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/intel_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_fdiv = true,                                                        \
34   .lower_scmp = true,                                                        \
35   .lower_flrp16 = true,                                                      \
36   .lower_fmod = true,                                                        \
37   .lower_bitfield_extract = true,                                            \
38   .lower_bitfield_insert = true,                                             \
39   .lower_uadd_carry = true,                                                  \
40   .lower_usub_borrow = true,                                                 \
41   .lower_flrp64 = true,                                                      \
42   .lower_isign = true,                                                       \
43   .lower_ldexp = true,                                                       \
44   .lower_device_index_to_zero = true,                                        \
45   .vectorize_io = true,                                                      \
46   .use_interpolated_input_intrinsics = true,                                 \
47   .lower_insert_byte = true,                                                 \
48   .lower_insert_word = true,                                                 \
49   .vertex_id_zero_based = true,                                              \
50   .lower_base_vertex = true,                                                 \
51   .use_scoped_barrier = true,                                                \
52   .support_16bit_alu = true,                                                 \
53   .lower_uniforms_to_ubo = true,                                             \
54   .has_txs = true
55
56#define COMMON_SCALAR_OPTIONS                                                 \
57   .lower_to_scalar = true,                                                   \
58   .lower_pack_half_2x16 = true,                                              \
59   .lower_pack_snorm_2x16 = true,                                             \
60   .lower_pack_snorm_4x8 = true,                                              \
61   .lower_pack_unorm_2x16 = true,                                             \
62   .lower_pack_unorm_4x8 = true,                                              \
63   .lower_unpack_half_2x16 = true,                                            \
64   .lower_unpack_snorm_2x16 = true,                                           \
65   .lower_unpack_snorm_4x8 = true,                                            \
66   .lower_unpack_unorm_2x16 = true,                                           \
67   .lower_unpack_unorm_4x8 = true,                                            \
68   .lower_usub_sat64 = true,                                                  \
69   .lower_hadd64 = true,                                                      \
70   .avoid_ternary_with_two_constants = true,                                  \
71   .has_pack_32_4x8 = true,                                                   \
72   .max_unroll_iterations = 32,                                               \
73   .force_indirect_unrolling = nir_var_function_temp
74
75static const struct nir_shader_compiler_options scalar_nir_options = {
76   COMMON_OPTIONS,
77   COMMON_SCALAR_OPTIONS,
78};
79
80static const struct nir_shader_compiler_options vector_nir_options = {
81   COMMON_OPTIONS,
82
83   /* In the vec4 backend, our dpN instruction replicates its result to all the
84    * components of a vec4.  We would like NIR to give us replicated fdot
85    * instructions because it can optimize better for us.
86    */
87   .fdot_replicates = true,
88
89   .lower_pack_snorm_2x16 = true,
90   .lower_pack_unorm_2x16 = true,
91   .lower_unpack_snorm_2x16 = true,
92   .lower_unpack_unorm_2x16 = true,
93   .lower_extract_byte = true,
94   .lower_extract_word = true,
95   .intel_vec4 = true,
96   .max_unroll_iterations = 32,
97};
98
99struct brw_compiler *
100brw_compiler_create(void *mem_ctx, const struct intel_device_info *devinfo)
101{
102   struct brw_compiler *compiler = rzalloc(mem_ctx, struct brw_compiler);
103
104   compiler->devinfo = devinfo;
105
106   brw_fs_alloc_reg_sets(compiler);
107   brw_vec4_alloc_reg_set(compiler);
108
109   compiler->precise_trig = env_var_as_boolean("INTEL_PRECISE_TRIG", false);
110
111   compiler->use_tcs_8_patch =
112      devinfo->ver >= 12 ||
113      (devinfo->ver >= 9 && INTEL_DEBUG(DEBUG_TCS_EIGHT_PATCH));
114
115   /* Default to the sampler since that's what we've done since forever */
116   compiler->indirect_ubos_use_sampler = true;
117
118   /* There is no vec4 mode on Gfx10+, and we don't use it at all on Gfx8+. */
119   for (int i = MESA_SHADER_VERTEX; i < MESA_ALL_SHADER_STAGES; i++) {
120      compiler->scalar_stage[i] = devinfo->ver >= 8 ||
121         i == MESA_SHADER_FRAGMENT || i == MESA_SHADER_COMPUTE;
122   }
123
124   for (int i = MESA_SHADER_TASK; i < MESA_VULKAN_SHADER_STAGES; i++)
125      compiler->scalar_stage[i] = true;
126
127   nir_lower_int64_options int64_options =
128      nir_lower_imul64 |
129      nir_lower_isign64 |
130      nir_lower_divmod64 |
131      nir_lower_imul_high64;
132   nir_lower_doubles_options fp64_options =
133      nir_lower_drcp |
134      nir_lower_dsqrt |
135      nir_lower_drsq |
136      nir_lower_dtrunc |
137      nir_lower_dfloor |
138      nir_lower_dceil |
139      nir_lower_dfract |
140      nir_lower_dround_even |
141      nir_lower_dmod |
142      nir_lower_dsub |
143      nir_lower_ddiv;
144
145   if (!devinfo->has_64bit_float || INTEL_DEBUG(DEBUG_SOFT64)) {
146      int64_options |= (nir_lower_int64_options)~0;
147      fp64_options |= nir_lower_fp64_full_software;
148   }
149
150   /* The Bspec's section tittled "Instruction_multiply[DevBDW+]" claims that
151    * destination type can be Quadword and source type Doubleword for Gfx8 and
152    * Gfx9. So, lower 64 bit multiply instruction on rest of the platforms.
153    */
154   if (devinfo->ver < 8 || devinfo->ver > 9)
155      int64_options |= nir_lower_imul_2x32_64;
156
157   /* We want the GLSL compiler to emit code that uses condition codes */
158   for (int i = 0; i < MESA_ALL_SHADER_STAGES; i++) {
159      compiler->glsl_compiler_options[i].MaxUnrollIterations = 0;
160      compiler->glsl_compiler_options[i].MaxIfDepth =
161         devinfo->ver < 6 ? 16 : UINT_MAX;
162
163      /* We handle this in NIR */
164      compiler->glsl_compiler_options[i].EmitNoIndirectInput = false;
165      compiler->glsl_compiler_options[i].EmitNoIndirectOutput = false;
166      compiler->glsl_compiler_options[i].EmitNoIndirectUniform = false;
167      compiler->glsl_compiler_options[i].EmitNoIndirectTemp = false;
168
169      bool is_scalar = compiler->scalar_stage[i];
170      compiler->glsl_compiler_options[i].OptimizeForAOS = !is_scalar;
171
172      struct nir_shader_compiler_options *nir_options =
173         rzalloc(compiler, struct nir_shader_compiler_options);
174      if (is_scalar) {
175         *nir_options = scalar_nir_options;
176      } else {
177         *nir_options = vector_nir_options;
178      }
179
180      /* Prior to Gfx6, there are no three source operations, and Gfx11 loses
181       * LRP.
182       */
183      nir_options->lower_ffma16 = devinfo->ver < 6;
184      nir_options->lower_ffma32 = devinfo->ver < 6;
185      nir_options->lower_ffma64 = devinfo->ver < 6;
186      nir_options->lower_flrp32 = devinfo->ver < 6 || devinfo->ver >= 11;
187      nir_options->lower_fpow = devinfo->ver >= 12;
188
189      nir_options->lower_rotate = devinfo->ver < 11;
190      nir_options->lower_bitfield_reverse = devinfo->ver < 7;
191      nir_options->has_iadd3 = devinfo->verx10 >= 125;
192
193      nir_options->has_dot_4x8 = devinfo->ver >= 12;
194      nir_options->has_sudot_4x8 = devinfo->ver >= 12;
195
196      nir_options->lower_int64_options = int64_options;
197      nir_options->lower_doubles_options = fp64_options;
198
199      nir_options->unify_interfaces = i < MESA_SHADER_FRAGMENT;
200
201      nir_options->force_indirect_unrolling |=
202         brw_nir_no_indirect_mask(compiler, i);
203
204      compiler->glsl_compiler_options[i].NirOptions = nir_options;
205
206      compiler->glsl_compiler_options[i].ClampBlockIndicesToArrayBounds = true;
207   }
208
209   return compiler;
210}
211
212static void
213insert_u64_bit(uint64_t *val, bool add)
214{
215   *val = (*val << 1) | !!add;
216}
217
218uint64_t
219brw_get_compiler_config_value(const struct brw_compiler *compiler)
220{
221   uint64_t config = 0;
222   insert_u64_bit(&config, compiler->precise_trig);
223   if (compiler->devinfo->ver >= 8 && compiler->devinfo->ver < 10) {
224      insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_VERTEX]);
225      insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_TESS_CTRL]);
226      insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_TESS_EVAL]);
227      insert_u64_bit(&config, compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
228   }
229   uint64_t mask = DEBUG_DISK_CACHE_MASK;
230   while (mask != 0) {
231      const uint64_t bit = 1ULL << (ffsll(mask) - 1);
232      insert_u64_bit(&config, INTEL_DEBUG(bit));
233      mask &= ~bit;
234   }
235   return config;
236}
237
238unsigned
239brw_prog_data_size(gl_shader_stage stage)
240{
241   static const size_t stage_sizes[] = {
242      [MESA_SHADER_VERTEX]       = sizeof(struct brw_vs_prog_data),
243      [MESA_SHADER_TESS_CTRL]    = sizeof(struct brw_tcs_prog_data),
244      [MESA_SHADER_TESS_EVAL]    = sizeof(struct brw_tes_prog_data),
245      [MESA_SHADER_GEOMETRY]     = sizeof(struct brw_gs_prog_data),
246      [MESA_SHADER_FRAGMENT]     = sizeof(struct brw_wm_prog_data),
247      [MESA_SHADER_COMPUTE]      = sizeof(struct brw_cs_prog_data),
248      [MESA_SHADER_RAYGEN]       = sizeof(struct brw_bs_prog_data),
249      [MESA_SHADER_ANY_HIT]      = sizeof(struct brw_bs_prog_data),
250      [MESA_SHADER_CLOSEST_HIT]  = sizeof(struct brw_bs_prog_data),
251      [MESA_SHADER_MISS]         = sizeof(struct brw_bs_prog_data),
252      [MESA_SHADER_INTERSECTION] = sizeof(struct brw_bs_prog_data),
253      [MESA_SHADER_CALLABLE]     = sizeof(struct brw_bs_prog_data),
254      [MESA_SHADER_KERNEL]       = sizeof(struct brw_cs_prog_data),
255   };
256   assert((int)stage >= 0 && stage < ARRAY_SIZE(stage_sizes));
257   return stage_sizes[stage];
258}
259
260unsigned
261brw_prog_key_size(gl_shader_stage stage)
262{
263   static const size_t stage_sizes[] = {
264      [MESA_SHADER_VERTEX]       = sizeof(struct brw_vs_prog_key),
265      [MESA_SHADER_TESS_CTRL]    = sizeof(struct brw_tcs_prog_key),
266      [MESA_SHADER_TESS_EVAL]    = sizeof(struct brw_tes_prog_key),
267      [MESA_SHADER_GEOMETRY]     = sizeof(struct brw_gs_prog_key),
268      [MESA_SHADER_FRAGMENT]     = sizeof(struct brw_wm_prog_key),
269      [MESA_SHADER_COMPUTE]      = sizeof(struct brw_cs_prog_key),
270      [MESA_SHADER_RAYGEN]       = sizeof(struct brw_bs_prog_key),
271      [MESA_SHADER_ANY_HIT]      = sizeof(struct brw_bs_prog_key),
272      [MESA_SHADER_CLOSEST_HIT]  = sizeof(struct brw_bs_prog_key),
273      [MESA_SHADER_MISS]         = sizeof(struct brw_bs_prog_key),
274      [MESA_SHADER_INTERSECTION] = sizeof(struct brw_bs_prog_key),
275      [MESA_SHADER_CALLABLE]     = sizeof(struct brw_bs_prog_key),
276      [MESA_SHADER_KERNEL]       = sizeof(struct brw_cs_prog_key),
277   };
278   assert((int)stage >= 0 && stage < ARRAY_SIZE(stage_sizes));
279   return stage_sizes[stage];
280}
281
282void
283brw_write_shader_relocs(const struct intel_device_info *devinfo,
284                        void *program,
285                        const struct brw_stage_prog_data *prog_data,
286                        struct brw_shader_reloc_value *values,
287                        unsigned num_values)
288{
289   for (unsigned i = 0; i < prog_data->num_relocs; i++) {
290      assert(prog_data->relocs[i].offset % 8 == 0);
291      void *dst = program + prog_data->relocs[i].offset;
292      for (unsigned j = 0; j < num_values; j++) {
293         if (prog_data->relocs[i].id == values[j].id) {
294            uint32_t value = values[j].value + prog_data->relocs[i].delta;
295            switch (prog_data->relocs[i].type) {
296            case BRW_SHADER_RELOC_TYPE_U32:
297               *(uint32_t *)dst = value;
298               break;
299            case BRW_SHADER_RELOC_TYPE_MOV_IMM:
300               brw_update_reloc_imm(devinfo, dst, value);
301               break;
302            default:
303               unreachable("Invalid relocation type");
304            }
305            break;
306         }
307      }
308   }
309}
310