ID */
 171    key->program_string_id = cp->id;
 172 }
 173
 174
 175 void
 176 brw_upload_cs_prog(struct brw_context *brw)
 177 {
 178    struct gl_context *ctx = &brw->ctx;
 179    struct brw_cs_prog_key key;
 180    struct brw_program *cp =
 181       (struct brw_program *) brw->programs[MESA_SHADER_COMPUTE];
 182
 183    if (!cp)
 184       return;
 185
 186    if (!brw_state_dirty(brw, _NEW_TEXTURE, BRW_NEW_COMPUTE_PROGRAM))
 187       return;
 188
 189    brw->cs.base.sampler_count =
 190       util_last_bit(ctx->ComputeProgram._Current->SamplersUsed);
 191
 192    brw_cs_populate_key(brw, &key);
 193
 194    if (!brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG,
 195                          &key, sizeof(key),
 196                          &brw->cs.base.prog_offset,
 197                          &brw->cs.base.prog_data)) {
 198       bool success = brw_codegen_cs_prog(brw, cp, &key);
 199       (void) success;
 200       assert(success);
 201    }
 202 }
 203
 204
 205 bool
 206 brw_cs_precompile(struct gl_context *ctx, struct gl_program *prog)
 207 {
 208    struct brw_context *brw = brw_context(ctx);
 209    struct brw_cs_prog_key key;
 210
 211    struct brw_program *bcp = brw_program(prog);
 212
 213    memset(&key, 0, sizeof(key));
 214    key.program_string_id = bcp->id;
 215
 216    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 217
 218    uint32_t old_prog_offset = brw->cs.base.prog_offset;
 219    struct brw_stage_prog_data *old_prog_data = brw->cs.base.prog_data;
 220
 221    bool success = brw_codegen_cs_prog(brw, bcp, &key);
 222
 223    brw->cs.base.prog_offset = old_prog_offset;
 224    brw->cs.base.prog_data = old_prog_data;
 225
 226    return success;
 227 }