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 brw_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_compute_program *cp,
  58                     struct brw_cs_prog_key *key)
  59 {
  60    struct gl_context *ctx = &brw->ctx;
  61    const GLuint *program;
  62    void *mem_ctx = ralloc_context(NULL);
  63    GLuint program_size;
  64    struct brw_cs_prog_data prog_data;
  65    bool start_busy = false;
  66    double start_time = 0;
  67
  68    struct brw_shader *cs =
  69       (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_COMPUTE];
  70    assert (cs);
  71
  72    memset(&prog_data, 0, sizeof(prog_data));
  73
  74    if (prog->Comp.SharedSize > 64 * 1024) {
  75       prog->LinkStatus = false;
  76       const char *error_str =
  77          "Compute shader used more than 64KB of shared variables";
  78       ralloc_strcat(&prog->InfoLog, error_str);
  79       _mesa_problem(NULL, "Failed to link compute shader: %s\n", error_str);
  80
  81       ralloc_free(mem_ctx);
  82       return false;
  83    } else {
  84       prog_data.base.total_shared = prog->Comp.SharedSize;
  85    }
  86
  87    assign_cs_binding_table_offsets(brw->intelScreen->devinfo, prog,
  88                                    &cp->program.Base, &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.Base.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, cs->base.NumImages);
 107    prog_data.base.nr_params = param_count;
 108    prog_data.base.nr_image_params = cs->base.NumImages;
 109
 110    brw_nir_setup_glsl_uniforms(cp->program.Base.nir, prog, &cp->program.Base,
 111                                &prog_data.base, true);
 112
 113    if (unlikely(brw->perf_debug)) {
 114       start_busy = (brw->batch.last_bo &&
 115                     drm_intel_bo_busy(brw->batch.last_bo));
 116       start_time = get_time();
 117    }
 118
 119    if (unlikely(INTEL_DEBUG & DEBUG_CS))
 120       brw_dump_ir("compute", prog, &cs->base, &cp->program.Base);
 121
 122    int st_index = -1;
 123    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
 124       st_index = brw_get_shader_time_index(brw, prog, &cp->program.Base, ST_CS);
 125
 126    char *error_str;
 127    program = brw_compile_cs(brw->intelScreen->compiler, brw, mem_ctx,
 128                             key, &prog_data, cp->program.Base.nir,
 129                             st_index, &program_size, &error_str);
 130    if (program == NULL) {
 131       prog->LinkStatus = false;
 132       ralloc_strcat(&prog->InfoLog, error_str);
 133       _mesa_problem(NULL, "Failed to compile compute shader: %s\n", error_str);
 134
 135       ralloc_free(mem_ctx);
 136       return false;
 137    }
 138
 139    if (unlikely(brw->perf_debug) && cs) {
 140       if (cs->compiled_once) {
 141          _mesa_problem(&brw->ctx, "CS programs shouldn't need recompiles");
 142       }
 143       cs->compiled_once = true;
 144
 145       if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
 146          perf_debug("CS compile took %.03f ms and stalled the GPU\n",
 147                     (get_time() - start_time) * 1000);
 148       }
 149    }
 150
 151    if (prog_data.base.total_scratch) {
 152       const unsigned subslices = MAX2(brw->intelScreen->subslice_total, 1);
 153
 154       /* WaCSScratchSize:hsw
 155        *
 156        * Haswell's scratch space address calculation appears to be sparse
 157        * rather than tightly packed.  The Thread ID has bits indicating
 158        * which subslice, EU within a subslice, and thread within an EU
 159        * it is.  There's a maximum of two slices and two subslices, so these
 160        * can be stored with a single bit.  Even though there are only 10 EUs
 161        * per subslice, this is stored in 4 bits, so there's an effective
 162        * maximum value of 16 EUs.  Similarly, although there are only 7
 163        * threads per EU, this is stored in a 3 bit number, giving an effective
 164        * maximum value of 8 threads per EU.
 165        *
 166        * This means that we need to use 16 * 8 instead of 10 * 7 for the
 167        * number of threads per subslice.
 168        */
 169       const unsigned scratch_ids_per_subslice =
 170          brw->is_haswell ? 16 * 8 : brw->max_cs_threads;
 171
 172       brw_get_scratch_bo(brw, &brw->cs.base.scratch_bo,
 173                          prog_data.base.total_scratch *
 174                          scratch_ids_per_subslice * subslices);
 175    }
 176
 177    if (unlikely(INTEL_DEBUG & DEBUG_CS))
 178       fprintf(stderr, "\n");
 179
 180    brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
 181                     key, sizeof(*key),
 182                     program, program_size,
 183                     &prog_data, sizeof(prog_data),
 184                     &brw->cs.base.prog_offset, &brw->cs.prog_data);
 185    ralloc_free(mem_ctx);
 186
 187    return true;
 188 }
 189
 190
 191 static void
 192 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
 193 {
 194    struct gl_context *ctx = &brw->ctx;
 195    /* BRW_NEW_COMPUTE_PROGRAM */
 196    const struct brw_compute_program *cp =
 197       (struct brw_compute_program *) brw->compute_program;
 198    const struct gl_program *prog = (struct gl_program *) cp;
 199
 200    memset(key, 0, sizeof(*key));
 201
 202    /* _NEW_TEXTURE */
 203    brw_populate_sampler_prog_key_data(ctx, prog, brw->cs.base.sampler_count,
 204                                       &key->tex);
 205
 206    /* The unique compute program ID */
 207    key->program_string_id = cp->id;
 208 }
 209
 210
 211 void
 212 brw_upload_cs_prog(struct brw_context *brw)
 213 {
 214    struct gl_context *ctx = &brw->ctx;
 215    struct brw_cs_prog_key key;
 216    struct brw_compute_program *cp = (struct brw_compute_program *)
 217       brw->compute_program;
 218
 219    if (!cp)
 220       return;
 221
 222    if (!brw_state_dirty(brw, _NEW_TEXTURE, BRW_NEW_COMPUTE_PROGRAM))
 223       return;
 224
 225    brw->cs.base.sampler_count =
 226       _mesa_fls(ctx->ComputeProgram._Current->Base.SamplersUsed);
 227
 228    brw_cs_populate_key(brw, &key);
 229
 230    if (!brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG,
 231                          &key, sizeof(key),
 232                          &brw->cs.base.prog_offset, &brw->cs.prog_data)) {
 233       bool success =
 234          brw_codegen_cs_prog(brw,
 235                              ctx->Shader.CurrentProgram[MESA_SHADER_COMPUTE],
 236                              cp, &key);
 237       (void) success;
 238       assert(success);
 239    }
 240    brw->cs.base.prog_data = &brw->cs.prog_data->base;
 241 }
 242
 243
 244 bool
 245 brw_cs_precompile(struct gl_context *ctx,
 246                   struct gl_shader_program *shader_prog,
 247                   struct gl_program *prog)
 248 {
 249    struct brw_context *brw = brw_context(ctx);
 250    struct brw_cs_prog_key key;
 251
 252    struct gl_compute_program *cp = (struct gl_compute_program *) prog;
 253    struct brw_compute_program *bcp = brw_compute_program(cp);
 254
 255    memset(&key, 0, sizeof(key));
 256    key.program_string_id = bcp->id;
 257
 258    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 259
 260    uint32_t old_prog_offset = brw->cs.base.prog_offset;
 261    struct brw_cs_prog_data *old_prog_data = brw->cs.prog_data;
 262
 263    bool success = brw_codegen_cs_prog(brw, shader_prog, bcp, &key);
 264
 265    brw->cs.base.prog_offset = old_prog_offset;
 266    brw->cs.prog_data = old_prog_data;
 267
 268    return success;
 269 }