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_eu.h"
  28 #include "brw_wm.h"
  29 #include "brw_shader.h"
  30 #include "intel_mipmap_tree.h"
  31 #include "brw_state.h"
  32 #include "intel_batchbuffer.h"
  33 #include "brw_nir.h"
  34 #include "brw_program.h"
  35 #include "compiler/glsl/ir_uniform.h"
  36
  37 static void
  38 assign_cs_binding_table_offsets(const struct gen_device_info *devinfo,
  39                                 const struct gl_shader_program *shader_prog,
  40                                 const struct gl_program *prog,
  41                                 struct brw_cs_prog_data *prog_data)
  42 {
  43    uint32_t next_binding_table_offset = 0;
  44
  45    /* May not be used if the gl_NumWorkGroups variable is not accessed. */
  46    prog_data->binding_table.work_groups_start = next_binding_table_offset;
  47    next_binding_table_offset++;
  48
  49    brw_assign_common_binding_table_offsets(MESA_SHADER_COMPUTE, devinfo,
  50                                            shader_prog, prog, &prog_data->base,
  51                                            next_binding_table_offset);
  52 }
  53
  54 static bool
  55 brw_codegen_cs_prog(struct brw_context *brw,
  56                     struct gl_shader_program *prog,
  57                     struct brw_program *cp,
  58                     struct brw_cs_prog_key *key)
  59 {
  60    const struct gen_device_info *devinfo = &brw->screen->devinfo;
  61    struct gl_context *ctx = &brw->ctx;
  62    const GLuint *program;
  63    void *mem_ctx = ralloc_context(NULL);
  64    GLuint program_size;
  65    struct brw_cs_prog_data prog_data;
  66    bool start_busy = false;
  67    double start_time = 0;
  68
  69    struct brw_shader *cs =
  70       (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_COMPUTE];
  71    assert (cs);
  72
  73    memset(&prog_data, 0, sizeof(prog_data));
  74
  75    if (prog->Comp.SharedSize > 64 * 1024) {
  76       prog->LinkStatus = false;
  77       const char *error_str =
  78          "Compute shader used more than 64KB of shared variables";
  79       ralloc_strcat(&prog->InfoLog, error_str);
  80       _mesa_problem(NULL, "Failed to link compute shader: %s\n", error_str);
  81
  82       ralloc_free(mem_ctx);
  83       return false;
  84    } else {
  85       prog_data.base.total_shared = prog->Comp.SharedSize;
  86    }
  87
  88    assign_cs_binding_table_offsets(devinfo, prog, &cp->program, &prog_data);
  89
  90    /* Allocate the references to the uniforms that will end up in the
  91     * prog_data associated with the compiled program, and which will be freed
  92     * by the state cache.
  93     */
  94    int param_count = cp->program.nir->num_uniforms / 4;
  95
  96    /* The backend also sometimes add a param for the thread local id. */
  97    prog_data.thread_local_id_index = param_count++;
  98
  99    /* The backend also sometimes adds params for texture size. */
 100    param_count += 2 * ctx->Const.Program[MESA_SHADER_COMPUTE].MaxTextureImageUnits;
 101    prog_data.base.param =
 102       rzalloc_array(NULL, const gl_constant_value *, param_count);
 103    prog_data.base.pull_param =
 104       rzalloc_array(NULL, const gl_constant_value *, param_count);
 105    prog_data.base.image_param =
 106       rzalloc_array(NULL, struct brw_image_param,
 107                     cp->program.info.num_images);
 108    prog_data.base.nr_params = param_count;
 109    prog_data.base.nr_image_params = cp->program.info.num_images;
 110
 111    brw_nir_setup_glsl_uniforms(cp->program.nir, prog, &cp->program,
 112                                &prog_data.base, true);
 113
 114    if (unlikely(brw->perf_debug)) {
 115       start_busy = (brw->batch.last_bo &&
 116                     drm_intel_bo_busy(brw->batch.last_bo));
 117       start_time = get_time();
 118    }
 119
 120    int st_index = -1;
 121    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
 122       st_index = brw_get_shader_time_index(brw, prog, &cp->program, ST_CS);
 123
 124    char *error_str;
 125    program = brw_compile_cs(brw->screen->compiler, brw, mem_ctx, key,
 126                             &prog_data, cp->program.nir, st_index,
 127                             &program_size, &error_str);
 128    if (program == NULL) {
 129       prog->LinkStatus = false;
 130       ralloc_strcat(&prog->InfoLog, error_str);
 131       _mesa_problem(NULL, "Failed to compile compute shader: %s\n", error_str);
 132
 133       ralloc_free(mem_ctx);
 134       return false;
 135    }
 136
 137    if (unlikely(brw->perf_debug) && cs) {
 138       if (cs->compiled_once) {
 139          _mesa_problem(&brw->ctx, "CS programs shouldn't need recompiles");
 140       }
 141       cs->compiled_once = true;
 142
 143       if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
 144          perf_debug("CS compile took %.03f ms and stalled the GPU\n",
 145                     (get_time() - start_time) * 1000);
 146       }
 147    }
 148
 149    const unsigned subslices = MAX2(brw->screen->subslice_total, 1);
 150
 151    /* WaCSScratchSize:hsw
 152     *
 153     * Haswell's scratch space address calculation appears to be sparse
 154     * rather than tightly packed.  The Thread ID has bits indicating
 155     * which subslice, EU within a subslice, and thread within an EU
 156     * it is.  There's a maximum of two slices and two subslices, so these
 157     * can be stored with a single bit.  Even though there are only 10 EUs
 158     * per subslice, this is stored in 4 bits, so there's an effective
 159     * maximum value of 16 EUs.  Similarly, although there are only 7
 160     * threads per EU, this is stored in a 3 bit number, giving an effective
 161     * maximum value of 8 threads per EU.
 162     *
 163     * This means that we need to use 16 * 8 instead of 10 * 7 for the
 164     * number of threads per subslice.
 165     */
 166    const unsigned scratch_ids_per_subslice =
 167       brw->is_haswell ? 16 * 8 : devinfo->max_cs_threads;
 168
 169    brw_alloc_stage_scratch(brw, &brw->cs.base,
 170                            prog_data.base.total_scratch,
 171                            scratch_ids_per_subslice * subslices);
 172
 173    brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
 174                     key, sizeof(*key),
 175                     program, program_size,
 176                     &prog_data, sizeof(prog_data),
 177                     &brw->cs.base.prog_offset, &brw->cs.base.prog_data);
 178    ralloc_free(mem_ctx);
 179
 180    return true;
 181 }
 182
 183
 184 static void
 185 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
 186 {
 187    struct gl_context *ctx = &brw->ctx;
 188    /* BRW_NEW_COMPUTE_PROGRAM */
 189    const struct brw_program *cp = (struct brw_program *) brw->compute_program;
 190    const struct gl_program *prog = (struct gl_program *) cp;
 191
 192    memset(key, 0, sizeof(*key));
 193
 194    /* _NEW_TEXTURE */
 195    brw_populate_sampler_prog_key_data(ctx, prog, &key->tex);
 196
 197    /* The unique compute program ID */
 198    key->program_string_id = cp->id;
 199 }
 200
 201
 202 void
 203 brw_upload_cs_prog(struct brw_context *brw)
 204 {
 205    struct gl_context *ctx = &brw->ctx;
 206    struct brw_cs_prog_key key;
 207    struct brw_program *cp = (struct brw_program *) brw->compute_program;
 208
 209    if (!cp)
 210       return;
 211
 212    if (!brw_state_dirty(brw, _NEW_TEXTURE, BRW_NEW_COMPUTE_PROGRAM))
 213       return;
 214
 215    brw->cs.base.sampler_count =
 216       util_last_bit(ctx->ComputeProgram._Current->SamplersUsed);
 217
 218    brw_cs_populate_key(brw, &key);
 219
 220    if (!brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG,
 221                          &key, sizeof(key),
 222                          &brw->cs.base.prog_offset,
 223                          &brw->cs.base.prog_data)) {
 224       bool success =
 225          brw_codegen_cs_prog(brw,
 226                              ctx->Shader.CurrentProgram[MESA_SHADER_COMPUTE],
 227                              cp, &key);
 228       (void) success;
 229       assert(success);
 230    }
 231 }
 232
 233
 234 bool
 235 brw_cs_precompile(struct gl_context *ctx,
 236                   struct gl_shader_program *shader_prog,
 237                   struct gl_program *prog)
 238 {
 239    struct brw_context *brw = brw_context(ctx);
 240    struct brw_cs_prog_key key;
 241
 242    struct brw_program *bcp = brw_program(prog);
 243
 244    memset(&key, 0, sizeof(key));
 245    key.program_string_id = bcp->id;
 246
 247    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 248
 249    uint32_t old_prog_offset = brw->cs.base.prog_offset;
 250    struct brw_stage_prog_data *old_prog_data = brw->cs.base.prog_data;
 251
 252    bool success = brw_codegen_cs_prog(brw, shader_prog, bcp, &key);
 253
 254    brw->cs.base.prog_offset = old_prog_offset;
 255    brw->cs.base.prog_data = old_prog_data;
 256
 257    return success;
 258 }