llvmpipe: pass number of samplers into llvm sampler code.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_state_cs.c
1 /**************************************************************************
2 *
3 * Copyright 2019 Red Hat.
4 * All Rights Reserved.
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 shall be included
14 * in all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
17 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 **************************************************************************/
25 #include "util/u_memory.h"
26 #include "util/simple_list.h"
27 #include "util/os_time.h"
28 #include "util/u_dump.h"
29 #include "util/u_string.h"
30 #include "tgsi/tgsi_dump.h"
31 #include "tgsi/tgsi_parse.h"
32 #include "gallivm/lp_bld_const.h"
33 #include "gallivm/lp_bld_debug.h"
34 #include "gallivm/lp_bld_intr.h"
35 #include "gallivm/lp_bld_flow.h"
36 #include "gallivm/lp_bld_gather.h"
37 #include "gallivm/lp_bld_coro.h"
38 #include "gallivm/lp_bld_nir.h"
39 #include "lp_state_cs.h"
40 #include "lp_context.h"
41 #include "lp_debug.h"
42 #include "lp_state.h"
43 #include "lp_perf.h"
44 #include "lp_screen.h"
45 #include "lp_memory.h"
46 #include "lp_cs_tpool.h"
47 #include "frontend/sw_winsys.h"
48 #include "nir/nir_to_tgsi_info.h"
49 #include "util/mesa-sha1.h"
50 #include "nir_serialize.h"
51
52 /** Fragment shader number (for debugging) */
53 static unsigned cs_no = 0;
54
55 struct lp_cs_job_info {
56 unsigned grid_size[3];
57 unsigned block_size[3];
58 unsigned req_local_mem;
59 unsigned work_dim;
60 struct lp_cs_exec *current;
61 };
62
63 static void
64 generate_compute(struct llvmpipe_context *lp,
65 struct lp_compute_shader *shader,
66 struct lp_compute_shader_variant *variant)
67 {
68 struct gallivm_state *gallivm = variant->gallivm;
69 const struct lp_compute_shader_variant_key *key = &variant->key;
70 char func_name[64], func_name_coro[64];
71 LLVMTypeRef arg_types[17];
72 LLVMTypeRef func_type, coro_func_type;
73 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
74 LLVMValueRef context_ptr;
75 LLVMValueRef x_size_arg, y_size_arg, z_size_arg;
76 LLVMValueRef grid_x_arg, grid_y_arg, grid_z_arg;
77 LLVMValueRef grid_size_x_arg, grid_size_y_arg, grid_size_z_arg;
78 LLVMValueRef work_dim_arg, thread_data_ptr;
79 LLVMBasicBlockRef block;
80 LLVMBuilderRef builder;
81 struct lp_build_sampler_soa *sampler;
82 struct lp_build_image_soa *image;
83 LLVMValueRef function, coro;
84 struct lp_type cs_type;
85 unsigned i;
86
87 /*
88 * This function has two parts
89 * a) setup the coroutine execution environment loop.
90 * b) build the compute shader llvm for use inside the coroutine.
91 */
92 assert(lp_native_vector_width / 32 >= 4);
93
94 memset(&cs_type, 0, sizeof cs_type);
95 cs_type.floating = TRUE; /* floating point values */
96 cs_type.sign = TRUE; /* values are signed */
97 cs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
98 cs_type.width = 32; /* 32-bit float */
99 cs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
100 snprintf(func_name, sizeof(func_name), "cs_variant");
101
102 snprintf(func_name_coro, sizeof(func_name), "cs_co_variant");
103
104 arg_types[0] = variant->jit_cs_context_ptr_type; /* context */
105 arg_types[1] = int32_type; /* block_x_size */
106 arg_types[2] = int32_type; /* block_y_size */
107 arg_types[3] = int32_type; /* block_z_size */
108 arg_types[4] = int32_type; /* grid_x */
109 arg_types[5] = int32_type; /* grid_y */
110 arg_types[6] = int32_type; /* grid_z */
111 arg_types[7] = int32_type; /* grid_size_x */
112 arg_types[8] = int32_type; /* grid_size_y */
113 arg_types[9] = int32_type; /* grid_size_z */
114 arg_types[10] = int32_type; /* work dim */
115 arg_types[11] = variant->jit_cs_thread_data_ptr_type; /* per thread data */
116 arg_types[12] = int32_type; /* coro only - num X loops */
117 arg_types[13] = int32_type; /* coro only - partials */
118 arg_types[14] = int32_type; /* coro block_x_size */
119 arg_types[15] = int32_type; /* coro block_y_size */
120 arg_types[16] = int32_type; /* coro block_z_size */
121 func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
122 arg_types, ARRAY_SIZE(arg_types) - 5, 0);
123
124 coro_func_type = LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0),
125 arg_types, ARRAY_SIZE(arg_types), 0);
126
127 function = LLVMAddFunction(gallivm->module, func_name, func_type);
128 LLVMSetFunctionCallConv(function, LLVMCCallConv);
129
130 coro = LLVMAddFunction(gallivm->module, func_name_coro, coro_func_type);
131 LLVMSetFunctionCallConv(coro, LLVMCCallConv);
132
133 variant->function = function;
134
135 for(i = 0; i < ARRAY_SIZE(arg_types); ++i) {
136 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
137 lp_add_function_attr(coro, i + 1, LP_FUNC_ATTR_NOALIAS);
138 lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
139 }
140 }
141
142 lp_build_coro_declare_malloc_hooks(gallivm);
143
144 if (variant->gallivm->cache->data_size)
145 return;
146
147 context_ptr = LLVMGetParam(function, 0);
148 x_size_arg = LLVMGetParam(function, 1);
149 y_size_arg = LLVMGetParam(function, 2);
150 z_size_arg = LLVMGetParam(function, 3);
151 grid_x_arg = LLVMGetParam(function, 4);
152 grid_y_arg = LLVMGetParam(function, 5);
153 grid_z_arg = LLVMGetParam(function, 6);
154 grid_size_x_arg = LLVMGetParam(function, 7);
155 grid_size_y_arg = LLVMGetParam(function, 8);
156 grid_size_z_arg = LLVMGetParam(function, 9);
157 work_dim_arg = LLVMGetParam(function, 10);
158 thread_data_ptr = LLVMGetParam(function, 11);
159
160 lp_build_name(context_ptr, "context");
161 lp_build_name(x_size_arg, "x_size");
162 lp_build_name(y_size_arg, "y_size");
163 lp_build_name(z_size_arg, "z_size");
164 lp_build_name(grid_x_arg, "grid_x");
165 lp_build_name(grid_y_arg, "grid_y");
166 lp_build_name(grid_z_arg, "grid_z");
167 lp_build_name(grid_size_x_arg, "grid_size_x");
168 lp_build_name(grid_size_y_arg, "grid_size_y");
169 lp_build_name(grid_size_z_arg, "grid_size_z");
170 lp_build_name(work_dim_arg, "work_dim");
171 lp_build_name(thread_data_ptr, "thread_data");
172
173 block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
174 builder = gallivm->builder;
175 assert(builder);
176 LLVMPositionBuilderAtEnd(builder, block);
177 sampler = lp_llvm_sampler_soa_create(key->samplers, key->nr_samplers);
178 image = lp_llvm_image_soa_create(lp_cs_variant_key_images(key));
179
180 struct lp_build_loop_state loop_state[4];
181 LLVMValueRef num_x_loop;
182 LLVMValueRef vec_length = lp_build_const_int32(gallivm, cs_type.length);
183 num_x_loop = LLVMBuildAdd(gallivm->builder, x_size_arg, vec_length, "");
184 num_x_loop = LLVMBuildSub(gallivm->builder, num_x_loop, lp_build_const_int32(gallivm, 1), "");
185 num_x_loop = LLVMBuildUDiv(gallivm->builder, num_x_loop, vec_length, "");
186 LLVMValueRef partials = LLVMBuildURem(gallivm->builder, x_size_arg, vec_length, "");
187
188 LLVMValueRef coro_num_hdls = LLVMBuildMul(gallivm->builder, num_x_loop, y_size_arg, "");
189 coro_num_hdls = LLVMBuildMul(gallivm->builder, coro_num_hdls, z_size_arg, "");
190
191 LLVMTypeRef hdl_ptr_type = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
192 LLVMValueRef coro_hdls = LLVMBuildArrayAlloca(gallivm->builder, hdl_ptr_type, coro_num_hdls, "coro_hdls");
193
194 unsigned end_coroutine = INT_MAX;
195
196 /*
197 * This is the main coroutine execution loop. It iterates over the dimensions
198 * and calls the coroutine main entrypoint on the first pass, but in subsequent
199 * passes it checks if the coroutine has completed and resumes it if not.
200 */
201 /* take x_width - round up to type.length width */
202 lp_build_loop_begin(&loop_state[3], gallivm,
203 lp_build_const_int32(gallivm, 0)); /* coroutine reentry loop */
204 lp_build_loop_begin(&loop_state[2], gallivm,
205 lp_build_const_int32(gallivm, 0)); /* z loop */
206 lp_build_loop_begin(&loop_state[1], gallivm,
207 lp_build_const_int32(gallivm, 0)); /* y loop */
208 lp_build_loop_begin(&loop_state[0], gallivm,
209 lp_build_const_int32(gallivm, 0)); /* x loop */
210 {
211 LLVMValueRef args[17];
212 args[0] = context_ptr;
213 args[1] = loop_state[0].counter;
214 args[2] = loop_state[1].counter;
215 args[3] = loop_state[2].counter;
216 args[4] = grid_x_arg;
217 args[5] = grid_y_arg;
218 args[6] = grid_z_arg;
219 args[7] = grid_size_x_arg;
220 args[8] = grid_size_y_arg;
221 args[9] = grid_size_z_arg;
222 args[10] = work_dim_arg;
223 args[11] = thread_data_ptr;
224 args[12] = num_x_loop;
225 args[13] = partials;
226 args[14] = x_size_arg;
227 args[15] = y_size_arg;
228 args[16] = z_size_arg;
229
230 /* idx = (z * (size_x * size_y) + y * size_x + x */
231 LLVMValueRef coro_hdl_idx = LLVMBuildMul(gallivm->builder, loop_state[2].counter,
232 LLVMBuildMul(gallivm->builder, num_x_loop, y_size_arg, ""), "");
233 coro_hdl_idx = LLVMBuildAdd(gallivm->builder, coro_hdl_idx,
234 LLVMBuildMul(gallivm->builder, loop_state[1].counter,
235 num_x_loop, ""), "");
236 coro_hdl_idx = LLVMBuildAdd(gallivm->builder, coro_hdl_idx,
237 loop_state[0].counter, "");
238
239 LLVMValueRef coro_entry = LLVMBuildGEP(gallivm->builder, coro_hdls, &coro_hdl_idx, 1, "");
240
241 LLVMValueRef coro_hdl = LLVMBuildLoad(gallivm->builder, coro_entry, "coro_hdl");
242
243 struct lp_build_if_state ifstate;
244 LLVMValueRef cmp = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, loop_state[3].counter,
245 lp_build_const_int32(gallivm, 0), "");
246 /* first time here - call the coroutine function entry point */
247 lp_build_if(&ifstate, gallivm, cmp);
248 LLVMValueRef coro_ret = LLVMBuildCall(gallivm->builder, coro, args, 17, "");
249 LLVMBuildStore(gallivm->builder, coro_ret, coro_entry);
250 lp_build_else(&ifstate);
251 /* subsequent calls for this invocation - check if done. */
252 LLVMValueRef coro_done = lp_build_coro_done(gallivm, coro_hdl);
253 struct lp_build_if_state ifstate2;
254 lp_build_if(&ifstate2, gallivm, coro_done);
255 /* if done destroy and force loop exit */
256 lp_build_coro_destroy(gallivm, coro_hdl);
257 lp_build_loop_force_set_counter(&loop_state[3], lp_build_const_int32(gallivm, end_coroutine - 1));
258 lp_build_else(&ifstate2);
259 /* otherwise resume the coroutine */
260 lp_build_coro_resume(gallivm, coro_hdl);
261 lp_build_endif(&ifstate2);
262 lp_build_endif(&ifstate);
263 lp_build_loop_force_reload_counter(&loop_state[3]);
264 }
265 lp_build_loop_end_cond(&loop_state[0],
266 num_x_loop,
267 NULL, LLVMIntUGE);
268 lp_build_loop_end_cond(&loop_state[1],
269 y_size_arg,
270 NULL, LLVMIntUGE);
271 lp_build_loop_end_cond(&loop_state[2],
272 z_size_arg,
273 NULL, LLVMIntUGE);
274 lp_build_loop_end_cond(&loop_state[3],
275 lp_build_const_int32(gallivm, end_coroutine),
276 NULL, LLVMIntEQ);
277 LLVMBuildRetVoid(builder);
278
279 /* This is stage (b) - generate the compute shader code inside the coroutine. */
280 LLVMValueRef block_x_size_arg, block_y_size_arg, block_z_size_arg;
281 context_ptr = LLVMGetParam(coro, 0);
282 x_size_arg = LLVMGetParam(coro, 1);
283 y_size_arg = LLVMGetParam(coro, 2);
284 z_size_arg = LLVMGetParam(coro, 3);
285 grid_x_arg = LLVMGetParam(coro, 4);
286 grid_y_arg = LLVMGetParam(coro, 5);
287 grid_z_arg = LLVMGetParam(coro, 6);
288 grid_size_x_arg = LLVMGetParam(coro, 7);
289 grid_size_y_arg = LLVMGetParam(coro, 8);
290 grid_size_z_arg = LLVMGetParam(coro, 9);
291 work_dim_arg = LLVMGetParam(coro, 10);
292 thread_data_ptr = LLVMGetParam(coro, 11);
293 num_x_loop = LLVMGetParam(coro, 12);
294 partials = LLVMGetParam(coro, 13);
295 block_x_size_arg = LLVMGetParam(coro, 14);
296 block_y_size_arg = LLVMGetParam(coro, 15);
297 block_z_size_arg = LLVMGetParam(coro, 16);
298 block = LLVMAppendBasicBlockInContext(gallivm->context, coro, "entry");
299 LLVMPositionBuilderAtEnd(builder, block);
300 {
301 LLVMValueRef consts_ptr, num_consts_ptr;
302 LLVMValueRef ssbo_ptr, num_ssbo_ptr;
303 LLVMValueRef shared_ptr;
304 LLVMValueRef kernel_args_ptr;
305 struct lp_build_mask_context mask;
306 struct lp_bld_tgsi_system_values system_values;
307
308 memset(&system_values, 0, sizeof(system_values));
309 consts_ptr = lp_jit_cs_context_constants(gallivm, context_ptr);
310 num_consts_ptr = lp_jit_cs_context_num_constants(gallivm, context_ptr);
311 ssbo_ptr = lp_jit_cs_context_ssbos(gallivm, context_ptr);
312 num_ssbo_ptr = lp_jit_cs_context_num_ssbos(gallivm, context_ptr);
313 kernel_args_ptr = lp_jit_cs_context_kernel_args(gallivm, context_ptr);
314
315 shared_ptr = lp_jit_cs_thread_data_shared(gallivm, thread_data_ptr);
316
317 /* these are coroutine entrypoint necessities */
318 LLVMValueRef coro_id = lp_build_coro_id(gallivm);
319 LLVMValueRef coro_hdl = lp_build_coro_begin_alloc_mem(gallivm, coro_id);
320
321 LLVMValueRef has_partials = LLVMBuildICmp(gallivm->builder, LLVMIntNE, partials, lp_build_const_int32(gallivm, 0), "");
322 LLVMValueRef tid_vals[3];
323 LLVMValueRef tids_x[LP_MAX_VECTOR_LENGTH], tids_y[LP_MAX_VECTOR_LENGTH], tids_z[LP_MAX_VECTOR_LENGTH];
324 LLVMValueRef base_val = LLVMBuildMul(gallivm->builder, x_size_arg, vec_length, "");
325 for (i = 0; i < cs_type.length; i++) {
326 tids_x[i] = LLVMBuildAdd(gallivm->builder, base_val, lp_build_const_int32(gallivm, i), "");
327 tids_y[i] = y_size_arg;
328 tids_z[i] = z_size_arg;
329 }
330 tid_vals[0] = lp_build_gather_values(gallivm, tids_x, cs_type.length);
331 tid_vals[1] = lp_build_gather_values(gallivm, tids_y, cs_type.length);
332 tid_vals[2] = lp_build_gather_values(gallivm, tids_z, cs_type.length);
333 system_values.thread_id = LLVMGetUndef(LLVMArrayType(LLVMVectorType(int32_type, cs_type.length), 3));
334 for (i = 0; i < 3; i++)
335 system_values.thread_id = LLVMBuildInsertValue(builder, system_values.thread_id, tid_vals[i], i, "");
336
337 LLVMValueRef gtids[3] = { grid_x_arg, grid_y_arg, grid_z_arg };
338 system_values.block_id = LLVMGetUndef(LLVMVectorType(int32_type, 3));
339 for (i = 0; i < 3; i++)
340 system_values.block_id = LLVMBuildInsertElement(builder, system_values.block_id, gtids[i], lp_build_const_int32(gallivm, i), "");
341
342 LLVMValueRef gstids[3] = { grid_size_x_arg, grid_size_y_arg, grid_size_z_arg };
343 system_values.grid_size = LLVMGetUndef(LLVMVectorType(int32_type, 3));
344 for (i = 0; i < 3; i++)
345 system_values.grid_size = LLVMBuildInsertElement(builder, system_values.grid_size, gstids[i], lp_build_const_int32(gallivm, i), "");
346
347 system_values.work_dim = work_dim_arg;
348
349 LLVMValueRef bsize[3] = { block_x_size_arg, block_y_size_arg, block_z_size_arg };
350 system_values.block_size = LLVMGetUndef(LLVMVectorType(int32_type, 3));
351 for (i = 0; i < 3; i++)
352 system_values.block_size = LLVMBuildInsertElement(builder, system_values.block_size, bsize[i], lp_build_const_int32(gallivm, i), "");
353
354 LLVMValueRef last_x_loop = LLVMBuildICmp(gallivm->builder, LLVMIntEQ, x_size_arg, LLVMBuildSub(gallivm->builder, num_x_loop, lp_build_const_int32(gallivm, 1), ""), "");
355 LLVMValueRef use_partial_mask = LLVMBuildAnd(gallivm->builder, last_x_loop, has_partials, "");
356 struct lp_build_if_state if_state;
357 LLVMValueRef mask_val = lp_build_alloca(gallivm, LLVMVectorType(int32_type, cs_type.length), "mask");
358 LLVMValueRef full_mask_val = lp_build_const_int_vec(gallivm, cs_type, ~0);
359 LLVMBuildStore(gallivm->builder, full_mask_val, mask_val);
360
361 lp_build_if(&if_state, gallivm, use_partial_mask);
362 struct lp_build_loop_state mask_loop_state;
363 lp_build_loop_begin(&mask_loop_state, gallivm, partials);
364 LLVMValueRef tmask_val = LLVMBuildLoad(gallivm->builder, mask_val, "");
365 tmask_val = LLVMBuildInsertElement(gallivm->builder, tmask_val, lp_build_const_int32(gallivm, 0), mask_loop_state.counter, "");
366 LLVMBuildStore(gallivm->builder, tmask_val, mask_val);
367 lp_build_loop_end_cond(&mask_loop_state, vec_length, NULL, LLVMIntUGE);
368 lp_build_endif(&if_state);
369
370 mask_val = LLVMBuildLoad(gallivm->builder, mask_val, "");
371 lp_build_mask_begin(&mask, gallivm, cs_type, mask_val);
372
373 struct lp_build_coro_suspend_info coro_info;
374
375 LLVMBasicBlockRef sus_block = LLVMAppendBasicBlockInContext(gallivm->context, coro, "suspend");
376 LLVMBasicBlockRef clean_block = LLVMAppendBasicBlockInContext(gallivm->context, coro, "cleanup");
377
378 coro_info.suspend = sus_block;
379 coro_info.cleanup = clean_block;
380
381 struct lp_build_tgsi_params params;
382 memset(&params, 0, sizeof(params));
383
384 params.type = cs_type;
385 params.mask = &mask;
386 params.consts_ptr = consts_ptr;
387 params.const_sizes_ptr = num_consts_ptr;
388 params.system_values = &system_values;
389 params.context_ptr = context_ptr;
390 params.sampler = sampler;
391 params.info = &shader->info.base;
392 params.ssbo_ptr = ssbo_ptr;
393 params.ssbo_sizes_ptr = num_ssbo_ptr;
394 params.image = image;
395 params.shared_ptr = shared_ptr;
396 params.coro = &coro_info;
397 params.kernel_args = kernel_args_ptr;
398
399 if (shader->base.type == PIPE_SHADER_IR_TGSI)
400 lp_build_tgsi_soa(gallivm, shader->base.tokens, &params, NULL);
401 else
402 lp_build_nir_soa(gallivm, shader->base.ir.nir, &params,
403 NULL);
404
405 mask_val = lp_build_mask_end(&mask);
406
407 lp_build_coro_suspend_switch(gallivm, &coro_info, NULL, true);
408 LLVMPositionBuilderAtEnd(builder, clean_block);
409
410 lp_build_coro_free_mem(gallivm, coro_id, coro_hdl);
411
412 LLVMBuildBr(builder, sus_block);
413 LLVMPositionBuilderAtEnd(builder, sus_block);
414
415 lp_build_coro_end(gallivm, coro_hdl);
416 LLVMBuildRet(builder, coro_hdl);
417 }
418
419 sampler->destroy(sampler);
420 image->destroy(image);
421
422 gallivm_verify_function(gallivm, coro);
423 gallivm_verify_function(gallivm, function);
424 }
425
426 static void *
427 llvmpipe_create_compute_state(struct pipe_context *pipe,
428 const struct pipe_compute_state *templ)
429 {
430 struct lp_compute_shader *shader;
431 int nr_samplers, nr_sampler_views;
432
433 shader = CALLOC_STRUCT(lp_compute_shader);
434 if (!shader)
435 return NULL;
436
437 shader->no = cs_no++;
438
439 shader->base.type = templ->ir_type;
440 if (templ->ir_type == PIPE_SHADER_IR_NIR_SERIALIZED) {
441 struct blob_reader reader;
442 const struct pipe_binary_program_header *hdr = templ->prog;
443
444 blob_reader_init(&reader, hdr->blob, hdr->num_bytes);
445 shader->base.ir.nir = nir_deserialize(NULL, pipe->screen->get_compiler_options(pipe->screen, PIPE_SHADER_IR_NIR, PIPE_SHADER_COMPUTE), &reader);
446 shader->base.type = PIPE_SHADER_IR_NIR;
447
448 pipe->screen->finalize_nir(pipe->screen, shader->base.ir.nir, false);
449 } else if (templ->ir_type == PIPE_SHADER_IR_NIR)
450 shader->base.ir.nir = (struct nir_shader *)templ->prog;
451
452 if (shader->base.type == PIPE_SHADER_IR_TGSI) {
453 /* get/save the summary info for this shader */
454 lp_build_tgsi_info(templ->prog, &shader->info);
455
456 /* we need to keep a local copy of the tokens */
457 shader->base.tokens = tgsi_dup_tokens(templ->prog);
458 } else {
459 nir_tgsi_scan_shader(shader->base.ir.nir, &shader->info.base, false);
460 }
461
462 shader->req_local_mem = templ->req_local_mem;
463 make_empty_list(&shader->variants);
464
465 nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
466 nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
467 int nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
468 shader->variant_key_size = lp_cs_variant_key_size(MAX2(nr_samplers, nr_sampler_views), nr_images);
469
470 return shader;
471 }
472
473 static void
474 llvmpipe_bind_compute_state(struct pipe_context *pipe,
475 void *cs)
476 {
477 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
478
479 if (llvmpipe->cs == cs)
480 return;
481
482 llvmpipe->cs = (struct lp_compute_shader *)cs;
483 llvmpipe->cs_dirty |= LP_CSNEW_CS;
484 }
485
486 /**
487 * Remove shader variant from two lists: the shader's variant list
488 * and the context's variant list.
489 */
490 static void
491 llvmpipe_remove_cs_shader_variant(struct llvmpipe_context *lp,
492 struct lp_compute_shader_variant *variant)
493 {
494 if ((LP_DEBUG & DEBUG_CS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
495 debug_printf("llvmpipe: del cs #%u var %u v created %u v cached %u "
496 "v total cached %u inst %u total inst %u\n",
497 variant->shader->no, variant->no,
498 variant->shader->variants_created,
499 variant->shader->variants_cached,
500 lp->nr_cs_variants, variant->nr_instrs, lp->nr_cs_instrs);
501 }
502
503 gallivm_destroy(variant->gallivm);
504
505 /* remove from shader's list */
506 remove_from_list(&variant->list_item_local);
507 variant->shader->variants_cached--;
508
509 /* remove from context's list */
510 remove_from_list(&variant->list_item_global);
511 lp->nr_fs_variants--;
512 lp->nr_fs_instrs -= variant->nr_instrs;
513
514 FREE(variant);
515 }
516
517 static void
518 llvmpipe_delete_compute_state(struct pipe_context *pipe,
519 void *cs)
520 {
521 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
522 struct lp_compute_shader *shader = cs;
523 struct lp_cs_variant_list_item *li;
524
525 if (llvmpipe->cs == cs)
526 llvmpipe->cs = NULL;
527 for (unsigned i = 0; i < shader->max_global_buffers; i++)
528 pipe_resource_reference(&shader->global_buffers[i], NULL);
529 FREE(shader->global_buffers);
530
531 /* Delete all the variants */
532 li = first_elem(&shader->variants);
533 while(!at_end(&shader->variants, li)) {
534 struct lp_cs_variant_list_item *next = next_elem(li);
535 llvmpipe_remove_cs_shader_variant(llvmpipe, li->base);
536 li = next;
537 }
538 if (shader->base.ir.nir)
539 ralloc_free(shader->base.ir.nir);
540 tgsi_free_tokens(shader->base.tokens);
541 FREE(shader);
542 }
543
544 static struct lp_compute_shader_variant_key *
545 make_variant_key(struct llvmpipe_context *lp,
546 struct lp_compute_shader *shader,
547 char *store)
548 {
549 int i;
550 struct lp_compute_shader_variant_key *key;
551 key = (struct lp_compute_shader_variant_key *)store;
552 memset(key, 0, offsetof(struct lp_compute_shader_variant_key, samplers[1]));
553
554 /* This value will be the same for all the variants of a given shader:
555 */
556 key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
557
558 struct lp_sampler_static_state *cs_sampler;
559
560 cs_sampler = key->samplers;
561 for(i = 0; i < key->nr_samplers; ++i) {
562 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
563 lp_sampler_static_sampler_state(&cs_sampler[i].sampler_state,
564 lp->samplers[PIPE_SHADER_COMPUTE][i]);
565 }
566 }
567
568 /*
569 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
570 * are dx10-style? Can't really have mixed opcodes, at least not
571 * if we want to skip the holes here (without rescanning tgsi).
572 */
573 if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
574 key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
575 for(i = 0; i < key->nr_sampler_views; ++i) {
576 /*
577 * Note sview may exceed what's representable by file_mask.
578 * This will still work, the only downside is that not actually
579 * used views may be included in the shader key.
580 */
581 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
582 lp_sampler_static_texture_state(&cs_sampler[i].texture_state,
583 lp->sampler_views[PIPE_SHADER_COMPUTE][i]);
584 }
585 }
586 }
587 else {
588 key->nr_sampler_views = key->nr_samplers;
589 for(i = 0; i < key->nr_sampler_views; ++i) {
590 if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
591 lp_sampler_static_texture_state(&cs_sampler[i].texture_state,
592 lp->sampler_views[PIPE_SHADER_COMPUTE][i]);
593 }
594 }
595 }
596
597 struct lp_image_static_state *lp_image;
598 lp_image = lp_cs_variant_key_images(key);
599 key->nr_images = shader->info.base.file_max[TGSI_FILE_IMAGE] + 1;
600 for (i = 0; i < key->nr_images; ++i) {
601 if (shader->info.base.file_mask[TGSI_FILE_IMAGE] & (1 << i)) {
602 lp_sampler_static_texture_state_image(&lp_image[i].image_state,
603 &lp->images[PIPE_SHADER_COMPUTE][i]);
604 }
605 }
606 return key;
607 }
608
609 static void
610 dump_cs_variant_key(const struct lp_compute_shader_variant_key *key)
611 {
612 int i;
613 debug_printf("cs variant %p:\n", (void *) key);
614
615 for (i = 0; i < key->nr_samplers; ++i) {
616 const struct lp_static_sampler_state *sampler = &key->samplers[i].sampler_state;
617 debug_printf("sampler[%u] = \n", i);
618 debug_printf(" .wrap = %s %s %s\n",
619 util_str_tex_wrap(sampler->wrap_s, TRUE),
620 util_str_tex_wrap(sampler->wrap_t, TRUE),
621 util_str_tex_wrap(sampler->wrap_r, TRUE));
622 debug_printf(" .min_img_filter = %s\n",
623 util_str_tex_filter(sampler->min_img_filter, TRUE));
624 debug_printf(" .min_mip_filter = %s\n",
625 util_str_tex_mipfilter(sampler->min_mip_filter, TRUE));
626 debug_printf(" .mag_img_filter = %s\n",
627 util_str_tex_filter(sampler->mag_img_filter, TRUE));
628 if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
629 debug_printf(" .compare_func = %s\n", util_str_func(sampler->compare_func, TRUE));
630 debug_printf(" .normalized_coords = %u\n", sampler->normalized_coords);
631 debug_printf(" .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
632 debug_printf(" .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
633 debug_printf(" .apply_min_lod = %u\n", sampler->apply_min_lod);
634 debug_printf(" .apply_max_lod = %u\n", sampler->apply_max_lod);
635 }
636 for (i = 0; i < key->nr_sampler_views; ++i) {
637 const struct lp_static_texture_state *texture = &key->samplers[i].texture_state;
638 debug_printf("texture[%u] = \n", i);
639 debug_printf(" .format = %s\n",
640 util_format_name(texture->format));
641 debug_printf(" .target = %s\n",
642 util_str_tex_target(texture->target, TRUE));
643 debug_printf(" .level_zero_only = %u\n",
644 texture->level_zero_only);
645 debug_printf(" .pot = %u %u %u\n",
646 texture->pot_width,
647 texture->pot_height,
648 texture->pot_depth);
649 }
650 struct lp_image_static_state *images = lp_cs_variant_key_images(key);
651 for (i = 0; i < key->nr_images; ++i) {
652 const struct lp_static_texture_state *image = &images[i].image_state;
653 debug_printf("image[%u] = \n", i);
654 debug_printf(" .format = %s\n",
655 util_format_name(image->format));
656 debug_printf(" .target = %s\n",
657 util_str_tex_target(image->target, TRUE));
658 debug_printf(" .level_zero_only = %u\n",
659 image->level_zero_only);
660 debug_printf(" .pot = %u %u %u\n",
661 image->pot_width,
662 image->pot_height,
663 image->pot_depth);
664 }
665 }
666
667 static void
668 lp_debug_cs_variant(const struct lp_compute_shader_variant *variant)
669 {
670 debug_printf("llvmpipe: Compute shader #%u variant #%u:\n",
671 variant->shader->no, variant->no);
672 if (variant->shader->base.type == PIPE_SHADER_IR_TGSI)
673 tgsi_dump(variant->shader->base.tokens, 0);
674 else
675 nir_print_shader(variant->shader->base.ir.nir, stderr);
676 dump_cs_variant_key(&variant->key);
677 debug_printf("\n");
678 }
679
680 static void
681 lp_cs_get_ir_cache_key(struct lp_compute_shader_variant *variant,
682 unsigned char ir_sha1_cache_key[20])
683 {
684 struct blob blob = { 0 };
685 unsigned ir_size;
686 void *ir_binary;
687
688 blob_init(&blob);
689 nir_serialize(&blob, variant->shader->base.ir.nir, true);
690 ir_binary = blob.data;
691 ir_size = blob.size;
692
693 struct mesa_sha1 ctx;
694 _mesa_sha1_init(&ctx);
695 _mesa_sha1_update(&ctx, &variant->key, variant->shader->variant_key_size);
696 _mesa_sha1_update(&ctx, ir_binary, ir_size);
697 _mesa_sha1_final(&ctx, ir_sha1_cache_key);
698
699 blob_finish(&blob);
700 }
701
702 static struct lp_compute_shader_variant *
703 generate_variant(struct llvmpipe_context *lp,
704 struct lp_compute_shader *shader,
705 const struct lp_compute_shader_variant_key *key)
706 {
707 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
708 struct lp_compute_shader_variant *variant;
709 char module_name[64];
710 unsigned char ir_sha1_cache_key[20];
711 struct lp_cached_code cached = { 0 };
712 bool needs_caching = false;
713 variant = MALLOC(sizeof *variant + shader->variant_key_size - sizeof variant->key);
714 if (!variant)
715 return NULL;
716
717 memset(variant, 0, sizeof(*variant));
718 snprintf(module_name, sizeof(module_name), "cs%u_variant%u",
719 shader->no, shader->variants_created);
720
721 variant->shader = shader;
722 memcpy(&variant->key, key, shader->variant_key_size);
723
724 if (shader->base.ir.nir) {
725 lp_cs_get_ir_cache_key(variant, ir_sha1_cache_key);
726
727 lp_disk_cache_find_shader(screen, &cached, ir_sha1_cache_key);
728 if (!cached.data_size)
729 needs_caching = true;
730 }
731 variant->gallivm = gallivm_create(module_name, lp->context, &cached);
732 if (!variant->gallivm) {
733 FREE(variant);
734 return NULL;
735 }
736
737 variant->list_item_global.base = variant;
738 variant->list_item_local.base = variant;
739 variant->no = shader->variants_created++;
740
741
742
743 if ((LP_DEBUG & DEBUG_CS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
744 lp_debug_cs_variant(variant);
745 }
746
747 lp_jit_init_cs_types(variant);
748
749 generate_compute(lp, shader, variant);
750
751 gallivm_compile_module(variant->gallivm);
752
753 lp_build_coro_add_malloc_hooks(variant->gallivm);
754 variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
755
756 variant->jit_function = (lp_jit_cs_func)gallivm_jit_function(variant->gallivm, variant->function);
757
758 if (needs_caching) {
759 lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);
760 }
761 gallivm_free_ir(variant->gallivm);
762 return variant;
763 }
764
765 static void
766 lp_cs_ctx_set_cs_variant( struct lp_cs_context *csctx,
767 struct lp_compute_shader_variant *variant)
768 {
769 csctx->cs.current.variant = variant;
770 }
771
772 static void
773 llvmpipe_update_cs(struct llvmpipe_context *lp)
774 {
775 struct lp_compute_shader *shader = lp->cs;
776
777 struct lp_compute_shader_variant_key *key;
778 struct lp_compute_shader_variant *variant = NULL;
779 struct lp_cs_variant_list_item *li;
780 char store[LP_CS_MAX_VARIANT_KEY_SIZE];
781
782 key = make_variant_key(lp, shader, store);
783
784 /* Search the variants for one which matches the key */
785 li = first_elem(&shader->variants);
786 while(!at_end(&shader->variants, li)) {
787 if(memcmp(&li->base->key, key, shader->variant_key_size) == 0) {
788 variant = li->base;
789 break;
790 }
791 li = next_elem(li);
792 }
793
794 if (variant) {
795 /* Move this variant to the head of the list to implement LRU
796 * deletion of shader's when we have too many.
797 */
798 move_to_head(&lp->cs_variants_list, &variant->list_item_global);
799 }
800 else {
801 /* variant not found, create it now */
802 int64_t t0, t1, dt;
803 unsigned i;
804 unsigned variants_to_cull;
805
806 if (LP_DEBUG & DEBUG_CS) {
807 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
808 lp->nr_cs_variants,
809 lp->nr_cs_instrs,
810 lp->nr_cs_variants ? lp->nr_cs_instrs / lp->nr_cs_variants : 0);
811 }
812
813 /* First, check if we've exceeded the max number of shader variants.
814 * If so, free 6.25% of them (the least recently used ones).
815 */
816 variants_to_cull = lp->nr_cs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 16 : 0;
817
818 if (variants_to_cull ||
819 lp->nr_cs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
820 if (gallivm_debug & GALLIVM_DEBUG_PERF) {
821 debug_printf("Evicting CS: %u cs variants,\t%u total variants,"
822 "\t%u instrs,\t%u instrs/variant\n",
823 shader->variants_cached,
824 lp->nr_cs_variants, lp->nr_cs_instrs,
825 lp->nr_cs_instrs / lp->nr_cs_variants);
826 }
827
828 /*
829 * We need to re-check lp->nr_cs_variants because an arbitrarliy large
830 * number of shader variants (potentially all of them) could be
831 * pending for destruction on flush.
832 */
833
834 for (i = 0; i < variants_to_cull || lp->nr_cs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
835 struct lp_cs_variant_list_item *item;
836 if (is_empty_list(&lp->cs_variants_list)) {
837 break;
838 }
839 item = last_elem(&lp->cs_variants_list);
840 assert(item);
841 assert(item->base);
842 llvmpipe_remove_cs_shader_variant(lp, item->base);
843 }
844 }
845 /*
846 * Generate the new variant.
847 */
848 t0 = os_time_get();
849 variant = generate_variant(lp, shader, key);
850 t1 = os_time_get();
851 dt = t1 - t0;
852 LP_COUNT_ADD(llvm_compile_time, dt);
853 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
854
855 /* Put the new variant into the list */
856 if (variant) {
857 insert_at_head(&shader->variants, &variant->list_item_local);
858 insert_at_head(&lp->cs_variants_list, &variant->list_item_global);
859 lp->nr_cs_variants++;
860 lp->nr_cs_instrs += variant->nr_instrs;
861 shader->variants_cached++;
862 }
863 }
864 /* Bind this variant */
865 lp_cs_ctx_set_cs_variant(lp->csctx, variant);
866 }
867
868 /**
869 * Called during state validation when LP_CSNEW_SAMPLER_VIEW is set.
870 */
871 static void
872 lp_csctx_set_sampler_views(struct lp_cs_context *csctx,
873 unsigned num,
874 struct pipe_sampler_view **views)
875 {
876 unsigned i, max_tex_num;
877
878 LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
879
880 assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
881
882 max_tex_num = MAX2(num, csctx->cs.current_tex_num);
883
884 for (i = 0; i < max_tex_num; i++) {
885 struct pipe_sampler_view *view = i < num ? views[i] : NULL;
886
887 if (view) {
888 struct pipe_resource *res = view->texture;
889 struct llvmpipe_resource *lp_tex = llvmpipe_resource(res);
890 struct lp_jit_texture *jit_tex;
891 jit_tex = &csctx->cs.current.jit_context.textures[i];
892
893 /* We're referencing the texture's internal data, so save a
894 * reference to it.
895 */
896 pipe_resource_reference(&csctx->cs.current_tex[i], res);
897
898 if (!lp_tex->dt) {
899 /* regular texture - csctx array of mipmap level offsets */
900 int j;
901 unsigned first_level = 0;
902 unsigned last_level = 0;
903
904 if (llvmpipe_resource_is_texture(res)) {
905 first_level = view->u.tex.first_level;
906 last_level = view->u.tex.last_level;
907 assert(first_level <= last_level);
908 assert(last_level <= res->last_level);
909 jit_tex->base = lp_tex->tex_data;
910 }
911 else {
912 jit_tex->base = lp_tex->data;
913 }
914 if (LP_PERF & PERF_TEX_MEM) {
915 /* use dummy tile memory */
916 jit_tex->base = lp_dummy_tile;
917 jit_tex->width = TILE_SIZE/8;
918 jit_tex->height = TILE_SIZE/8;
919 jit_tex->depth = 1;
920 jit_tex->first_level = 0;
921 jit_tex->last_level = 0;
922 jit_tex->mip_offsets[0] = 0;
923 jit_tex->row_stride[0] = 0;
924 jit_tex->img_stride[0] = 0;
925 jit_tex->num_samples = 0;
926 jit_tex->sample_stride = 0;
927 }
928 else {
929 jit_tex->width = res->width0;
930 jit_tex->height = res->height0;
931 jit_tex->depth = res->depth0;
932 jit_tex->first_level = first_level;
933 jit_tex->last_level = last_level;
934 jit_tex->num_samples = res->nr_samples;
935 jit_tex->sample_stride = 0;
936
937 if (llvmpipe_resource_is_texture(res)) {
938 for (j = first_level; j <= last_level; j++) {
939 jit_tex->mip_offsets[j] = lp_tex->mip_offsets[j];
940 jit_tex->row_stride[j] = lp_tex->row_stride[j];
941 jit_tex->img_stride[j] = lp_tex->img_stride[j];
942 }
943 jit_tex->sample_stride = lp_tex->sample_stride;
944
945 if (res->target == PIPE_TEXTURE_1D_ARRAY ||
946 res->target == PIPE_TEXTURE_2D_ARRAY ||
947 res->target == PIPE_TEXTURE_CUBE ||
948 res->target == PIPE_TEXTURE_CUBE_ARRAY) {
949 /*
950 * For array textures, we don't have first_layer, instead
951 * adjust last_layer (stored as depth) plus the mip level offsets
952 * (as we have mip-first layout can't just adjust base ptr).
953 * XXX For mip levels, could do something similar.
954 */
955 jit_tex->depth = view->u.tex.last_layer - view->u.tex.first_layer + 1;
956 for (j = first_level; j <= last_level; j++) {
957 jit_tex->mip_offsets[j] += view->u.tex.first_layer *
958 lp_tex->img_stride[j];
959 }
960 if (view->target == PIPE_TEXTURE_CUBE ||
961 view->target == PIPE_TEXTURE_CUBE_ARRAY) {
962 assert(jit_tex->depth % 6 == 0);
963 }
964 assert(view->u.tex.first_layer <= view->u.tex.last_layer);
965 assert(view->u.tex.last_layer < res->array_size);
966 }
967 }
968 else {
969 /*
970 * For buffers, we don't have "offset", instead adjust
971 * the size (stored as width) plus the base pointer.
972 */
973 unsigned view_blocksize = util_format_get_blocksize(view->format);
974 /* probably don't really need to fill that out */
975 jit_tex->mip_offsets[0] = 0;
976 jit_tex->row_stride[0] = 0;
977 jit_tex->img_stride[0] = 0;
978
979 /* everything specified in number of elements here. */
980 jit_tex->width = view->u.buf.size / view_blocksize;
981 jit_tex->base = (uint8_t *)jit_tex->base + view->u.buf.offset;
982 /* XXX Unsure if we need to sanitize parameters? */
983 assert(view->u.buf.offset + view->u.buf.size <= res->width0);
984 }
985 }
986 }
987 else {
988 /* display target texture/surface */
989 /*
990 * XXX: Where should this be unmapped?
991 */
992 struct llvmpipe_screen *screen = llvmpipe_screen(res->screen);
993 struct sw_winsys *winsys = screen->winsys;
994 jit_tex->base = winsys->displaytarget_map(winsys, lp_tex->dt,
995 PIPE_TRANSFER_READ);
996 jit_tex->row_stride[0] = lp_tex->row_stride[0];
997 jit_tex->img_stride[0] = lp_tex->img_stride[0];
998 jit_tex->mip_offsets[0] = 0;
999 jit_tex->width = res->width0;
1000 jit_tex->height = res->height0;
1001 jit_tex->depth = res->depth0;
1002 jit_tex->first_level = jit_tex->last_level = 0;
1003 jit_tex->num_samples = res->nr_samples;
1004 jit_tex->sample_stride = 0;
1005 assert(jit_tex->base);
1006 }
1007 }
1008 else {
1009 pipe_resource_reference(&csctx->cs.current_tex[i], NULL);
1010 }
1011 }
1012 csctx->cs.current_tex_num = num;
1013 }
1014
1015
1016 /**
1017 * Called during state validation when LP_NEW_SAMPLER is set.
1018 */
1019 static void
1020 lp_csctx_set_sampler_state(struct lp_cs_context *csctx,
1021 unsigned num,
1022 struct pipe_sampler_state **samplers)
1023 {
1024 unsigned i;
1025
1026 LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
1027
1028 assert(num <= PIPE_MAX_SAMPLERS);
1029
1030 for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
1031 const struct pipe_sampler_state *sampler = i < num ? samplers[i] : NULL;
1032
1033 if (sampler) {
1034 struct lp_jit_sampler *jit_sam;
1035 jit_sam = &csctx->cs.current.jit_context.samplers[i];
1036
1037 jit_sam->min_lod = sampler->min_lod;
1038 jit_sam->max_lod = sampler->max_lod;
1039 jit_sam->lod_bias = sampler->lod_bias;
1040 COPY_4V(jit_sam->border_color, sampler->border_color.f);
1041 }
1042 }
1043 }
1044
1045 static void
1046 lp_csctx_set_cs_constants(struct lp_cs_context *csctx,
1047 unsigned num,
1048 struct pipe_constant_buffer *buffers)
1049 {
1050 unsigned i;
1051
1052 LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) buffers);
1053
1054 assert(num <= ARRAY_SIZE(csctx->constants));
1055
1056 for (i = 0; i < num; ++i) {
1057 util_copy_constant_buffer(&csctx->constants[i].current, &buffers[i]);
1058 }
1059 for (; i < ARRAY_SIZE(csctx->constants); i++) {
1060 util_copy_constant_buffer(&csctx->constants[i].current, NULL);
1061 }
1062 }
1063
1064 static void
1065 lp_csctx_set_cs_ssbos(struct lp_cs_context *csctx,
1066 unsigned num,
1067 struct pipe_shader_buffer *buffers)
1068 {
1069 int i;
1070 LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *)buffers);
1071
1072 assert (num <= ARRAY_SIZE(csctx->ssbos));
1073
1074 for (i = 0; i < num; ++i) {
1075 util_copy_shader_buffer(&csctx->ssbos[i].current, &buffers[i]);
1076 }
1077 for (; i < ARRAY_SIZE(csctx->ssbos); i++) {
1078 util_copy_shader_buffer(&csctx->ssbos[i].current, NULL);
1079 }
1080 }
1081
1082 static void
1083 lp_csctx_set_cs_images(struct lp_cs_context *csctx,
1084 unsigned num,
1085 struct pipe_image_view *images)
1086 {
1087 unsigned i;
1088
1089 LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) images);
1090
1091 assert(num <= ARRAY_SIZE(csctx->images));
1092
1093 for (i = 0; i < num; ++i) {
1094 struct pipe_image_view *image = &images[i];
1095 util_copy_image_view(&csctx->images[i].current, &images[i]);
1096
1097 struct pipe_resource *res = image->resource;
1098 struct llvmpipe_resource *lp_res = llvmpipe_resource(res);
1099 struct lp_jit_image *jit_image;
1100
1101 jit_image = &csctx->cs.current.jit_context.images[i];
1102 if (!lp_res)
1103 continue;
1104 if (!lp_res->dt) {
1105 /* regular texture - csctx array of mipmap level offsets */
1106 if (llvmpipe_resource_is_texture(res)) {
1107 jit_image->base = lp_res->tex_data;
1108 } else
1109 jit_image->base = lp_res->data;
1110
1111 jit_image->width = res->width0;
1112 jit_image->height = res->height0;
1113 jit_image->depth = res->depth0;
1114 jit_image->num_samples = res->nr_samples;
1115
1116 if (llvmpipe_resource_is_texture(res)) {
1117 uint32_t mip_offset = lp_res->mip_offsets[image->u.tex.level];
1118
1119 jit_image->width = u_minify(jit_image->width, image->u.tex.level);
1120 jit_image->height = u_minify(jit_image->height, image->u.tex.level);
1121
1122 if (res->target == PIPE_TEXTURE_1D_ARRAY ||
1123 res->target == PIPE_TEXTURE_2D_ARRAY ||
1124 res->target == PIPE_TEXTURE_3D ||
1125 res->target == PIPE_TEXTURE_CUBE ||
1126 res->target == PIPE_TEXTURE_CUBE_ARRAY) {
1127 /*
1128 * For array textures, we don't have first_layer, instead
1129 * adjust last_layer (stored as depth) plus the mip level offsets
1130 * (as we have mip-first layout can't just adjust base ptr).
1131 * XXX For mip levels, could do something similar.
1132 */
1133 jit_image->depth = image->u.tex.last_layer - image->u.tex.first_layer + 1;
1134 mip_offset += image->u.tex.first_layer * lp_res->img_stride[image->u.tex.level];
1135 } else
1136 jit_image->depth = u_minify(jit_image->depth, image->u.tex.level);
1137
1138 jit_image->row_stride = lp_res->row_stride[image->u.tex.level];
1139 jit_image->img_stride = lp_res->img_stride[image->u.tex.level];
1140 jit_image->sample_stride = lp_res->sample_stride;
1141 jit_image->base = (uint8_t *)jit_image->base + mip_offset;
1142 } else {
1143 unsigned view_blocksize = util_format_get_blocksize(image->format);
1144 jit_image->width = image->u.buf.size / view_blocksize;
1145 jit_image->base = (uint8_t *)jit_image->base + image->u.buf.offset;
1146 }
1147 }
1148 }
1149 for (; i < ARRAY_SIZE(csctx->images); i++) {
1150 util_copy_image_view(&csctx->images[i].current, NULL);
1151 }
1152 }
1153
1154 static void
1155 update_csctx_consts(struct llvmpipe_context *llvmpipe)
1156 {
1157 struct lp_cs_context *csctx = llvmpipe->csctx;
1158 int i;
1159
1160 for (i = 0; i < ARRAY_SIZE(csctx->constants); ++i) {
1161 struct pipe_resource *buffer = csctx->constants[i].current.buffer;
1162 const ubyte *current_data = NULL;
1163
1164 if (buffer) {
1165 /* resource buffer */
1166 current_data = (ubyte *) llvmpipe_resource_data(buffer);
1167 }
1168 else if (csctx->constants[i].current.user_buffer) {
1169 /* user-space buffer */
1170 current_data = (ubyte *) csctx->constants[i].current.user_buffer;
1171 }
1172
1173 if (current_data) {
1174 current_data += csctx->constants[i].current.buffer_offset;
1175
1176 csctx->cs.current.jit_context.constants[i] = (const float *)current_data;
1177 csctx->cs.current.jit_context.num_constants[i] = csctx->constants[i].current.buffer_size;
1178 } else {
1179 csctx->cs.current.jit_context.constants[i] = NULL;
1180 csctx->cs.current.jit_context.num_constants[i] = 0;
1181 }
1182 }
1183 }
1184
1185 static void
1186 update_csctx_ssbo(struct llvmpipe_context *llvmpipe)
1187 {
1188 struct lp_cs_context *csctx = llvmpipe->csctx;
1189 int i;
1190 for (i = 0; i < ARRAY_SIZE(csctx->ssbos); ++i) {
1191 struct pipe_resource *buffer = csctx->ssbos[i].current.buffer;
1192 const ubyte *current_data = NULL;
1193
1194 if (!buffer)
1195 continue;
1196 /* resource buffer */
1197 current_data = (ubyte *) llvmpipe_resource_data(buffer);
1198 if (current_data) {
1199 current_data += csctx->ssbos[i].current.buffer_offset;
1200
1201 csctx->cs.current.jit_context.ssbos[i] = (const uint32_t *)current_data;
1202 csctx->cs.current.jit_context.num_ssbos[i] = csctx->ssbos[i].current.buffer_size;
1203 } else {
1204 csctx->cs.current.jit_context.ssbos[i] = NULL;
1205 csctx->cs.current.jit_context.num_ssbos[i] = 0;
1206 }
1207 }
1208 }
1209
1210 static void
1211 llvmpipe_cs_update_derived(struct llvmpipe_context *llvmpipe, void *input)
1212 {
1213 if (llvmpipe->cs_dirty & (LP_CSNEW_CS))
1214 llvmpipe_update_cs(llvmpipe);
1215
1216 if (llvmpipe->cs_dirty & LP_CSNEW_CONSTANTS) {
1217 lp_csctx_set_cs_constants(llvmpipe->csctx,
1218 ARRAY_SIZE(llvmpipe->constants[PIPE_SHADER_COMPUTE]),
1219 llvmpipe->constants[PIPE_SHADER_COMPUTE]);
1220 update_csctx_consts(llvmpipe);
1221 }
1222
1223 if (llvmpipe->cs_dirty & LP_CSNEW_SSBOS) {
1224 lp_csctx_set_cs_ssbos(llvmpipe->csctx,
1225 ARRAY_SIZE(llvmpipe->ssbos[PIPE_SHADER_COMPUTE]),
1226 llvmpipe->ssbos[PIPE_SHADER_COMPUTE]);
1227 update_csctx_ssbo(llvmpipe);
1228 }
1229
1230 if (llvmpipe->cs_dirty & LP_CSNEW_SAMPLER_VIEW)
1231 lp_csctx_set_sampler_views(llvmpipe->csctx,
1232 llvmpipe->num_sampler_views[PIPE_SHADER_COMPUTE],
1233 llvmpipe->sampler_views[PIPE_SHADER_COMPUTE]);
1234
1235 if (llvmpipe->cs_dirty & LP_CSNEW_SAMPLER)
1236 lp_csctx_set_sampler_state(llvmpipe->csctx,
1237 llvmpipe->num_samplers[PIPE_SHADER_COMPUTE],
1238 llvmpipe->samplers[PIPE_SHADER_COMPUTE]);
1239
1240 if (llvmpipe->cs_dirty & LP_CSNEW_IMAGES)
1241 lp_csctx_set_cs_images(llvmpipe->csctx,
1242 ARRAY_SIZE(llvmpipe->images[PIPE_SHADER_COMPUTE]),
1243 llvmpipe->images[PIPE_SHADER_COMPUTE]);
1244
1245 if (input) {
1246 struct lp_cs_context *csctx = llvmpipe->csctx;
1247 csctx->input = input;
1248 csctx->cs.current.jit_context.kernel_args = input;
1249 }
1250
1251 llvmpipe->cs_dirty = 0;
1252 }
1253
1254 static void
1255 cs_exec_fn(void *init_data, int iter_idx, struct lp_cs_local_mem *lmem)
1256 {
1257 struct lp_cs_job_info *job_info = init_data;
1258 struct lp_jit_cs_thread_data thread_data;
1259
1260 memset(&thread_data, 0, sizeof(thread_data));
1261
1262 if (lmem->local_size < job_info->req_local_mem) {
1263 lmem->local_mem_ptr = REALLOC(lmem->local_mem_ptr, lmem->local_size,
1264 job_info->req_local_mem);
1265 lmem->local_size = job_info->req_local_mem;
1266 }
1267 thread_data.shared = lmem->local_mem_ptr;
1268
1269 unsigned grid_z = iter_idx / (job_info->grid_size[0] * job_info->grid_size[1]);
1270 unsigned grid_y = (iter_idx - (grid_z * (job_info->grid_size[0] * job_info->grid_size[1]))) / job_info->grid_size[0];
1271 unsigned grid_x = (iter_idx - (grid_z * (job_info->grid_size[0] * job_info->grid_size[1])) - (grid_y * job_info->grid_size[0]));
1272 struct lp_compute_shader_variant *variant = job_info->current->variant;
1273 variant->jit_function(&job_info->current->jit_context,
1274 job_info->block_size[0], job_info->block_size[1], job_info->block_size[2],
1275 grid_x, grid_y, grid_z,
1276 job_info->grid_size[0], job_info->grid_size[1], job_info->grid_size[2], job_info->work_dim,
1277 &thread_data);
1278 }
1279
1280 static void
1281 fill_grid_size(struct pipe_context *pipe,
1282 const struct pipe_grid_info *info,
1283 uint32_t grid_size[3])
1284 {
1285 struct pipe_transfer *transfer;
1286 uint32_t *params;
1287 if (!info->indirect) {
1288 grid_size[0] = info->grid[0];
1289 grid_size[1] = info->grid[1];
1290 grid_size[2] = info->grid[2];
1291 return;
1292 }
1293 params = pipe_buffer_map_range(pipe, info->indirect,
1294 info->indirect_offset,
1295 3 * sizeof(uint32_t),
1296 PIPE_TRANSFER_READ,
1297 &transfer);
1298
1299 if (!transfer)
1300 return;
1301
1302 grid_size[0] = params[0];
1303 grid_size[1] = params[1];
1304 grid_size[2] = params[2];
1305 pipe_buffer_unmap(pipe, transfer);
1306 }
1307
1308 static void llvmpipe_launch_grid(struct pipe_context *pipe,
1309 const struct pipe_grid_info *info)
1310 {
1311 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1312 struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
1313 struct lp_cs_job_info job_info;
1314
1315 memset(&job_info, 0, sizeof(job_info));
1316
1317 llvmpipe_cs_update_derived(llvmpipe, info->input);
1318
1319 fill_grid_size(pipe, info, job_info.grid_size);
1320
1321 job_info.block_size[0] = info->block[0];
1322 job_info.block_size[1] = info->block[1];
1323 job_info.block_size[2] = info->block[2];
1324 job_info.work_dim = info->work_dim;
1325 job_info.req_local_mem = llvmpipe->cs->req_local_mem;
1326 job_info.current = &llvmpipe->csctx->cs.current;
1327
1328 int num_tasks = job_info.grid_size[2] * job_info.grid_size[1] * job_info.grid_size[0];
1329 if (num_tasks) {
1330 struct lp_cs_tpool_task *task;
1331 mtx_lock(&screen->cs_mutex);
1332 task = lp_cs_tpool_queue_task(screen->cs_tpool, cs_exec_fn, &job_info, num_tasks);
1333
1334 lp_cs_tpool_wait_for_task(screen->cs_tpool, &task);
1335 mtx_unlock(&screen->cs_mutex);
1336 }
1337 llvmpipe->pipeline_statistics.cs_invocations += num_tasks * info->block[0] * info->block[1] * info->block[2];
1338 }
1339
1340 static void
1341 llvmpipe_set_compute_resources(struct pipe_context *pipe,
1342 unsigned start, unsigned count,
1343 struct pipe_surface **resources)
1344 {
1345
1346
1347 }
1348
1349 static void
1350 llvmpipe_set_global_binding(struct pipe_context *pipe,
1351 unsigned first, unsigned count,
1352 struct pipe_resource **resources,
1353 uint32_t **handles)
1354 {
1355 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1356 struct lp_compute_shader *cs = llvmpipe->cs;
1357 unsigned i;
1358
1359 if (first + count > cs->max_global_buffers) {
1360 unsigned old_max = cs->max_global_buffers;
1361 cs->max_global_buffers = first + count;
1362 cs->global_buffers = realloc(cs->global_buffers,
1363 cs->max_global_buffers * sizeof(cs->global_buffers[0]));
1364 if (!cs->global_buffers) {
1365 return;
1366 }
1367
1368 memset(&cs->global_buffers[old_max], 0, (cs->max_global_buffers - old_max) * sizeof(cs->global_buffers[0]));
1369 }
1370
1371 if (!resources) {
1372 for (i = 0; i < count; i++)
1373 pipe_resource_reference(&cs->global_buffers[first + i], NULL);
1374 return;
1375 }
1376
1377 for (i = 0; i < count; i++) {
1378 uintptr_t va;
1379 uint32_t offset;
1380 pipe_resource_reference(&cs->global_buffers[first + i], resources[i]);
1381 struct llvmpipe_resource *lp_res = llvmpipe_resource(resources[i]);
1382 offset = *handles[i];
1383 va = (uintptr_t)((char *)lp_res->data + offset);
1384 memcpy(handles[i], &va, sizeof(va));
1385 }
1386 }
1387
1388 void
1389 llvmpipe_init_compute_funcs(struct llvmpipe_context *llvmpipe)
1390 {
1391 llvmpipe->pipe.create_compute_state = llvmpipe_create_compute_state;
1392 llvmpipe->pipe.bind_compute_state = llvmpipe_bind_compute_state;
1393 llvmpipe->pipe.delete_compute_state = llvmpipe_delete_compute_state;
1394 llvmpipe->pipe.set_compute_resources = llvmpipe_set_compute_resources;
1395 llvmpipe->pipe.set_global_binding = llvmpipe_set_global_binding;
1396 llvmpipe->pipe.launch_grid = llvmpipe_launch_grid;
1397 }
1398
1399 void
1400 lp_csctx_destroy(struct lp_cs_context *csctx)
1401 {
1402 unsigned i;
1403 for (i = 0; i < ARRAY_SIZE(csctx->cs.current_tex); i++) {
1404 pipe_resource_reference(&csctx->cs.current_tex[i], NULL);
1405 }
1406 for (i = 0; i < ARRAY_SIZE(csctx->constants); i++) {
1407 pipe_resource_reference(&csctx->constants[i].current.buffer, NULL);
1408 }
1409 for (i = 0; i < ARRAY_SIZE(csctx->ssbos); i++) {
1410 pipe_resource_reference(&csctx->ssbos[i].current.buffer, NULL);
1411 }
1412 FREE(csctx);
1413 }
1414
1415 struct lp_cs_context *lp_csctx_create(struct pipe_context *pipe)
1416 {
1417 struct lp_cs_context *csctx;
1418
1419 csctx = CALLOC_STRUCT(lp_cs_context);
1420 if (!csctx)
1421 return NULL;
1422
1423 csctx->pipe = pipe;
1424 return csctx;
1425 }