f220846aa2cfe738c945a0cbe825d06bb9c05140
[mesa.git] / src / mesa / drivers / dri / i965 / brw_cs.c
1 /*
2 * Copyright (c) 2014 - 2015 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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24 #include "util/ralloc.h"
25 #include "brw_context.h"
26 #include "brw_cs.h"
27 #include "brw_eu.h"
28 #include "brw_wm.h"
29 #include "brw_shader.h"
30 #include "intel_mipmap_tree.h"
31 #include "brw_state.h"
32 #include "intel_batchbuffer.h"
33 #include "brw_nir.h"
34 #include "brw_program.h"
35 #include "compiler/glsl/ir_uniform.h"
36
37 static void
38 assign_cs_binding_table_offsets(const struct gen_device_info *devinfo,
39 const struct gl_shader_program *shader_prog,
40 const struct gl_program *prog,
41 struct brw_cs_prog_data *prog_data)
42 {
43 uint32_t next_binding_table_offset = 0;
44
45 /* May not be used if the gl_NumWorkGroups variable is not accessed. */
46 prog_data->binding_table.work_groups_start = next_binding_table_offset;
47 next_binding_table_offset++;
48
49 brw_assign_common_binding_table_offsets(MESA_SHADER_COMPUTE, devinfo,
50 shader_prog, prog, &prog_data->base,
51 next_binding_table_offset);
52 }
53
54 static bool
55 brw_codegen_cs_prog(struct brw_context *brw,
56 struct gl_shader_program *prog,
57 struct brw_program *cp,
58 struct brw_cs_prog_key *key)
59 {
60 const struct gen_device_info *devinfo = &brw->screen->devinfo;
61 struct gl_context *ctx = &brw->ctx;
62 const GLuint *program;
63 void *mem_ctx = ralloc_context(NULL);
64 GLuint program_size;
65 struct brw_cs_prog_data prog_data;
66 bool start_busy = false;
67 double start_time = 0;
68
69 memset(&prog_data, 0, sizeof(prog_data));
70
71 if (cp->program.info.cs.shared_size > 64 * 1024) {
72 cp->program.sh.data->LinkStatus = false;
73 const char *error_str =
74 "Compute shader used more than 64KB of shared variables";
75 ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
76 _mesa_problem(NULL, "Failed to link compute shader: %s\n", error_str);
77
78 ralloc_free(mem_ctx);
79 return false;
80 } else {
81 prog_data.base.total_shared = cp->program.info.cs.shared_size;
82 }
83
84 assign_cs_binding_table_offsets(devinfo, prog, &cp->program, &prog_data);
85
86 /* Allocate the references to the uniforms that will end up in the
87 * prog_data associated with the compiled program, and which will be freed
88 * by the state cache.
89 */
90 int param_count = cp->program.nir->num_uniforms / 4;
91
92 /* The backend also sometimes add a param for the thread local id. */
93 prog_data.thread_local_id_index = param_count++;
94
95 /* The backend also sometimes adds params for texture size. */
96 param_count += 2 * ctx->Const.Program[MESA_SHADER_COMPUTE].MaxTextureImageUnits;
97 prog_data.base.param =
98 rzalloc_array(NULL, const gl_constant_value *, param_count);
99 prog_data.base.pull_param =
100 rzalloc_array(NULL, const gl_constant_value *, param_count);
101 prog_data.base.image_param =
102 rzalloc_array(NULL, struct brw_image_param,
103 cp->program.info.num_images);
104 prog_data.base.nr_params = param_count;
105 prog_data.base.nr_image_params = cp->program.info.num_images;
106
107 brw_nir_setup_glsl_uniforms(cp->program.nir, prog, &cp->program,
108 &prog_data.base, true);
109
110 if (unlikely(brw->perf_debug)) {
111 start_busy = (brw->batch.last_bo &&
112 drm_intel_bo_busy(brw->batch.last_bo));
113 start_time = get_time();
114 }
115
116 int st_index = -1;
117 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
118 st_index = brw_get_shader_time_index(brw, &cp->program, ST_CS, true);
119
120 char *error_str;
121 program = brw_compile_cs(brw->screen->compiler, brw, mem_ctx, key,
122 &prog_data, cp->program.nir, st_index,
123 &program_size, &error_str);
124 if (program == NULL) {
125 cp->program.sh.data->LinkStatus = false;
126 ralloc_strcat(&cp->program.sh.data->InfoLog, error_str);
127 _mesa_problem(NULL, "Failed to compile compute shader: %s\n", error_str);
128
129 ralloc_free(mem_ctx);
130 return false;
131 }
132
133 if (unlikely(brw->perf_debug)) {
134 if (cp->compiled_once) {
135 _mesa_problem(&brw->ctx, "CS programs shouldn't need recompiles");
136 }
137 cp->compiled_once = true;
138
139 if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
140 perf_debug("CS compile took %.03f ms and stalled the GPU\n",
141 (get_time() - start_time) * 1000);
142 }
143 }
144
145 const unsigned subslices = MAX2(brw->screen->subslice_total, 1);
146
147 /* WaCSScratchSize:hsw
148 *
149 * Haswell's scratch space address calculation appears to be sparse
150 * rather than tightly packed. The Thread ID has bits indicating
151 * which subslice, EU within a subslice, and thread within an EU
152 * it is. There's a maximum of two slices and two subslices, so these
153 * can be stored with a single bit. Even though there are only 10 EUs
154 * per subslice, this is stored in 4 bits, so there's an effective
155 * maximum value of 16 EUs. Similarly, although there are only 7
156 * threads per EU, this is stored in a 3 bit number, giving an effective
157 * maximum value of 8 threads per EU.
158 *
159 * This means that we need to use 16 * 8 instead of 10 * 7 for the
160 * number of threads per subslice.
161 */
162 const unsigned scratch_ids_per_subslice =
163 brw->is_haswell ? 16 * 8 : devinfo->max_cs_threads;
164
165 brw_alloc_stage_scratch(brw, &brw->cs.base,
166 prog_data.base.total_scratch,
167 scratch_ids_per_subslice * subslices);
168
169 brw_upload_cache(&brw->cache, BRW_CACHE_CS_PROG,
170 key, sizeof(*key),
171 program, program_size,
172 &prog_data, sizeof(prog_data),
173 &brw->cs.base.prog_offset, &brw->cs.base.prog_data);
174 ralloc_free(mem_ctx);
175
176 return true;
177 }
178
179
180 static void
181 brw_cs_populate_key(struct brw_context *brw, struct brw_cs_prog_key *key)
182 {
183 struct gl_context *ctx = &brw->ctx;
184 /* BRW_NEW_COMPUTE_PROGRAM */
185 const struct brw_program *cp = (struct brw_program *) brw->compute_program;
186 const struct gl_program *prog = (struct gl_program *) cp;
187
188 memset(key, 0, sizeof(*key));
189
190 /* _NEW_TEXTURE */
191 brw_populate_sampler_prog_key_data(ctx, prog, &key->tex);
192
193 /* The unique compute program ID */
194 key->program_string_id = cp->id;
195 }
196
197
198 void
199 brw_upload_cs_prog(struct brw_context *brw)
200 {
201 struct gl_context *ctx = &brw->ctx;
202 struct brw_cs_prog_key key;
203 struct brw_program *cp = (struct brw_program *) brw->compute_program;
204
205 if (!cp)
206 return;
207
208 if (!brw_state_dirty(brw, _NEW_TEXTURE, BRW_NEW_COMPUTE_PROGRAM))
209 return;
210
211 brw->cs.base.sampler_count =
212 util_last_bit(ctx->ComputeProgram._Current->SamplersUsed);
213
214 brw_cs_populate_key(brw, &key);
215
216 if (!brw_search_cache(&brw->cache, BRW_CACHE_CS_PROG,
217 &key, sizeof(key),
218 &brw->cs.base.prog_offset,
219 &brw->cs.base.prog_data)) {
220 bool success =
221 brw_codegen_cs_prog(brw,
222 ctx->_Shader->CurrentProgram[MESA_SHADER_COMPUTE],
223 cp, &key);
224 (void) success;
225 assert(success);
226 }
227 }
228
229
230 bool
231 brw_cs_precompile(struct gl_context *ctx,
232 struct gl_shader_program *shader_prog,
233 struct gl_program *prog)
234 {
235 struct brw_context *brw = brw_context(ctx);
236 struct brw_cs_prog_key key;
237
238 struct brw_program *bcp = brw_program(prog);
239
240 memset(&key, 0, sizeof(key));
241 key.program_string_id = bcp->id;
242
243 brw_setup_tex_for_precompile(brw, &key.tex, prog);
244
245 uint32_t old_prog_offset = brw->cs.base.prog_offset;
246 struct brw_stage_prog_data *old_prog_data = brw->cs.base.prog_data;
247
248 bool success = brw_codegen_cs_prog(brw, shader_prog, bcp, &key);
249
250 brw->cs.base.prog_offset = old_prog_offset;
251 brw->cs.base.prog_data = old_prog_data;
252
253 return success;
254 }