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

   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
  25 #include "util/ralloc.h"
  26 #include "brw_context.h"
  27 #include "brw_cs.h"
  28 #include "brw_fs.h"
  29 #include "brw_eu.h"
  30 #include "brw_wm.h"
  31 #include "intel_mipmap_tree.h"
  32 #include "brw_state.h"
  33 #include "intel_batchbuffer.h"
  34
  35 extern "C"
  36 bool
  37 brw_cs_prog_data_compare(const void *in_a, const void *in_b)
  38 {
  39    const struct brw_cs_prog_data *a =
  40       (const struct brw_cs_prog_data *)in_a;
  41    const struct brw_cs_prog_data *b =
  42       (const struct brw_cs_prog_data *)in_b;
  43
  44    /* Compare the base structure. */
  45    if (!brw_stage_prog_data_compare(&a->base, &b->base))
  46       return false;
  47
  48    /* Compare the rest of the structure. */
  49    const unsigned offset = sizeof(struct brw_stage_prog_data);
  50    if (memcmp(((char *) a) + offset, ((char *) b) + offset,
  51               sizeof(struct brw_cs_prog_data) - offset))
  52       return false;
  53
  54    return true;
  55 }
  56
  57
  58 static const unsigned *
  59 brw_cs_emit(struct brw_context *brw,
  60             void *mem_ctx,
  61             const struct brw_cs_prog_key *key,
  62             struct brw_cs_prog_data *prog_data,
  63             struct gl_compute_program *cp,
  64             struct gl_shader_program *prog,
  65             unsigned *final_assembly_size)
  66 {
  67    bool start_busy = false;
  68    double start_time = 0;
  69
  70    if (unlikely(brw->perf_debug)) {
  71       start_busy = (brw->batch.last_bo &&
  72                     drm_intel_bo_busy(brw->batch.last_bo));
  73       start_time = get_time();
  74    }
  75
  76    struct brw_shader *shader =
  77       (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_COMPUTE];
  78
  79    if (unlikely(INTEL_DEBUG & DEBUG_CS))
  80       brw_dump_ir("compute", prog, &shader->base, &cp->Base);
  81
  82    prog_data->local_size[0] = cp->LocalSize[0];
  83    prog_data->local_size[1] = cp->LocalSize[1];
  84    prog_data->local_size[2] = cp->LocalSize[2];
  85    int local_workgroup_size =
  86       cp->LocalSize[0] * cp->LocalSize[1] * cp->LocalSize[2];
  87
  88    cfg_t *cfg = NULL;
  89    const char *fail_msg = NULL;
  90
  91    /* Now the main event: Visit the shader IR and generate our CS IR for it.
  92     */
  93    fs_visitor v8(brw, mem_ctx, MESA_SHADER_COMPUTE, key, &prog_data->base, prog,
  94                  &cp->Base, 8);
  95    if (!v8.run_cs()) {
  96       fail_msg = v8.fail_msg;
  97    } else if (local_workgroup_size <= 8 * brw->max_cs_threads) {
  98       cfg = v8.cfg;
  99       prog_data->simd_size = 8;
 100    }
 101
 102    fs_visitor v16(brw, mem_ctx, MESA_SHADER_COMPUTE, key, &prog_data->base, prog,
 103                   &cp->Base, 16);
 104    if (likely(!(INTEL_DEBUG & DEBUG_NO16)) &&
 105        !fail_msg && !v8.simd16_unsupported &&
 106        local_workgroup_size <= 16 * brw->max_cs_threads) {
 107       /* Try a SIMD16 compile */
 108       v16.import_uniforms(&v8);
 109       if (!v16.run_cs()) {
 110          perf_debug("SIMD16 shader failed to compile: %s", v16.fail_msg);
 111          if (!cfg) {
 112             fail_msg =
 113                "Couldn't generate SIMD16 program and not "
 114                "enough threads for SIMD8";
 115          }
 116       } else {
 117          cfg = v16.cfg;
 118          prog_data->simd_size = 16;
 119       }
 120    }
 121
 122    if (unlikely(cfg == NULL)) {
 123       assert(fail_msg);
 124       prog->LinkStatus = false;
 125       ralloc_strcat(&prog->InfoLog, fail_msg);
 126       _mesa_problem(NULL, "Failed to compile compute shader: %s\n",
 127                     fail_msg);
 128       return NULL;
 129    }
 130
 131    fs_generator g(brw->intelScreen->compiler,
 132                   mem_ctx, (void*) key, &prog_data->base, &cp->Base,
 133                   v8.promoted_constants, v8.runtime_check_aads_emit, "CS");
 134    if (INTEL_DEBUG & DEBUG_CS) {
 135       char *name = ralloc_asprintf(mem_ctx, "%s compute shader %d",
 136                                    prog->Label ? prog->Label : "unnamed",
 137                                    prog->Name);
 138       g.enable_debug(name);
 139    }
 140
 141    g.generate_code(cfg, prog_data->simd_size);
 142
 143    if (unlikely(brw->perf_debug) && shader) {
 144       if (shader->compiled_once) {
 145          _mesa_problem(&brw->ctx, "CS programs shouldn't need recompiles");
 146       }
 147       shader->compiled_once = true;
 148
 149       if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
 150          perf_debug("CS compile took %.03f ms and stalled the GPU\n",
 151                     (get_time() - start_time) * 1000);
 152       }
 153    }
 154
 155    return g.get_assembly(final_assembly_size);
 156 }
 157
 158 static bool
 159 brw_codegen_cs_prog(struct brw_context *brw,
 160                     struct gl_shader_program *prog,
 161                     struct brw_compute_program *cp,
 162                     struct brw_cs_prog_key *key)
 163 {
 164    struct gl_context *ctx = &brw->ctx;
 165    const GLuint *program;
 166    void *mem_ctx = ralloc_context(NULL);
 167    GLuint program_size;
 168    struct brw_cs_prog_data prog_data;
 169
 170    struct gl_shader *cs = prog->_LinkedShaders[MESA_SHADER_COMPUTE];
 171    assert (cs);
 172
 173    memset(&prog_data, 0, sizeof(prog_data));
 174
 175    /* Allocate the references to the uniforms that will end up in the
 176     * prog_data associated with the compiled program, and which will be freed
 177     * by the state cache.
 178     */
 179    int param_count = cs->num_uniform_components;
 180
 181    /* The backend also sometimes adds params for texture size. */
 182    param_count += 2 * ctx->Const.Program[MESA_SHADER_COMPUTE].MaxTextureImageUnits;
 183    prog_data.base.param =
 184       rzalloc_array(NULL, const gl_constant_value *, param_count);
 185    prog_data.base.pull_param =
 186       rzalloc_array(NULL, const gl_constant_value *, param_count);
 187    prog_data.base.nr_params = param_count;
 188
 189    program = brw_cs_emit(brw, mem_ctx, key, &prog_data,
 190                          &cp->program, prog, &program_size);
 191    if (program == NULL) {
 192       ralloc_free(mem_ctx);
 193       return false;
 194    }
 195
 196    if (prog_data.base.total_scratch) {
 197       brw_get_scratch_bo(brw, &brw->cs.base.scratch_bo,
 198                          prog_data.base.total_scratch * brw->max_cs_threads);
 199    }
 200
 201    if (unlikely(INTEL_DEBUG & DEBUG_CS))
 202       fprintf(stderr, "\n");
 203
 204    brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
 205                     key, sizeof(*key),
 206                     program, program_size,
 207                     &prog_data, sizeof(prog_data),
 208                     &brw->cs.base.prog_offset, &brw->cs.prog_data);
 209    ralloc_free(mem_ctx);
 210
 211    return true;
 212 }
 213
 214
 215 static void
 216 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
 217 {
 218    /* BRW_NEW_COMPUTE_PROGRAM */
 219    const struct brw_compute_program *cp =
 220       (struct brw_compute_program *) brw->compute_program;
 221
 222    memset(key, 0, sizeof(*key));
 223
 224    /* The unique compute program ID */
 225    key->program_string_id = cp->id;
 226 }
 227
 228
 229 extern "C"
 230 void
 231 brw_upload_cs_prog(struct brw_context *brw)
 232 {
 233    struct gl_context *ctx = &brw->ctx;
 234    struct brw_cs_prog_key key;
 235    struct brw_compute_program *cp = (struct brw_compute_program *)
 236       brw->compute_program;
 237
 238    if (!cp)
 239       return;
 240
 241    if (!brw_state_dirty(brw, 0, BRW_NEW_COMPUTE_PROGRAM))
 242       return;
 243
 244    brw_cs_populate_key(brw, &key);
 245
 246    if (!brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG,
 247                          &key, sizeof(key),
 248                          &brw->cs.base.prog_offset, &brw->cs.prog_data)) {
 249       bool success =
 250          brw_codegen_cs_prog(brw,
 251                              ctx->Shader.CurrentProgram[MESA_SHADER_COMPUTE],
 252                              cp, &key);
 253       (void) success;
 254       assert(success);
 255    }
 256    brw->cs.base.prog_data = &brw->cs.prog_data->base;
 257 }
 258
 259
 260 extern "C" bool
 261 brw_cs_precompile(struct gl_context *ctx,
 262                   struct gl_shader_program *shader_prog,
 263                   struct gl_program *prog)
 264 {
 265    struct brw_context *brw = brw_context(ctx);
 266    struct brw_cs_prog_key key;
 267
 268    struct gl_compute_program *cp = (struct gl_compute_program *) prog;
 269    struct brw_compute_program *bcp = brw_compute_program(cp);
 270
 271    memset(&key, 0, sizeof(key));
 272    key.program_string_id = bcp->id;
 273
 274    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 275
 276    uint32_t old_prog_offset = brw->cs.base.prog_offset;
 277    struct brw_cs_prog_data *old_prog_data = brw->cs.prog_data;
 278
 279    bool success = brw_codegen_cs_prog(brw, shader_prog, bcp, &key);
 280
 281    brw->cs.base.prog_offset = old_prog_offset;
 282    brw->cs.prog_data = old_prog_data;
 283
 284    return success;
 285 }
 286
 287
 288 static void
 289 brw_upload_cs_state(struct brw_context *brw)
 290 {
 291    if (!brw->cs.prog_data)
 292       return;
 293
 294    uint32_t offset;
 295    uint32_t *desc = (uint32_t*) brw_state_batch(brw, AUB_TRACE_SURFACE_STATE,
 296                                                 8 * 4, 64, &offset);
 297    struct brw_stage_state *stage_state = &brw->cs.base;
 298    struct brw_cs_prog_data *cs_prog_data = brw->cs.prog_data;
 299    struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
 300
 301    if (INTEL_DEBUG & DEBUG_SHADER_TIME) {
 302       brw->vtbl.emit_buffer_surface_state(
 303          brw, &stage_state->surf_offset[
 304                  prog_data->binding_table.shader_time_start],
 305          brw->shader_time.bo, 0, BRW_SURFACEFORMAT_RAW,
 306          brw->shader_time.bo->size, 1, true);
 307    }
 308
 309    uint32_t *bind = (uint32_t*) brw_state_batch(brw, AUB_TRACE_BINDING_TABLE,
 310                                             prog_data->binding_table.size_bytes,
 311                                             32, &stage_state->bind_bo_offset);
 312
 313    uint32_t dwords = brw->gen < 8 ? 8 : 9;
 314    BEGIN_BATCH(dwords);
 315    OUT_BATCH(MEDIA_VFE_STATE << 16 | (dwords - 2));
 316
 317    if (prog_data->total_scratch) {
 318       if (brw->gen >= 8)
 319          OUT_RELOC64(stage_state->scratch_bo,
 320                      I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 321                      ffs(prog_data->total_scratch) - 11);
 322       else
 323          OUT_RELOC(stage_state->scratch_bo,
 324                    I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
 325                    ffs(prog_data->total_scratch) - 11);
 326    } else {
 327       OUT_BATCH(0);
 328       if (brw->gen >= 8)
 329          OUT_BATCH(0);
 330    }
 331
 332    const uint32_t vfe_num_urb_entries = brw->gen >= 8 ? 2 : 0;
 333    const uint32_t vfe_gpgpu_mode =
 334       brw->gen == 7 ? SET_FIELD(1, GEN7_MEDIA_VFE_STATE_GPGPU_MODE) : 0;
 335    OUT_BATCH(SET_FIELD(brw->max_cs_threads - 1, MEDIA_VFE_STATE_MAX_THREADS) |
 336              SET_FIELD(vfe_num_urb_entries, MEDIA_VFE_STATE_URB_ENTRIES) |
 337              SET_FIELD(1, MEDIA_VFE_STATE_RESET_GTW_TIMER) |
 338              SET_FIELD(1, MEDIA_VFE_STATE_BYPASS_GTW) |
 339              vfe_gpgpu_mode);
 340
 341    OUT_BATCH(0);
 342    const uint32_t vfe_urb_allocation = brw->gen >= 8 ? 2 : 0;
 343    OUT_BATCH(SET_FIELD(vfe_urb_allocation, MEDIA_VFE_STATE_URB_ALLOC));
 344    OUT_BATCH(0);
 345    OUT_BATCH(0);
 346    OUT_BATCH(0);
 347    ADVANCE_BATCH();
 348
 349    /* BRW_NEW_SURFACES and BRW_NEW_*_CONSTBUF */
 350    memcpy(bind, stage_state->surf_offset,
 351           prog_data->binding_table.size_bytes);
 352
 353    memset(desc, 0, 8 * 4);
 354
 355    int dw = 0;
 356    desc[dw++] = brw->cs.base.prog_offset;
 357    if (brw->gen >= 8)
 358       desc[dw++] = 0; /* Kernel Start Pointer High */
 359    desc[dw++] = 0;
 360    desc[dw++] = 0;
 361    desc[dw++] = stage_state->bind_bo_offset;
 362
 363    BEGIN_BATCH(4);
 364    OUT_BATCH(MEDIA_INTERFACE_DESCRIPTOR_LOAD << 16 | (4 - 2));
 365    OUT_BATCH(0);
 366    OUT_BATCH(8 * 4);
 367    OUT_BATCH(offset);
 368    ADVANCE_BATCH();
 369 }
 370
 371
 372 extern "C"
 373 const struct brw_tracked_state brw_cs_state = {
 374    /* explicit initialisers aren't valid C++, comment
 375     * them for documentation purposes */
 376    /* .dirty = */{
 377       /* .mesa = */ 0,
 378       /* .brw = */  BRW_NEW_CS_PROG_DATA,
 379    },
 380    /* .emit = */ brw_upload_cs_state
 381 };