src/gallium/drivers/freedreno/ir3/ir3_shader.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_string.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_inlines.h"
  33 #include "util/u_format.h"
  34 #include "tgsi/tgsi_dump.h"
  35 #include "tgsi/tgsi_parse.h"
  36
  37 #include "freedreno_context.h"
  38 #include "freedreno_util.h"
  39
  40 #include "ir3_shader.h"
  41 #include "ir3_compiler.h"
  42
  43
  44 static void
  45 delete_variant(struct ir3_shader_variant *v)
  46 {
  47         if (v->ir)
  48                 ir3_destroy(v->ir);
  49         fd_bo_del(v->bo);
  50         free(v);
  51 }
  52
  53 /* for vertex shader, the inputs are loaded into registers before the shader
  54  * is executed, so max_regs from the shader instructions might not properly
  55  * reflect the # of registers actually used, especially in case passthrough
  56  * varyings.
  57  *
  58  * Likewise, for fragment shader, we can have some regs which are passed
  59  * input values but never touched by the resulting shader (ie. as result
  60  * of dead code elimination or simply because we don't know how to turn
  61  * the reg off.
  62  */
  63 static void
  64 fixup_regfootprint(struct ir3_shader_variant *v)
  65 {
  66         if (v->type == SHADER_VERTEX) {
  67                 unsigned i;
  68                 for (i = 0; i < v->inputs_count; i++) {
  69                         /* skip frag inputs fetch via bary.f since their reg's are
  70                          * not written by gpu before shader starts (and in fact the
  71                          * regid's might not even be valid)
  72                          */
  73                         if (v->inputs[i].bary)
  74                                 continue;
  75
  76                         if (v->inputs[i].compmask) {
  77                                 int32_t regid = (v->inputs[i].regid + 3) >> 2;
  78                                 v->info.max_reg = MAX2(v->info.max_reg, regid);
  79                         }
  80                 }
  81                 for (i = 0; i < v->outputs_count; i++) {
  82                         int32_t regid = (v->outputs[i].regid + 3) >> 2;
  83                         v->info.max_reg = MAX2(v->info.max_reg, regid);
  84                 }
  85         } else if (v->type == SHADER_FRAGMENT) {
  86                 /* NOTE: not sure how to turn pos_regid off..  but this could
  87                  * be, for example, r1.x while max reg used by the shader is
  88                  * r0.*, in which case we need to fixup the reg footprint:
  89                  */
  90                 v->info.max_reg = MAX2(v->info.max_reg, v->pos_regid >> 2);
  91                 if (v->frag_coord)
  92                         debug_assert(v->info.max_reg >= 0); /* hard coded r0.x */
  93                 if (v->frag_face)
  94                         debug_assert(v->info.max_half_reg >= 0); /* hr0.x */
  95         }
  96 }
  97
  98 /* wrapper for ir3_assemble() which does some info fixup based on
  99  * shader state.  Non-static since used by ir3_cmdline too.
 100  */
 101 void * ir3_shader_assemble(struct ir3_shader_variant *v, uint32_t gpu_id)
 102 {
 103         void *bin;
 104
 105         bin = ir3_assemble(v->ir, &v->info, gpu_id);
 106         if (!bin)
 107                 return NULL;
 108
 109         if (gpu_id >= 400) {
 110                 v->instrlen = v->info.sizedwords / (2 * 16);
 111         } else {
 112                 v->instrlen = v->info.sizedwords / (2 * 4);
 113         }
 114
 115         /* NOTE: if relative addressing is used, we set constlen in
 116          * the compiler (to worst-case value) since we don't know in
 117          * the assembler what the max addr reg value can be:
 118          */
 119         v->constlen = MIN2(255, MAX2(v->constlen, v->info.max_const + 1));
 120
 121         fixup_regfootprint(v);
 122
 123         return bin;
 124 }
 125
 126 static void
 127 assemble_variant(struct ir3_shader_variant *v)
 128 {
 129         struct fd_context *ctx = fd_context(v->shader->pctx);
 130         uint32_t gpu_id = v->shader->compiler->gpu_id;
 131         uint32_t sz, *bin;
 132
 133         bin = ir3_shader_assemble(v, gpu_id);
 134         sz = v->info.sizedwords * 4;
 135
 136         v->bo = fd_bo_new(ctx->dev, sz,
 137                         DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
 138                         DRM_FREEDRENO_GEM_TYPE_KMEM);
 139
 140         memcpy(fd_bo_map(v->bo), bin, sz);
 141
 142         free(bin);
 143
 144         /* no need to keep the ir around beyond this point: */
 145         ir3_destroy(v->ir);
 146         v->ir = NULL;
 147 }
 148
 149 static struct ir3_shader_variant *
 150 create_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 151 {
 152         struct ir3_shader_variant *v = CALLOC_STRUCT(ir3_shader_variant);
 153         const struct tgsi_token *tokens = shader->tokens;
 154         int ret;
 155
 156         if (!v)
 157                 return NULL;
 158
 159         v->shader = shader;
 160         v->key = key;
 161         v->type = shader->type;
 162
 163         if (fd_mesa_debug & FD_DBG_DISASM) {
 164                 DBG("dump tgsi: type=%d, k={bp=%u,cts=%u,hp=%u}", shader->type,
 165                         key.binning_pass, key.color_two_side, key.half_precision);
 166                 tgsi_dump(tokens, 0);
 167         }
 168
 169         ret = ir3_compile_shader_nir(shader->compiler, v, tokens, key);
 170         if (ret) {
 171                 debug_error("compile failed!");
 172                 goto fail;
 173         }
 174
 175         assemble_variant(v);
 176         if (!v->bo) {
 177                 debug_error("assemble failed!");
 178                 goto fail;
 179         }
 180
 181         if (fd_mesa_debug & FD_DBG_DISASM) {
 182                 DBG("disassemble: type=%d, k={bp=%u,cts=%u,hp=%u}", v->type,
 183                         key.binning_pass, key.color_two_side, key.half_precision);
 184                 disasm_a3xx(fd_bo_map(v->bo), v->info.sizedwords, 0, v->type);
 185         }
 186
 187         return v;
 188
 189 fail:
 190         delete_variant(v);
 191         return NULL;
 192 }
 193
 194 struct ir3_shader_variant *
 195 ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 196 {
 197         struct ir3_shader_variant *v;
 198
 199         /* some shader key values only apply to vertex or frag shader,
 200          * so normalize the key to avoid constructing multiple identical
 201          * variants:
 202          */
 203         switch (shader->type) {
 204         case SHADER_FRAGMENT:
 205         case SHADER_COMPUTE:
 206                 key.binning_pass = false;
 207                 if (key.has_per_samp) {
 208                         key.vsaturate_s = 0;
 209                         key.vsaturate_t = 0;
 210                         key.vsaturate_r = 0;
 211                 }
 212                 break;
 213         case SHADER_VERTEX:
 214                 key.color_two_side = false;
 215                 key.half_precision = false;
 216                 key.rasterflat = false;
 217                 if (key.has_per_samp) {
 218                         key.fsaturate_s = 0;
 219                         key.fsaturate_t = 0;
 220                         key.fsaturate_r = 0;
 221                 }
 222                 break;
 223         }
 224
 225         for (v = shader->variants; v; v = v->next)
 226                 if (ir3_shader_key_equal(&key, &v->key))
 227                         return v;
 228
 229         /* compile new variant if it doesn't exist already: */
 230         v = create_variant(shader, key);
 231         v->next = shader->variants;
 232         shader->variants = v;
 233
 234         return v;
 235 }
 236
 237
 238 void
 239 ir3_shader_destroy(struct ir3_shader *shader)
 240 {
 241         struct ir3_shader_variant *v, *t;
 242         for (v = shader->variants; v; ) {
 243                 t = v;
 244                 v = v->next;
 245                 delete_variant(t);
 246         }
 247         free((void *)shader->tokens);
 248         free(shader);
 249 }
 250
 251 struct ir3_shader *
 252 ir3_shader_create(struct pipe_context *pctx, const struct tgsi_token *tokens,
 253                 enum shader_t type)
 254 {
 255         struct ir3_shader *shader = CALLOC_STRUCT(ir3_shader);
 256         shader->compiler = fd_context(pctx)->screen->compiler;
 257         shader->pctx = pctx;
 258         shader->type = type;
 259         shader->tokens = tgsi_dup_tokens(tokens);
 260         return shader;
 261 }