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