src/intel/compiler/brw_compiler.c

   1 /*
   2  * Copyright © 2015-2016 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 DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "brw_compiler.h"
  25 #include "brw_shader.h"
  26 #include "brw_eu.h"
  27 #include "dev/gen_debug.h"
  28 #include "compiler/nir/nir.h"
  29 #include "main/errors.h"
  30 #include "util/debug.h"
  31
  32 #define COMMON_OPTIONS                                                        \
  33    .lower_sub = true,                                                         \
  34    .lower_fdiv = true,                                                        \
  35    .lower_scmp = true,                                                        \
  36    .lower_flrp16 = true,                                                      \
  37    .lower_fmod = true,                                                        \
  38    .lower_bitfield_extract = true,                                            \
  39    .lower_bitfield_insert = true,                                             \
  40    .lower_uadd_carry = true,                                                  \
  41    .lower_usub_borrow = true,                                                 \
  42    .lower_fdiv = true,                                                        \
  43    .lower_flrp64 = true,                                                      \
  44    .lower_isign = true,                                                       \
  45    .lower_ldexp = true,                                                       \
  46    .lower_device_index_to_zero = true,                                        \
  47    .vectorize_io = true,                                                      \
  48    .use_interpolated_input_intrinsics = true,                                 \
  49    .vertex_id_zero_based = true,                                              \
  50    .lower_base_vertex = true,                                                 \
  51    .use_scoped_memory_barrier = true,                                         \
  52    .support_8bit_alu = true,                                                  \
  53    .support_16bit_alu = true
  54
  55 #define COMMON_SCALAR_OPTIONS                                                 \
  56    .lower_to_scalar = true,                                                   \
  57    .lower_pack_half_2x16 = true,                                              \
  58    .lower_pack_snorm_2x16 = true,                                             \
  59    .lower_pack_snorm_4x8 = true,                                              \
  60    .lower_pack_unorm_2x16 = true,                                             \
  61    .lower_pack_unorm_4x8 = true,                                              \
  62    .lower_unpack_half_2x16 = true,                                            \
  63    .lower_unpack_snorm_2x16 = true,                                           \
  64    .lower_unpack_snorm_4x8 = true,                                            \
  65    .lower_unpack_unorm_2x16 = true,                                           \
  66    .lower_unpack_unorm_4x8 = true,                                            \
  67    .lower_usub_sat64 = true,                                                  \
  68    .lower_hadd64 = true,                                                      \
  69    .lower_bfe_with_two_constants = true,                                      \
  70    .max_unroll_iterations = 32
  71
  72 static const struct nir_shader_compiler_options scalar_nir_options = {
  73    COMMON_OPTIONS,
  74    COMMON_SCALAR_OPTIONS,
  75 };
  76
  77 static const struct nir_shader_compiler_options vector_nir_options = {
  78    COMMON_OPTIONS,
  79
  80    /* In the vec4 backend, our dpN instruction replicates its result to all the
  81     * components of a vec4.  We would like NIR to give us replicated fdot
  82     * instructions because it can optimize better for us.
  83     */
  84    .fdot_replicates = true,
  85
  86    .lower_pack_snorm_2x16 = true,
  87    .lower_pack_unorm_2x16 = true,
  88    .lower_unpack_snorm_2x16 = true,
  89    .lower_unpack_unorm_2x16 = true,
  90    .lower_extract_byte = true,
  91    .lower_extract_word = true,
  92    .intel_vec4 = true,
  93    .max_unroll_iterations = 32,
  94 };
  95
  96 struct brw_compiler *
  97 brw_compiler_create(void *mem_ctx, const struct gen_device_info *devinfo)
  98 {
  99    struct brw_compiler *compiler = rzalloc(mem_ctx, struct brw_compiler);
 100
 101    compiler->devinfo = devinfo;
 102
 103    brw_fs_alloc_reg_sets(compiler);
 104    brw_vec4_alloc_reg_set(compiler);
 105    brw_init_compaction_tables(devinfo);
 106
 107    compiler->precise_trig = env_var_as_boolean("INTEL_PRECISE_TRIG", false);
 108
 109    compiler->use_tcs_8_patch =
 110       devinfo->gen >= 12 ||
 111       (devinfo->gen >= 9 && (INTEL_DEBUG & DEBUG_TCS_EIGHT_PATCH));
 112
 113    if (devinfo->gen >= 10) {
 114       /* We don't support vec4 mode on Cannonlake. */
 115       for (int i = MESA_SHADER_VERTEX; i < MESA_SHADER_STAGES; i++)
 116          compiler->scalar_stage[i] = true;
 117    } else {
 118       compiler->scalar_stage[MESA_SHADER_VERTEX] =
 119          devinfo->gen >= 8 && env_var_as_boolean("INTEL_SCALAR_VS", true);
 120       compiler->scalar_stage[MESA_SHADER_TESS_CTRL] =
 121          devinfo->gen >= 8 && env_var_as_boolean("INTEL_SCALAR_TCS", true);
 122       compiler->scalar_stage[MESA_SHADER_TESS_EVAL] =
 123          devinfo->gen >= 8 && env_var_as_boolean("INTEL_SCALAR_TES", true);
 124       compiler->scalar_stage[MESA_SHADER_GEOMETRY] =
 125          devinfo->gen >= 8 && env_var_as_boolean("INTEL_SCALAR_GS", true);
 126       compiler->scalar_stage[MESA_SHADER_FRAGMENT] = true;
 127       compiler->scalar_stage[MESA_SHADER_COMPUTE] = true;
 128    }
 129
 130    nir_lower_int64_options int64_options =
 131       nir_lower_imul64 |
 132       nir_lower_isign64 |
 133       nir_lower_divmod64 |
 134       nir_lower_imul_high64;
 135    nir_lower_doubles_options fp64_options =
 136       nir_lower_drcp |
 137       nir_lower_dsqrt |
 138       nir_lower_drsq |
 139       nir_lower_dtrunc |
 140       nir_lower_dfloor |
 141       nir_lower_dceil |
 142       nir_lower_dfract |
 143       nir_lower_dround_even |
 144       nir_lower_dmod |
 145       nir_lower_dsub |
 146       nir_lower_ddiv;
 147
 148    if (!devinfo->has_64bit_float || (INTEL_DEBUG & DEBUG_SOFT64)) {
 149       int64_options |= nir_lower_mov64 |
 150                        nir_lower_icmp64 |
 151                        nir_lower_iadd64 |
 152                        nir_lower_iabs64 |
 153                        nir_lower_ineg64 |
 154                        nir_lower_logic64 |
 155                        nir_lower_minmax64 |
 156                        nir_lower_shift64 |
 157                        nir_lower_extract64;
 158       fp64_options |= nir_lower_fp64_full_software;
 159    }
 160
 161    /* The Bspec's section tittled "Instruction_multiply[DevBDW+]" claims that
 162     * destination type can be Quadword and source type Doubleword for Gen8 and
 163     * Gen9. So, lower 64 bit multiply instruction on rest of the platforms.
 164     */
 165    if (devinfo->gen < 8 || devinfo->gen > 9)
 166       int64_options |= nir_lower_imul_2x32_64;
 167
 168    /* We want the GLSL compiler to emit code that uses condition codes */
 169    for (int i = 0; i < MESA_SHADER_STAGES; i++) {
 170       compiler->glsl_compiler_options[i].MaxUnrollIterations = 0;
 171       compiler->glsl_compiler_options[i].MaxIfDepth =
 172          devinfo->gen < 6 ? 16 : UINT_MAX;
 173
 174       compiler->glsl_compiler_options[i].EmitNoIndirectInput = true;
 175       compiler->glsl_compiler_options[i].EmitNoIndirectUniform = false;
 176
 177       bool is_scalar = compiler->scalar_stage[i];
 178
 179       compiler->glsl_compiler_options[i].EmitNoIndirectOutput = is_scalar;
 180       compiler->glsl_compiler_options[i].EmitNoIndirectTemp = is_scalar;
 181       compiler->glsl_compiler_options[i].OptimizeForAOS = !is_scalar;
 182
 183       struct nir_shader_compiler_options *nir_options =
 184          rzalloc(compiler, struct nir_shader_compiler_options);
 185       if (is_scalar) {
 186          *nir_options = scalar_nir_options;
 187       } else {
 188          *nir_options = vector_nir_options;
 189       }
 190
 191       /* Prior to Gen6, there are no three source operations, and Gen11 loses
 192        * LRP.
 193        */
 194       nir_options->lower_ffma = devinfo->gen < 6;
 195       nir_options->lower_flrp32 = devinfo->gen < 6 || devinfo->gen >= 11;
 196       nir_options->lower_fpow = devinfo->gen >= 12;
 197
 198       nir_options->lower_rotate = devinfo->gen < 11;
 199       nir_options->lower_bitfield_reverse = devinfo->gen < 7;
 200
 201       nir_options->lower_int64_options = int64_options;
 202       nir_options->lower_doubles_options = fp64_options;
 203
 204       nir_options->unify_interfaces = i < MESA_SHADER_FRAGMENT;
 205
 206       compiler->glsl_compiler_options[i].NirOptions = nir_options;
 207
 208       compiler->glsl_compiler_options[i].ClampBlockIndicesToArrayBounds = true;
 209    }
 210
 211    compiler->glsl_compiler_options[MESA_SHADER_TESS_CTRL].EmitNoIndirectInput = false;
 212    compiler->glsl_compiler_options[MESA_SHADER_TESS_EVAL].EmitNoIndirectInput = false;
 213    compiler->glsl_compiler_options[MESA_SHADER_TESS_CTRL].EmitNoIndirectOutput = false;
 214
 215    if (compiler->scalar_stage[MESA_SHADER_GEOMETRY])
 216       compiler->glsl_compiler_options[MESA_SHADER_GEOMETRY].EmitNoIndirectInput = false;
 217
 218    return compiler;
 219 }
 220
 221 static void
 222 insert_u64_bit(uint64_t *val, bool add)
 223 {
 224    *val = (*val << 1) | !!add;
 225 }
 226
 227 uint64_t
 228 brw_get_compiler_config_value(const struct brw_compiler *compiler)
 229 {
 230    uint64_t config = 0;
 231    insert_u64_bit(&config, compiler->precise_trig);
 232    if (compiler->devinfo->gen >= 8 && compiler->devinfo->gen < 10) {
 233       insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_VERTEX]);
 234       insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_TESS_CTRL]);
 235       insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_TESS_EVAL]);
 236       insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
 237    }
 238    uint64_t debug_bits = INTEL_DEBUG;
 239    uint64_t mask = DEBUG_DISK_CACHE_MASK;
 240    while (mask != 0) {
 241       const uint64_t bit = 1ULL << (ffsll(mask) - 1);
 242       insert_u64_bit(&config, (debug_bits & bit) != 0);
 243       mask &= ~bit;
 244    }
 245    return config;
 246 }
 247
 248 unsigned
 249 brw_prog_data_size(gl_shader_stage stage)
 250 {
 251    STATIC_ASSERT(MESA_SHADER_VERTEX == 0);
 252    STATIC_ASSERT(MESA_SHADER_TESS_CTRL == 1);
 253    STATIC_ASSERT(MESA_SHADER_TESS_EVAL == 2);
 254    STATIC_ASSERT(MESA_SHADER_GEOMETRY == 3);
 255    STATIC_ASSERT(MESA_SHADER_FRAGMENT == 4);
 256    STATIC_ASSERT(MESA_SHADER_COMPUTE == 5);
 257    static const size_t stage_sizes[] = {
 258       sizeof(struct brw_vs_prog_data),
 259       sizeof(struct brw_tcs_prog_data),
 260       sizeof(struct brw_tes_prog_data),
 261       sizeof(struct brw_gs_prog_data),
 262       sizeof(struct brw_wm_prog_data),
 263       sizeof(struct brw_cs_prog_data),
 264    };
 265    assert((int)stage >= 0 && stage < ARRAY_SIZE(stage_sizes));
 266    return stage_sizes[stage];
 267 }
 268
 269 unsigned
 270 brw_prog_key_size(gl_shader_stage stage)
 271 {
 272    static const size_t stage_sizes[] = {
 273       sizeof(struct brw_vs_prog_key),
 274       sizeof(struct brw_tcs_prog_key),
 275       sizeof(struct brw_tes_prog_key),
 276       sizeof(struct brw_gs_prog_key),
 277       sizeof(struct brw_wm_prog_key),
 278       sizeof(struct brw_cs_prog_key),
 279    };
 280    assert((int)stage >= 0 && stage < ARRAY_SIZE(stage_sizes));
 281    return stage_sizes[stage];
 282 }