src/gallium/drivers/freedreno/ir3/ir3_gallium.c

   1 /*
   2  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "pipe/p_state.h"
  28 #include "pipe/p_screen.h"
  29 #include "util/u_string.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_inlines.h"
  32 #include "util/format/u_format.h"
  33 #include "tgsi/tgsi_dump.h"
  34 #include "tgsi/tgsi_parse.h"
  35
  36 #include "nir/tgsi_to_nir.h"
  37
  38 #include "freedreno_context.h"
  39 #include "freedreno_util.h"
  40
  41 #include "ir3/ir3_shader.h"
  42 #include "ir3/ir3_gallium.h"
  43 #include "ir3/ir3_compiler.h"
  44 #include "ir3/ir3_nir.h"
  45
  46 static void
  47 dump_shader_info(struct ir3_shader_variant *v, bool binning_pass,
  48                 struct pipe_debug_callback *debug)
  49 {
  50         if (!unlikely(fd_mesa_debug & FD_DBG_SHADERDB))
  51                 return;
  52
  53         pipe_debug_message(debug, SHADER_INFO,
  54                         "%s shader: %u inst, %u nops, %u non-nops, %u mov, %u cov, "
  55                         "%u dwords, %u last-baryf, %u half, %u full, %u constlen, "
  56                         "%u sstall, %u (ss), %u (sy), %d max_sun, %d loops\n",
  57                         ir3_shader_stage(v),
  58                         v->info.instrs_count,
  59                         v->info.nops_count,
  60                         v->info.instrs_count - v->info.nops_count,
  61                         v->info.mov_count,
  62                         v->info.cov_count,
  63                         v->info.sizedwords,
  64                         v->info.last_baryf,
  65                         v->info.max_half_reg + 1,
  66                         v->info.max_reg + 1,
  67                         v->constlen,
  68                         v->info.sstall,
  69                         v->info.ss, v->info.sy,
  70                         v->max_sun, v->loops);
  71 }
  72
  73 static void
  74 upload_shader_variant(struct ir3_shader_variant *v)
  75 {
  76         struct shader_info *info = &v->shader->nir->info;
  77         struct ir3_compiler *compiler = v->shader->compiler;
  78
  79         assert(!v->bo);
  80
  81         unsigned sz = v->info.sizedwords * 4;
  82
  83         v->bo = fd_bo_new(compiler->dev, sz,
  84                         DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
  85                         DRM_FREEDRENO_GEM_TYPE_KMEM,
  86                         "%s:%s", ir3_shader_stage(v), info->name);
  87
  88         /* Always include shaders in kernel crash dumps. */
  89         fd_bo_mark_for_dump(v->bo);
  90
  91         memcpy(fd_bo_map(v->bo), v->bin, sz);
  92 }
  93
  94 struct ir3_shader_variant *
  95 ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key,
  96                 bool binning_pass, struct pipe_debug_callback *debug)
  97 {
  98         struct ir3_shader_variant *v;
  99         bool created = false;
 100
 101         /* Some shader key values may not be used by a given ir3_shader (for
 102          * example, fragment shader saturates in the vertex shader), so clean out
 103          * those flags to avoid recompiling.
 104          */
 105         ir3_key_clear_unused(&key, shader);
 106
 107         v = ir3_shader_get_variant(shader, &key, binning_pass, &created);
 108
 109         if (created) {
 110                 if (shader->initial_variants_done) {
 111                         pipe_debug_message(debug, SHADER_INFO,
 112                                         "%s shader: recompiling at draw time: global 0x%08x, vsats %x/%x/%x, fsats %x/%x/%x, vfsamples %x/%x, astc %x/%x\n",
 113                                         ir3_shader_stage(v),
 114                                         key.global,
 115                                         key.vsaturate_s, key.vsaturate_t, key.vsaturate_r,
 116                                         key.fsaturate_s, key.fsaturate_t, key.fsaturate_r,
 117                                         key.vsamples, key.fsamples,
 118                                         key.vastc_srgb, key.fastc_srgb);
 119
 120                 }
 121                 dump_shader_info(v, binning_pass, debug);
 122
 123                 upload_shader_variant(v);
 124         }
 125
 126         return v;
 127 }
 128
 129 static void
 130 copy_stream_out(struct ir3_stream_output_info *i,
 131                 const struct pipe_stream_output_info *p)
 132 {
 133         STATIC_ASSERT(ARRAY_SIZE(i->stride) == ARRAY_SIZE(p->stride));
 134         STATIC_ASSERT(ARRAY_SIZE(i->output) == ARRAY_SIZE(p->output));
 135
 136         i->num_outputs = p->num_outputs;
 137         for (int n = 0; n < ARRAY_SIZE(i->stride); n++)
 138                 i->stride[n] = p->stride[n];
 139
 140         for (int n = 0; n < ARRAY_SIZE(i->output); n++) {
 141                 i->output[n].register_index  = p->output[n].register_index;
 142                 i->output[n].start_component = p->output[n].start_component;
 143                 i->output[n].num_components  = p->output[n].num_components;
 144                 i->output[n].output_buffer   = p->output[n].output_buffer;
 145                 i->output[n].dst_offset      = p->output[n].dst_offset;
 146                 i->output[n].stream          = p->output[n].stream;
 147         }
 148 }
 149
 150 struct ir3_shader *
 151 ir3_shader_create(struct ir3_compiler *compiler,
 152                 const struct pipe_shader_state *cso,
 153                 struct pipe_debug_callback *debug,
 154                 struct pipe_screen *screen)
 155 {
 156         nir_shader *nir;
 157         if (cso->type == PIPE_SHADER_IR_NIR) {
 158                 /* we take ownership of the reference: */
 159                 nir = cso->ir.nir;
 160         } else {
 161                 debug_assert(cso->type == PIPE_SHADER_IR_TGSI);
 162                 if (ir3_shader_debug & IR3_DBG_DISASM) {
 163                         tgsi_dump(cso->tokens, 0);
 164                 }
 165                 nir = tgsi_to_nir(cso->tokens, screen, false);
 166         }
 167
 168         struct ir3_stream_output_info stream_output;
 169         copy_stream_out(&stream_output, &cso->stream_output);
 170
 171         struct ir3_shader *shader = ir3_shader_from_nir(compiler, nir, &stream_output);
 172
 173         /* Compile standard variants immediately to try to avoid draw-time stalls
 174          * to run the compiler.
 175          */
 176         struct ir3_shader_key key = {
 177                 .tessellation = IR3_TESS_NONE,
 178                 .msaa = true,
 179         };
 180
 181         switch (nir->info.stage) {
 182         case MESA_SHADER_TESS_EVAL:
 183                 key.tessellation = ir3_tess_mode(nir->info.tess.primitive_mode);
 184                 break;
 185
 186         case MESA_SHADER_TESS_CTRL:
 187                 /* The primitive_mode field, while it exists for TCS, is not
 188                  * populated (since separable shaders between TCS/TES are legal,
 189                  * so TCS wouldn't have access to TES's declaration).  Make a
 190                  * guess so that we shader-db something plausible for TCS.
 191                  */
 192                 if (nir->info.outputs_written & VARYING_BIT_TESS_LEVEL_INNER)
 193                         key.tessellation = IR3_TESS_TRIANGLES;
 194                 else
 195                         key.tessellation = IR3_TESS_ISOLINES;
 196                 break;
 197
 198         case MESA_SHADER_GEOMETRY:
 199                 key.has_gs = true;
 200                 break;
 201
 202         default:
 203                 break;
 204         }
 205
 206         ir3_shader_variant(shader, key, false, debug);
 207
 208         if (nir->info.stage == MESA_SHADER_VERTEX)
 209                 ir3_shader_variant(shader, key, true, debug);
 210
 211         shader->initial_variants_done = true;
 212
 213         return shader;
 214 }
 215
 216 /* a bit annoying that compute-shader and normal shader state objects
 217  * aren't a bit more aligned.
 218  */
 219 struct ir3_shader *
 220 ir3_shader_create_compute(struct ir3_compiler *compiler,
 221                 const struct pipe_compute_state *cso,
 222                 struct pipe_debug_callback *debug,
 223                 struct pipe_screen *screen)
 224 {
 225         nir_shader *nir;
 226         if (cso->ir_type == PIPE_SHADER_IR_NIR) {
 227                 /* we take ownership of the reference: */
 228                 nir = (nir_shader *)cso->prog;
 229         } else {
 230                 debug_assert(cso->ir_type == PIPE_SHADER_IR_TGSI);
 231                 if (ir3_shader_debug & IR3_DBG_DISASM) {
 232                         tgsi_dump(cso->prog, 0);
 233                 }
 234                 nir = tgsi_to_nir(cso->prog, screen, false);
 235         }
 236
 237         struct ir3_shader *shader = ir3_shader_from_nir(compiler, nir, NULL);
 238
 239         /* Immediately compile a standard variant.  We have so few variants in our
 240          * shaders, that doing so almost eliminates draw-time recompiles.  (This
 241          * is also how we get data from shader-db's ./run)
 242          */
 243         static struct ir3_shader_key key; /* static is implicitly zeroed */
 244         ir3_shader_variant(shader, key, false, debug);
 245
 246         shader->initial_variants_done = true;
 247
 248         return shader;
 249 }
 250
 251 void *
 252 ir3_shader_state_create(struct pipe_context *pctx, const struct pipe_shader_state *cso)
 253 {
 254         struct fd_context *ctx = fd_context(pctx);
 255         struct ir3_compiler *compiler = ctx->screen->compiler;
 256         return ir3_shader_create(compiler, cso, &ctx->debug, pctx->screen);
 257 }
 258
 259 void
 260 ir3_shader_state_delete(struct pipe_context *pctx, void *hwcso)
 261 {
 262         struct ir3_shader *so = hwcso;
 263
 264         /* free the uploaded shaders, since this is handled outside of the
 265          * shared ir3 code (ie. not used by turnip):
 266          */
 267         for (struct ir3_shader_variant *v = so->variants; v; v = v->next) {
 268                 fd_bo_del(v->bo);
 269                 v->bo = NULL;
 270
 271                 if (v->binning) {
 272                         fd_bo_del(v->binning->bo);
 273                         v->binning->bo = NULL;
 274                 }
 275         }
 276
 277         ir3_shader_destroy(so);
 278 }
 279
 280 void
 281 ir3_prog_init(struct pipe_context *pctx)
 282 {
 283         pctx->create_vs_state = ir3_shader_state_create;
 284         pctx->delete_vs_state = ir3_shader_state_delete;
 285
 286         pctx->create_tcs_state = ir3_shader_state_create;
 287         pctx->delete_tcs_state = ir3_shader_state_delete;
 288
 289         pctx->create_tes_state = ir3_shader_state_create;
 290         pctx->delete_tes_state = ir3_shader_state_delete;
 291
 292         pctx->create_gs_state = ir3_shader_state_create;
 293         pctx->delete_gs_state = ir3_shader_state_delete;
 294
 295         pctx->create_fs_state = ir3_shader_state_create;
 296         pctx->delete_fs_state = ir3_shader_state_delete;
 297 }