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