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 struct brw_cs_parameters
  36 brw_cs_get_parameters(const struct brw_context *brw)
  37 {
  38    assert(brw->cs.base.prog_data);
  39    struct brw_cs_prog_data *cs_prog_data =
  40       brw_cs_prog_data(brw->cs.base.prog_data);
  41
  42    struct brw_cs_parameters params = {};
  43
  44    if (brw->compute.group_size) {
  45       /* With ARB_compute_variable_group_size the group size is set at
  46        * dispatch time, so we can't use the one provided by the compiler.
  47        */
  48       params.group_size = brw->compute.group_size[0] *
  49                           brw->compute.group_size[1] *
  50                           brw->compute.group_size[2];
  51    } else {
  52       params.group_size = cs_prog_data->local_size[0] *
  53                           cs_prog_data->local_size[1] *
  54                           cs_prog_data->local_size[2];
  55    }
  56
  57    params.simd_size =
  58       brw_cs_simd_size_for_group_size(&brw->screen->devinfo,
  59                                       cs_prog_data, params.group_size);
  60    params.threads = DIV_ROUND_UP(params.group_size, params.simd_size);
  61
  62    return params;
  63 }
  64
  65 static void
  66 assign_cs_binding_table_offsets(const struct gen_device_info *devinfo,
  67                                 const struct gl_program *prog,
  68                                 struct brw_cs_prog_data *prog_data)
  69 {
  70    uint32_t next_binding_table_offset = 0;
  71
  72    /* May not be used if the gl_NumWorkGroups variable is not accessed. */
  73    prog_data->binding_table.work_groups_start = next_binding_table_offset;
  74    next_binding_table_offset++;
  75
  76    brw_assign_common_binding_table_offsets(devinfo, prog, &prog_data->base,
  77                                            next_binding_table_offset);
  78 }
  79
  80 static bool
  81 brw_codegen_cs_prog(struct brw_context *brw,
  82                     struct brw_program *cp,
  83                     struct brw_cs_prog_key *key)
  84 {
  85    const struct gen_device_info *devinfo = &brw->screen->devinfo;
  86    const GLuint *program;
  87    void *mem_ctx = ralloc_context(NULL);
  88    struct brw_cs_prog_data prog_data;
  89    bool start_busy = false;
  90    double start_time = 0;
  91    struct gl_context *gl_ctx = &brw->ctx;
  92    nir_shader *nir = nir_shader_clone(mem_ctx, cp->program.nir);
  93
  94    memset(&prog_data, 0, sizeof(prog_data));
  95
  96    if (cp->program.info.cs.shared_size > 64 * 1024) {
  97       cp->program.sh.data->LinkStatus = LINKING_FAILURE;
  98       const char *error_str =
  99          "Compute shader used more than 64KB of shared variables";
 100       ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
 101       _mesa_problem(NULL, "Failed to link compute shader: %s\n", error_str);
 102
 103       ralloc_free(mem_ctx);
 104       return false;
 105    }
 106
 107    assign_cs_binding_table_offsets(devinfo, &cp->program, &prog_data);
 108
 109    brw_nir_setup_glsl_uniforms(mem_ctx, nir,
 110                                &cp->program, &prog_data.base, true);
 111
 112    if (unlikely(brw->perf_debug)) {
 113       start_busy = (brw->batch.last_bo &&
 114                     brw_bo_busy(brw->batch.last_bo));
 115       start_time = get_time();
 116    }
 117
 118    int st_index = -1;
 119    if (INTEL_DEBUG & DEBUG_SHADER_TIME)
 120       st_index = brw_get_shader_time_index(brw, &cp->program, ST_CS, true);
 121
 122    /* If the work group size is variable we set it to the maximum here since
 123     * the actual size is not known until the dispatch command is issued.
 124     */
 125    if (nir->info.cs.local_size_variable) {
 126       nir->info.cs.max_variable_local_size =
 127          gl_ctx->Const.MaxComputeVariableGroupInvocations;
 128    }
 129
 130    brw_nir_lower_cs_intrinsics(nir);
 131
 132    char *error_str;
 133    program = brw_compile_cs(brw->screen->compiler, brw, mem_ctx, key,
 134                             &prog_data, nir, st_index, NULL, &error_str);
 135    if (program == NULL) {
 136       cp->program.sh.data->LinkStatus = LINKING_FAILURE;
 137       ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
 138       _mesa_problem(NULL, "Failed to compile compute shader: %s\n", error_str);
 139
 140       ralloc_free(mem_ctx);
 141       return false;
 142    }
 143
 144    if (unlikely(brw->perf_debug)) {
 145       if (cp->compiled_once) {
 146          brw_debug_recompile(brw, MESA_SHADER_COMPUTE, cp->program.Id,
 147                              &key->base);
 148       }
 149       cp->compiled_once = true;
 150
 151       if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
 152          perf_debug("CS compile took %.03f ms and stalled the GPU\n",
 153                     (get_time() - start_time) * 1000);
 154       }
 155    }
 156
 157    brw_alloc_stage_scratch(brw, &brw->cs.base, prog_data.base.total_scratch);
 158
 159    /* The param and pull_param arrays will be freed by the shader cache. */
 160    ralloc_steal(NULL, prog_data.base.param);
 161    ralloc_steal(NULL, prog_data.base.pull_param);
 162    brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
 163                     key, sizeof(*key),
 164                     program, prog_data.base.program_size,
 165                     &prog_data, sizeof(prog_data),
 166                     &brw->cs.base.prog_offset, &brw->cs.base.prog_data);
 167    ralloc_free(mem_ctx);
 168
 169    return true;
 170 }
 171
 172
 173 void
 174 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
 175 {
 176    struct gl_context *ctx = &brw->ctx;
 177    /* BRW_NEW_COMPUTE_PROGRAM */
 178    const struct brw_program *cp =
 179       (struct brw_program *) brw->programs[MESA_SHADER_COMPUTE];
 180
 181    memset(key, 0, sizeof(*key));
 182
 183    /* _NEW_TEXTURE */
 184    brw_populate_base_prog_key(ctx, cp, &key->base);
 185 }
 186
 187
 188 void
 189 brw_upload_cs_prog(struct brw_context *brw)
 190 {
 191    struct gl_context *ctx = &brw->ctx;
 192    struct brw_cs_prog_key key;
 193    struct brw_program *cp =
 194       (struct brw_program *) brw->programs[MESA_SHADER_COMPUTE];
 195
 196    if (!cp)
 197       return;
 198
 199    if (!brw_state_dirty(brw, _NEW_TEXTURE, BRW_NEW_COMPUTE_PROGRAM))
 200       return;
 201
 202    brw->cs.base.sampler_count =
 203       util_last_bit(ctx->ComputeProgram._Current->SamplersUsed);
 204
 205    brw_cs_populate_key(brw, &key);
 206
 207    if (brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG, &key, sizeof(key),
 208                         &brw->cs.base.prog_offset, &brw->cs.base.prog_data,
 209                         true))
 210       return;
 211
 212    if (brw_disk_cache_upload_program(brw, MESA_SHADER_COMPUTE))
 213       return;
 214
 215    cp = (struct brw_program *) brw->programs[MESA_SHADER_COMPUTE];
 216    cp->id = key.base.program_string_id;
 217
 218    ASSERTED bool success = brw_codegen_cs_prog(brw, cp, &key);
 219    assert(success);
 220 }
 221
 222 void
 223 brw_cs_populate_default_key(const struct brw_compiler *compiler,
 224                             struct brw_cs_prog_key *key,
 225                             struct gl_program *prog)
 226 {
 227    const struct gen_device_info *devinfo = compiler->devinfo;
 228    memset(key, 0, sizeof(*key));
 229    brw_populate_default_base_prog_key(devinfo, brw_program(prog), &key->base);
 230 }
 231
 232 bool
 233 brw_cs_precompile(struct gl_context *ctx, struct gl_program *prog)
 234 {
 235    struct brw_context *brw = brw_context(ctx);
 236    struct brw_cs_prog_key key;
 237
 238    struct brw_program *bcp = brw_program(prog);
 239
 240    brw_cs_populate_default_key(brw->screen->compiler, &key, prog);
 241
 242    uint32_t old_prog_offset = brw->cs.base.prog_offset;
 243    struct brw_stage_prog_data *old_prog_data = brw->cs.base.prog_data;
 244
 245    bool success = brw_codegen_cs_prog(brw, bcp, &key);
 246
 247    brw->cs.base.prog_offset = old_prog_offset;
 248    brw->cs.base.prog_data = old_prog_data;
 249
 250    return success;
 251 }