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