• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* -*- mesa-c++  -*-
2  * Copyright 2019 Collabora LTD
3  * Author: Gert Wollny <gert.wollny@collabora.com>
4  * SPDX-License-Identifier: MIT
5  */
6 
7 #include "r600_sfn.h"
8 
9 #include "compiler/nir/nir.h"
10 #include "compiler/shader_enums.h"
11 #include "sfn/sfn_assembler.h"
12 #include "sfn/sfn_debug.h"
13 #include "sfn/sfn_memorypool.h"
14 #include "sfn/sfn_nir.h"
15 #include "sfn/sfn_shader.h"
16 #include "r600_asm.h"
17 #include "r600_pipe.h"
18 #include "util/macros.h"
19 #include "util/ralloc.h"
20 
21 #include <cassert>
22 #include <cstdio>
23 #include <cstring>
24 #include <iostream>
25 
26 char *
r600_finalize_nir(pipe_screen * screen,struct nir_shader * nir)27 r600_finalize_nir(pipe_screen *screen, struct nir_shader *nir)
28 {
29    auto rs = container_of(screen, r600_screen, b.b);
30    r600_finalize_nir_common(nir, rs->b.gfx_level);
31    return nullptr;
32 }
33 
34 class MallocPoolRelease {
35 public:
MallocPoolRelease()36    MallocPoolRelease() { r600::init_pool(); }
~MallocPoolRelease()37    ~MallocPoolRelease() { r600::release_pool(); }
38 };
39 
40 int
r600_shader_from_nir(struct r600_context * rctx,struct r600_pipe_shader * pipeshader,r600_shader_key * key)41 r600_shader_from_nir(struct r600_context *rctx,
42                      struct r600_pipe_shader *pipeshader,
43                      r600_shader_key *key)
44 {
45 
46    MallocPoolRelease pool_release;
47 
48    struct r600_pipe_shader_selector *sel = pipeshader->selector;
49 
50    if (rctx->screen->b.debug_flags & DBG_PREOPT_IR) {
51       fprintf(stderr, "PRE-OPT-NIR-----------.------------------------------\n");
52       nir_print_shader(sel->nir, stderr);
53       fprintf(stderr, "END PRE-OPT-NIR--------------------------------------\n\n");
54    }
55 
56    auto sh = nir_shader_clone(sel->nir, sel->nir);
57 
58    r600_lower_and_optimize_nir(sh, key, rctx->b.gfx_level, &sel->so);
59 
60    if (rctx->screen->b.debug_flags & DBG_ALL_SHADERS) {
61       fprintf(stderr,
62               "-- NIR --------------------------------------------------------\n");
63       struct nir_function *func =
64          (struct nir_function *)exec_list_get_head(&sh->functions);
65       nir_index_ssa_defs(func->impl);
66       nir_print_shader(sh, stderr);
67       fprintf(stderr,
68               "-- END --------------------------------------------------------\n");
69    }
70 
71    memset(&pipeshader->shader, 0, sizeof(r600_shader));
72    pipeshader->scratch_space_needed = sh->scratch_size;
73 
74    if (sh->info.stage == MESA_SHADER_TESS_EVAL || sh->info.stage == MESA_SHADER_VERTEX ||
75        sh->info.stage == MESA_SHADER_GEOMETRY) {
76       pipeshader->shader.clip_dist_write |=
77          ((1 << sh->info.clip_distance_array_size) - 1);
78       pipeshader->shader.cull_dist_write = ((1 << sh->info.cull_distance_array_size) - 1)
79                                            << sh->info.clip_distance_array_size;
80       pipeshader->shader.cc_dist_mask =
81          (1 << (sh->info.cull_distance_array_size + sh->info.clip_distance_array_size)) -
82          1;
83    }
84    struct r600_shader *gs_shader = nullptr;
85    if (rctx->gs_shader)
86       gs_shader = &rctx->gs_shader->current->shader;
87    r600_screen *rscreen = rctx->screen;
88 
89    r600::Shader *shader =
90       r600::Shader::translate_from_nir(sh, &sel->so, gs_shader, *key,
91                                        rctx->isa->hw_class, rscreen->b.family);
92 
93    assert(shader);
94    if (!shader)
95       return -2;
96 
97    pipeshader->enabled_stream_buffers_mask = shader->enabled_stream_buffers_mask();
98    pipeshader->selector->info.file_count[TGSI_FILE_HW_ATOMIC] +=
99       shader->atomic_file_count();
100    pipeshader->selector->info.writes_memory =
101       shader->has_flag(r600::Shader::sh_writes_memory);
102 
103    r600_finalize_and_optimize_shader(shader);
104 
105    auto scheduled_shader = r600_schedule_shader(shader);
106    if (!scheduled_shader) {
107       return -1;
108    }
109 
110    scheduled_shader->get_shader_info(&pipeshader->shader);
111    pipeshader->shader.uses_doubles = sh->info.bit_sizes_float & 64 ? 1 : 0;
112 
113    r600_bytecode_init(&pipeshader->shader.bc,
114                       rscreen->b.gfx_level,
115                       rscreen->b.family,
116                       rscreen->has_compressed_msaa_texturing);
117 
118    /* We already schedule the code with this in mind, no need to handle this
119     * in the backend assembler */
120    pipeshader->shader.bc.ar_handling = AR_HANDLE_NORMAL;
121    pipeshader->shader.bc.r6xx_nop_after_rel_dst = 0;
122 
123    r600::sfn_log << r600::SfnLog::shader_info << "pipeshader->shader.processor_type = "
124                  << pipeshader->shader.processor_type << "\n";
125 
126    pipeshader->shader.bc.type = pipeshader->shader.processor_type;
127    pipeshader->shader.bc.isa = rctx->isa;
128    pipeshader->shader.bc.ngpr = scheduled_shader->required_registers();
129 
130    r600::Assembler afs(&pipeshader->shader, *key);
131    if (!afs.lower(scheduled_shader)) {
132       R600_ERR("%s: Lowering to assembly failed\n", __func__);
133 
134       scheduled_shader->print(std::cerr);
135       /* For now crash if the shader could not be generated */
136       assert(0);
137       return -1;
138    }
139 
140    if (sh->info.stage == MESA_SHADER_VERTEX) {
141       pipeshader->shader.vs_position_window_space =
142             sh->info.vs.window_space_position;
143    }
144 
145    if (sh->info.stage == MESA_SHADER_FRAGMENT)
146       pipeshader->shader.ps_conservative_z =
147             sh->info.fs.depth_layout;
148 
149    if (sh->info.stage == MESA_SHADER_GEOMETRY) {
150       r600::sfn_log << r600::SfnLog::shader_info
151                     << "Geometry shader, create copy shader\n";
152       generate_gs_copy_shader(rctx, pipeshader, &sel->so);
153       assert(pipeshader->gs_copy_shader);
154    } else {
155       r600::sfn_log << r600::SfnLog::shader_info << "This is not a Geometry shader\n";
156    }
157    ralloc_free(sh);
158 
159    return 0;
160 }
161