1 /**************************************************************************
3 * Copyright 2019 Red Hat.
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:
13 * The above copyright notice and this permission notice shall be included
14 * in all copies or substantial portions of the Software.
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
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"
44 #include "lp_screen.h"
45 #include "lp_memory.h"
46 #include "lp_cs_tpool.h"
47 #include "state_tracker/sw_winsys.h"
48 #include "nir/nir_to_tgsi_info.h"
49 #include "nir_serialize.h"
50 struct lp_cs_job_info
{
51 unsigned grid_size
[3];
52 unsigned block_size
[3];
53 unsigned req_local_mem
;
55 struct lp_cs_exec
*current
;
59 generate_compute(struct llvmpipe_context
*lp
,
60 struct lp_compute_shader
*shader
,
61 struct lp_compute_shader_variant
*variant
)
63 struct gallivm_state
*gallivm
= variant
->gallivm
;
64 const struct lp_compute_shader_variant_key
*key
= &variant
->key
;
65 char func_name
[64], func_name_coro
[64];
66 LLVMTypeRef arg_types
[17];
67 LLVMTypeRef func_type
, coro_func_type
;
68 LLVMTypeRef int32_type
= LLVMInt32TypeInContext(gallivm
->context
);
69 LLVMValueRef context_ptr
;
70 LLVMValueRef x_size_arg
, y_size_arg
, z_size_arg
;
71 LLVMValueRef grid_x_arg
, grid_y_arg
, grid_z_arg
;
72 LLVMValueRef grid_size_x_arg
, grid_size_y_arg
, grid_size_z_arg
;
73 LLVMValueRef work_dim_arg
, thread_data_ptr
;
74 LLVMBasicBlockRef block
;
75 LLVMBuilderRef builder
;
76 struct lp_build_sampler_soa
*sampler
;
77 struct lp_build_image_soa
*image
;
78 LLVMValueRef function
, coro
;
79 struct lp_type cs_type
;
83 * This function has two parts
84 * a) setup the coroutine execution environment loop.
85 * b) build the compute shader llvm for use inside the coroutine.
87 assert(lp_native_vector_width
/ 32 >= 4);
89 memset(&cs_type
, 0, sizeof cs_type
);
90 cs_type
.floating
= TRUE
; /* floating point values */
91 cs_type
.sign
= TRUE
; /* values are signed */
92 cs_type
.norm
= FALSE
; /* values are not limited to [0,1] or [-1,1] */
93 cs_type
.width
= 32; /* 32-bit float */
94 cs_type
.length
= MIN2(lp_native_vector_width
/ 32, 16); /* n*4 elements per vector */
95 snprintf(func_name
, sizeof(func_name
), "cs%u_variant%u",
96 shader
->no
, variant
->no
);
98 snprintf(func_name_coro
, sizeof(func_name
), "cs_co_%u_variant%u",
99 shader
->no
, variant
->no
);
101 arg_types
[0] = variant
->jit_cs_context_ptr_type
; /* context */
102 arg_types
[1] = int32_type
; /* block_x_size */
103 arg_types
[2] = int32_type
; /* block_y_size */
104 arg_types
[3] = int32_type
; /* block_z_size */
105 arg_types
[4] = int32_type
; /* grid_x */
106 arg_types
[5] = int32_type
; /* grid_y */
107 arg_types
[6] = int32_type
; /* grid_z */
108 arg_types
[7] = int32_type
; /* grid_size_x */
109 arg_types
[8] = int32_type
; /* grid_size_y */
110 arg_types
[9] = int32_type
; /* grid_size_z */
111 arg_types
[10] = int32_type
; /* work dim */
112 arg_types
[11] = variant
->jit_cs_thread_data_ptr_type
; /* per thread data */
113 arg_types
[12] = int32_type
; /* coro only - num X loops */
114 arg_types
[13] = int32_type
; /* coro only - partials */
115 arg_types
[14] = int32_type
; /* coro block_x_size */
116 arg_types
[15] = int32_type
; /* coro block_y_size */
117 arg_types
[16] = int32_type
; /* coro block_z_size */
118 func_type
= LLVMFunctionType(LLVMVoidTypeInContext(gallivm
->context
),
119 arg_types
, ARRAY_SIZE(arg_types
) - 5, 0);
121 coro_func_type
= LLVMFunctionType(LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0),
122 arg_types
, ARRAY_SIZE(arg_types
), 0);
124 function
= LLVMAddFunction(gallivm
->module
, func_name
, func_type
);
125 LLVMSetFunctionCallConv(function
, LLVMCCallConv
);
127 coro
= LLVMAddFunction(gallivm
->module
, func_name_coro
, coro_func_type
);
128 LLVMSetFunctionCallConv(coro
, LLVMCCallConv
);
130 variant
->function
= function
;
132 for(i
= 0; i
< ARRAY_SIZE(arg_types
); ++i
) {
133 if(LLVMGetTypeKind(arg_types
[i
]) == LLVMPointerTypeKind
) {
134 lp_add_function_attr(coro
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
135 lp_add_function_attr(function
, i
+ 1, LP_FUNC_ATTR_NOALIAS
);
139 context_ptr
= LLVMGetParam(function
, 0);
140 x_size_arg
= LLVMGetParam(function
, 1);
141 y_size_arg
= LLVMGetParam(function
, 2);
142 z_size_arg
= LLVMGetParam(function
, 3);
143 grid_x_arg
= LLVMGetParam(function
, 4);
144 grid_y_arg
= LLVMGetParam(function
, 5);
145 grid_z_arg
= LLVMGetParam(function
, 6);
146 grid_size_x_arg
= LLVMGetParam(function
, 7);
147 grid_size_y_arg
= LLVMGetParam(function
, 8);
148 grid_size_z_arg
= LLVMGetParam(function
, 9);
149 work_dim_arg
= LLVMGetParam(function
, 10);
150 thread_data_ptr
= LLVMGetParam(function
, 11);
152 lp_build_name(context_ptr
, "context");
153 lp_build_name(x_size_arg
, "x_size");
154 lp_build_name(y_size_arg
, "y_size");
155 lp_build_name(z_size_arg
, "z_size");
156 lp_build_name(grid_x_arg
, "grid_x");
157 lp_build_name(grid_y_arg
, "grid_y");
158 lp_build_name(grid_z_arg
, "grid_z");
159 lp_build_name(grid_size_x_arg
, "grid_size_x");
160 lp_build_name(grid_size_y_arg
, "grid_size_y");
161 lp_build_name(grid_size_z_arg
, "grid_size_z");
162 lp_build_name(work_dim_arg
, "work_dim");
163 lp_build_name(thread_data_ptr
, "thread_data");
165 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, function
, "entry");
166 builder
= gallivm
->builder
;
168 LLVMPositionBuilderAtEnd(builder
, block
);
169 sampler
= lp_llvm_sampler_soa_create(key
->state
);
170 image
= lp_llvm_image_soa_create(key
->image_state
);
172 struct lp_build_loop_state loop_state
[4];
173 LLVMValueRef num_x_loop
;
174 LLVMValueRef vec_length
= lp_build_const_int32(gallivm
, cs_type
.length
);
175 num_x_loop
= LLVMBuildAdd(gallivm
->builder
, x_size_arg
, vec_length
, "");
176 num_x_loop
= LLVMBuildSub(gallivm
->builder
, num_x_loop
, lp_build_const_int32(gallivm
, 1), "");
177 num_x_loop
= LLVMBuildUDiv(gallivm
->builder
, num_x_loop
, vec_length
, "");
178 LLVMValueRef partials
= LLVMBuildURem(gallivm
->builder
, x_size_arg
, vec_length
, "");
180 LLVMValueRef coro_num_hdls
= LLVMBuildMul(gallivm
->builder
, num_x_loop
, y_size_arg
, "");
181 coro_num_hdls
= LLVMBuildMul(gallivm
->builder
, coro_num_hdls
, z_size_arg
, "");
183 LLVMTypeRef hdl_ptr_type
= LLVMPointerType(LLVMInt8TypeInContext(gallivm
->context
), 0);
184 LLVMValueRef coro_hdls
= LLVMBuildArrayAlloca(gallivm
->builder
, hdl_ptr_type
, coro_num_hdls
, "coro_hdls");
186 unsigned end_coroutine
= INT_MAX
;
189 * This is the main coroutine execution loop. It iterates over the dimensions
190 * and calls the coroutine main entrypoint on the first pass, but in subsequent
191 * passes it checks if the coroutine has completed and resumes it if not.
193 /* take x_width - round up to type.length width */
194 lp_build_loop_begin(&loop_state
[3], gallivm
,
195 lp_build_const_int32(gallivm
, 0)); /* coroutine reentry loop */
196 lp_build_loop_begin(&loop_state
[2], gallivm
,
197 lp_build_const_int32(gallivm
, 0)); /* z loop */
198 lp_build_loop_begin(&loop_state
[1], gallivm
,
199 lp_build_const_int32(gallivm
, 0)); /* y loop */
200 lp_build_loop_begin(&loop_state
[0], gallivm
,
201 lp_build_const_int32(gallivm
, 0)); /* x loop */
203 LLVMValueRef args
[17];
204 args
[0] = context_ptr
;
205 args
[1] = loop_state
[0].counter
;
206 args
[2] = loop_state
[1].counter
;
207 args
[3] = loop_state
[2].counter
;
208 args
[4] = grid_x_arg
;
209 args
[5] = grid_y_arg
;
210 args
[6] = grid_z_arg
;
211 args
[7] = grid_size_x_arg
;
212 args
[8] = grid_size_y_arg
;
213 args
[9] = grid_size_z_arg
;
214 args
[10] = work_dim_arg
;
215 args
[11] = thread_data_ptr
;
216 args
[12] = num_x_loop
;
218 args
[14] = x_size_arg
;
219 args
[15] = y_size_arg
;
220 args
[16] = z_size_arg
;
222 /* idx = (z * (size_x * size_y) + y * size_x + x */
223 LLVMValueRef coro_hdl_idx
= LLVMBuildMul(gallivm
->builder
, loop_state
[2].counter
,
224 LLVMBuildMul(gallivm
->builder
, num_x_loop
, y_size_arg
, ""), "");
225 coro_hdl_idx
= LLVMBuildAdd(gallivm
->builder
, coro_hdl_idx
,
226 LLVMBuildMul(gallivm
->builder
, loop_state
[1].counter
,
227 num_x_loop
, ""), "");
228 coro_hdl_idx
= LLVMBuildAdd(gallivm
->builder
, coro_hdl_idx
,
229 loop_state
[0].counter
, "");
231 LLVMValueRef coro_entry
= LLVMBuildGEP(gallivm
->builder
, coro_hdls
, &coro_hdl_idx
, 1, "");
233 LLVMValueRef coro_hdl
= LLVMBuildLoad(gallivm
->builder
, coro_entry
, "coro_hdl");
235 struct lp_build_if_state ifstate
;
236 LLVMValueRef cmp
= LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
, loop_state
[3].counter
,
237 lp_build_const_int32(gallivm
, 0), "");
238 /* first time here - call the coroutine function entry point */
239 lp_build_if(&ifstate
, gallivm
, cmp
);
240 LLVMValueRef coro_ret
= LLVMBuildCall(gallivm
->builder
, coro
, args
, 17, "");
241 LLVMBuildStore(gallivm
->builder
, coro_ret
, coro_entry
);
242 lp_build_else(&ifstate
);
243 /* subsequent calls for this invocation - check if done. */
244 LLVMValueRef coro_done
= lp_build_coro_done(gallivm
, coro_hdl
);
245 struct lp_build_if_state ifstate2
;
246 lp_build_if(&ifstate2
, gallivm
, coro_done
);
247 /* if done destroy and force loop exit */
248 lp_build_coro_destroy(gallivm
, coro_hdl
);
249 lp_build_loop_force_set_counter(&loop_state
[3], lp_build_const_int32(gallivm
, end_coroutine
- 1));
250 lp_build_else(&ifstate2
);
251 /* otherwise resume the coroutine */
252 lp_build_coro_resume(gallivm
, coro_hdl
);
253 lp_build_endif(&ifstate2
);
254 lp_build_endif(&ifstate
);
255 lp_build_loop_force_reload_counter(&loop_state
[3]);
257 lp_build_loop_end_cond(&loop_state
[0],
260 lp_build_loop_end_cond(&loop_state
[1],
263 lp_build_loop_end_cond(&loop_state
[2],
266 lp_build_loop_end_cond(&loop_state
[3],
267 lp_build_const_int32(gallivm
, end_coroutine
),
269 LLVMBuildRetVoid(builder
);
271 /* This is stage (b) - generate the compute shader code inside the coroutine. */
272 LLVMValueRef block_x_size_arg
, block_y_size_arg
, block_z_size_arg
;
273 context_ptr
= LLVMGetParam(coro
, 0);
274 x_size_arg
= LLVMGetParam(coro
, 1);
275 y_size_arg
= LLVMGetParam(coro
, 2);
276 z_size_arg
= LLVMGetParam(coro
, 3);
277 grid_x_arg
= LLVMGetParam(coro
, 4);
278 grid_y_arg
= LLVMGetParam(coro
, 5);
279 grid_z_arg
= LLVMGetParam(coro
, 6);
280 grid_size_x_arg
= LLVMGetParam(coro
, 7);
281 grid_size_y_arg
= LLVMGetParam(coro
, 8);
282 grid_size_z_arg
= LLVMGetParam(coro
, 9);
283 work_dim_arg
= LLVMGetParam(coro
, 10);
284 thread_data_ptr
= LLVMGetParam(coro
, 11);
285 num_x_loop
= LLVMGetParam(coro
, 12);
286 partials
= LLVMGetParam(coro
, 13);
287 block_x_size_arg
= LLVMGetParam(coro
, 14);
288 block_y_size_arg
= LLVMGetParam(coro
, 15);
289 block_z_size_arg
= LLVMGetParam(coro
, 16);
290 block
= LLVMAppendBasicBlockInContext(gallivm
->context
, coro
, "entry");
291 LLVMPositionBuilderAtEnd(builder
, block
);
293 LLVMValueRef consts_ptr
, num_consts_ptr
;
294 LLVMValueRef ssbo_ptr
, num_ssbo_ptr
;
295 LLVMValueRef shared_ptr
;
296 LLVMValueRef kernel_args_ptr
;
297 struct lp_build_mask_context mask
;
298 struct lp_bld_tgsi_system_values system_values
;
300 memset(&system_values
, 0, sizeof(system_values
));
301 consts_ptr
= lp_jit_cs_context_constants(gallivm
, context_ptr
);
302 num_consts_ptr
= lp_jit_cs_context_num_constants(gallivm
, context_ptr
);
303 ssbo_ptr
= lp_jit_cs_context_ssbos(gallivm
, context_ptr
);
304 num_ssbo_ptr
= lp_jit_cs_context_num_ssbos(gallivm
, context_ptr
);
305 kernel_args_ptr
= lp_jit_cs_context_kernel_args(gallivm
, context_ptr
);
307 shared_ptr
= lp_jit_cs_thread_data_shared(gallivm
, thread_data_ptr
);
309 /* these are coroutine entrypoint necessities */
310 LLVMValueRef coro_id
= lp_build_coro_id(gallivm
);
311 LLVMValueRef coro_hdl
= lp_build_coro_begin_alloc_mem(gallivm
, coro_id
);
313 LLVMValueRef has_partials
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
, partials
, lp_build_const_int32(gallivm
, 0), "");
314 LLVMValueRef tid_vals
[3];
315 LLVMValueRef tids_x
[LP_MAX_VECTOR_LENGTH
], tids_y
[LP_MAX_VECTOR_LENGTH
], tids_z
[LP_MAX_VECTOR_LENGTH
];
316 LLVMValueRef base_val
= LLVMBuildMul(gallivm
->builder
, x_size_arg
, vec_length
, "");
317 for (i
= 0; i
< cs_type
.length
; i
++) {
318 tids_x
[i
] = LLVMBuildAdd(gallivm
->builder
, base_val
, lp_build_const_int32(gallivm
, i
), "");
319 tids_y
[i
] = y_size_arg
;
320 tids_z
[i
] = z_size_arg
;
322 tid_vals
[0] = lp_build_gather_values(gallivm
, tids_x
, cs_type
.length
);
323 tid_vals
[1] = lp_build_gather_values(gallivm
, tids_y
, cs_type
.length
);
324 tid_vals
[2] = lp_build_gather_values(gallivm
, tids_z
, cs_type
.length
);
325 system_values
.thread_id
= LLVMGetUndef(LLVMArrayType(LLVMVectorType(int32_type
, cs_type
.length
), 3));
326 for (i
= 0; i
< 3; i
++)
327 system_values
.thread_id
= LLVMBuildInsertValue(builder
, system_values
.thread_id
, tid_vals
[i
], i
, "");
329 LLVMValueRef gtids
[3] = { grid_x_arg
, grid_y_arg
, grid_z_arg
};
330 system_values
.block_id
= LLVMGetUndef(LLVMVectorType(int32_type
, 3));
331 for (i
= 0; i
< 3; i
++)
332 system_values
.block_id
= LLVMBuildInsertElement(builder
, system_values
.block_id
, gtids
[i
], lp_build_const_int32(gallivm
, i
), "");
334 LLVMValueRef gstids
[3] = { grid_size_x_arg
, grid_size_y_arg
, grid_size_z_arg
};
335 system_values
.grid_size
= LLVMGetUndef(LLVMVectorType(int32_type
, 3));
336 for (i
= 0; i
< 3; i
++)
337 system_values
.grid_size
= LLVMBuildInsertElement(builder
, system_values
.grid_size
, gstids
[i
], lp_build_const_int32(gallivm
, i
), "");
339 system_values
.work_dim
= work_dim_arg
;
341 LLVMValueRef bsize
[3] = { block_x_size_arg
, block_y_size_arg
, block_z_size_arg
};
342 system_values
.block_size
= LLVMGetUndef(LLVMVectorType(int32_type
, 3));
343 for (i
= 0; i
< 3; i
++)
344 system_values
.block_size
= LLVMBuildInsertElement(builder
, system_values
.block_size
, bsize
[i
], lp_build_const_int32(gallivm
, i
), "");
346 LLVMValueRef last_x_loop
= LLVMBuildICmp(gallivm
->builder
, LLVMIntEQ
, x_size_arg
, LLVMBuildSub(gallivm
->builder
, num_x_loop
, lp_build_const_int32(gallivm
, 1), ""), "");
347 LLVMValueRef use_partial_mask
= LLVMBuildAnd(gallivm
->builder
, last_x_loop
, has_partials
, "");
348 struct lp_build_if_state if_state
;
349 LLVMValueRef mask_val
= lp_build_alloca(gallivm
, LLVMVectorType(int32_type
, cs_type
.length
), "mask");
350 LLVMValueRef full_mask_val
= lp_build_const_int_vec(gallivm
, cs_type
, ~0);
351 LLVMBuildStore(gallivm
->builder
, full_mask_val
, mask_val
);
353 lp_build_if(&if_state
, gallivm
, use_partial_mask
);
354 struct lp_build_loop_state mask_loop_state
;
355 lp_build_loop_begin(&mask_loop_state
, gallivm
, partials
);
356 LLVMValueRef tmask_val
= LLVMBuildLoad(gallivm
->builder
, mask_val
, "");
357 tmask_val
= LLVMBuildInsertElement(gallivm
->builder
, tmask_val
, lp_build_const_int32(gallivm
, 0), mask_loop_state
.counter
, "");
358 LLVMBuildStore(gallivm
->builder
, tmask_val
, mask_val
);
359 lp_build_loop_end_cond(&mask_loop_state
, vec_length
, NULL
, LLVMIntUGE
);
360 lp_build_endif(&if_state
);
362 mask_val
= LLVMBuildLoad(gallivm
->builder
, mask_val
, "");
363 lp_build_mask_begin(&mask
, gallivm
, cs_type
, mask_val
);
365 struct lp_build_coro_suspend_info coro_info
;
367 LLVMBasicBlockRef sus_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, coro
, "suspend");
368 LLVMBasicBlockRef clean_block
= LLVMAppendBasicBlockInContext(gallivm
->context
, coro
, "cleanup");
370 coro_info
.suspend
= sus_block
;
371 coro_info
.cleanup
= clean_block
;
373 struct lp_build_tgsi_params params
;
374 memset(¶ms
, 0, sizeof(params
));
376 params
.type
= cs_type
;
378 params
.consts_ptr
= consts_ptr
;
379 params
.const_sizes_ptr
= num_consts_ptr
;
380 params
.system_values
= &system_values
;
381 params
.context_ptr
= context_ptr
;
382 params
.sampler
= sampler
;
383 params
.info
= &shader
->info
.base
;
384 params
.ssbo_ptr
= ssbo_ptr
;
385 params
.ssbo_sizes_ptr
= num_ssbo_ptr
;
386 params
.image
= image
;
387 params
.shared_ptr
= shared_ptr
;
388 params
.coro
= &coro_info
;
389 params
.kernel_args
= kernel_args_ptr
;
391 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
392 lp_build_tgsi_soa(gallivm
, shader
->base
.tokens
, ¶ms
, NULL
);
394 lp_build_nir_soa(gallivm
, shader
->base
.ir
.nir
, ¶ms
,
397 mask_val
= lp_build_mask_end(&mask
);
399 lp_build_coro_suspend_switch(gallivm
, &coro_info
, NULL
, true);
400 LLVMPositionBuilderAtEnd(builder
, clean_block
);
402 lp_build_coro_free_mem(gallivm
, coro_id
, coro_hdl
);
404 LLVMBuildBr(builder
, sus_block
);
405 LLVMPositionBuilderAtEnd(builder
, sus_block
);
407 lp_build_coro_end(gallivm
, coro_hdl
);
408 LLVMBuildRet(builder
, coro_hdl
);
411 sampler
->destroy(sampler
);
412 image
->destroy(image
);
414 gallivm_verify_function(gallivm
, coro
);
415 gallivm_verify_function(gallivm
, function
);
419 llvmpipe_create_compute_state(struct pipe_context
*pipe
,
420 const struct pipe_compute_state
*templ
)
422 struct lp_compute_shader
*shader
;
423 int nr_samplers
, nr_sampler_views
;
424 shader
= CALLOC_STRUCT(lp_compute_shader
);
428 shader
->base
.type
= templ
->ir_type
;
429 if (templ
->ir_type
== PIPE_SHADER_IR_NIR_SERIALIZED
) {
430 struct blob_reader reader
;
431 const struct pipe_binary_program_header
*hdr
= templ
->prog
;
433 blob_reader_init(&reader
, hdr
->blob
, hdr
->num_bytes
);
434 shader
->base
.ir
.nir
= nir_deserialize(NULL
, pipe
->screen
->get_compiler_options(pipe
->screen
, PIPE_SHADER_IR_NIR
, PIPE_SHADER_COMPUTE
), &reader
);
435 shader
->base
.type
= PIPE_SHADER_IR_NIR
;
437 pipe
->screen
->finalize_nir(pipe
->screen
, shader
->base
.ir
.nir
, false);
438 } else if (templ
->ir_type
== PIPE_SHADER_IR_NIR
)
439 shader
->base
.ir
.nir
= (struct nir_shader
*)templ
->prog
;
441 if (shader
->base
.type
== PIPE_SHADER_IR_TGSI
) {
442 /* get/save the summary info for this shader */
443 lp_build_tgsi_info(templ
->prog
, &shader
->info
);
445 /* we need to keep a local copy of the tokens */
446 shader
->base
.tokens
= tgsi_dup_tokens(templ
->prog
);
448 nir_tgsi_scan_shader(shader
->base
.ir
.nir
, &shader
->info
.base
, false);
451 shader
->req_local_mem
= templ
->req_local_mem
;
452 make_empty_list(&shader
->variants
);
454 nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
455 nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
456 shader
->variant_key_size
= Offset(struct lp_compute_shader_variant_key
,
457 state
[MAX2(nr_samplers
, nr_sampler_views
)]);
462 llvmpipe_bind_compute_state(struct pipe_context
*pipe
,
465 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
467 if (llvmpipe
->cs
== cs
)
470 llvmpipe
->cs
= (struct lp_compute_shader
*)cs
;
471 llvmpipe
->cs_dirty
|= LP_CSNEW_CS
;
475 * Remove shader variant from two lists: the shader's variant list
476 * and the context's variant list.
479 llvmpipe_remove_cs_shader_variant(struct llvmpipe_context
*lp
,
480 struct lp_compute_shader_variant
*variant
)
482 if ((LP_DEBUG
& DEBUG_CS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
483 debug_printf("llvmpipe: del cs #%u var %u v created %u v cached %u "
484 "v total cached %u inst %u total inst %u\n",
485 variant
->shader
->no
, variant
->no
,
486 variant
->shader
->variants_created
,
487 variant
->shader
->variants_cached
,
488 lp
->nr_cs_variants
, variant
->nr_instrs
, lp
->nr_cs_instrs
);
491 gallivm_destroy(variant
->gallivm
);
493 /* remove from shader's list */
494 remove_from_list(&variant
->list_item_local
);
495 variant
->shader
->variants_cached
--;
497 /* remove from context's list */
498 remove_from_list(&variant
->list_item_global
);
499 lp
->nr_fs_variants
--;
500 lp
->nr_fs_instrs
-= variant
->nr_instrs
;
506 llvmpipe_delete_compute_state(struct pipe_context
*pipe
,
509 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
510 struct lp_compute_shader
*shader
= cs
;
511 struct lp_cs_variant_list_item
*li
;
513 if (llvmpipe
->cs
== cs
)
515 for (unsigned i
= 0; i
< shader
->max_global_buffers
; i
++)
516 pipe_resource_reference(&shader
->global_buffers
[i
], NULL
);
517 FREE(shader
->global_buffers
);
519 /* Delete all the variants */
520 li
= first_elem(&shader
->variants
);
521 while(!at_end(&shader
->variants
, li
)) {
522 struct lp_cs_variant_list_item
*next
= next_elem(li
);
523 llvmpipe_remove_cs_shader_variant(llvmpipe
, li
->base
);
526 tgsi_free_tokens(shader
->base
.tokens
);
531 make_variant_key(struct llvmpipe_context
*lp
,
532 struct lp_compute_shader
*shader
,
533 struct lp_compute_shader_variant_key
*key
)
537 memset(key
, 0, shader
->variant_key_size
);
539 /* This value will be the same for all the variants of a given shader:
541 key
->nr_samplers
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER
] + 1;
543 for(i
= 0; i
< key
->nr_samplers
; ++i
) {
544 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
545 lp_sampler_static_sampler_state(&key
->state
[i
].sampler_state
,
546 lp
->samplers
[PIPE_SHADER_COMPUTE
][i
]);
551 * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
552 * are dx10-style? Can't really have mixed opcodes, at least not
553 * if we want to skip the holes here (without rescanning tgsi).
555 if (shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] != -1) {
556 key
->nr_sampler_views
= shader
->info
.base
.file_max
[TGSI_FILE_SAMPLER_VIEW
] + 1;
557 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
559 * Note sview may exceed what's representable by file_mask.
560 * This will still work, the only downside is that not actually
561 * used views may be included in the shader key.
563 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER_VIEW
] & (1u << (i
& 31))) {
564 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
565 lp
->sampler_views
[PIPE_SHADER_COMPUTE
][i
]);
570 key
->nr_sampler_views
= key
->nr_samplers
;
571 for(i
= 0; i
< key
->nr_sampler_views
; ++i
) {
572 if(shader
->info
.base
.file_mask
[TGSI_FILE_SAMPLER
] & (1 << i
)) {
573 lp_sampler_static_texture_state(&key
->state
[i
].texture_state
,
574 lp
->sampler_views
[PIPE_SHADER_COMPUTE
][i
]);
579 key
->nr_images
= shader
->info
.base
.file_max
[TGSI_FILE_IMAGE
] + 1;
580 for (i
= 0; i
< key
->nr_images
; ++i
) {
581 if (shader
->info
.base
.file_mask
[TGSI_FILE_IMAGE
] & (1 << i
)) {
582 lp_sampler_static_texture_state_image(&key
->image_state
[i
].image_state
,
583 &lp
->images
[PIPE_SHADER_COMPUTE
][i
]);
589 dump_cs_variant_key(const struct lp_compute_shader_variant_key
*key
)
592 debug_printf("cs variant %p:\n", (void *) key
);
594 for (i
= 0; i
< key
->nr_samplers
; ++i
) {
595 const struct lp_static_sampler_state
*sampler
= &key
->state
[i
].sampler_state
;
596 debug_printf("sampler[%u] = \n", i
);
597 debug_printf(" .wrap = %s %s %s\n",
598 util_str_tex_wrap(sampler
->wrap_s
, TRUE
),
599 util_str_tex_wrap(sampler
->wrap_t
, TRUE
),
600 util_str_tex_wrap(sampler
->wrap_r
, TRUE
));
601 debug_printf(" .min_img_filter = %s\n",
602 util_str_tex_filter(sampler
->min_img_filter
, TRUE
));
603 debug_printf(" .min_mip_filter = %s\n",
604 util_str_tex_mipfilter(sampler
->min_mip_filter
, TRUE
));
605 debug_printf(" .mag_img_filter = %s\n",
606 util_str_tex_filter(sampler
->mag_img_filter
, TRUE
));
607 if (sampler
->compare_mode
!= PIPE_TEX_COMPARE_NONE
)
608 debug_printf(" .compare_func = %s\n", util_str_func(sampler
->compare_func
, TRUE
));
609 debug_printf(" .normalized_coords = %u\n", sampler
->normalized_coords
);
610 debug_printf(" .min_max_lod_equal = %u\n", sampler
->min_max_lod_equal
);
611 debug_printf(" .lod_bias_non_zero = %u\n", sampler
->lod_bias_non_zero
);
612 debug_printf(" .apply_min_lod = %u\n", sampler
->apply_min_lod
);
613 debug_printf(" .apply_max_lod = %u\n", sampler
->apply_max_lod
);
615 for (i
= 0; i
< key
->nr_sampler_views
; ++i
) {
616 const struct lp_static_texture_state
*texture
= &key
->state
[i
].texture_state
;
617 debug_printf("texture[%u] = \n", i
);
618 debug_printf(" .format = %s\n",
619 util_format_name(texture
->format
));
620 debug_printf(" .target = %s\n",
621 util_str_tex_target(texture
->target
, TRUE
));
622 debug_printf(" .level_zero_only = %u\n",
623 texture
->level_zero_only
);
624 debug_printf(" .pot = %u %u %u\n",
629 for (i
= 0; i
< key
->nr_images
; ++i
) {
630 const struct lp_static_texture_state
*image
= &key
->image_state
[i
].image_state
;
631 debug_printf("image[%u] = \n", i
);
632 debug_printf(" .format = %s\n",
633 util_format_name(image
->format
));
634 debug_printf(" .target = %s\n",
635 util_str_tex_target(image
->target
, TRUE
));
636 debug_printf(" .level_zero_only = %u\n",
637 image
->level_zero_only
);
638 debug_printf(" .pot = %u %u %u\n",
646 lp_debug_cs_variant(const struct lp_compute_shader_variant
*variant
)
648 debug_printf("llvmpipe: Compute shader #%u variant #%u:\n",
649 variant
->shader
->no
, variant
->no
);
650 if (variant
->shader
->base
.type
== PIPE_SHADER_IR_TGSI
)
651 tgsi_dump(variant
->shader
->base
.tokens
, 0);
653 nir_print_shader(variant
->shader
->base
.ir
.nir
, stderr
);
654 dump_cs_variant_key(&variant
->key
);
658 static struct lp_compute_shader_variant
*
659 generate_variant(struct llvmpipe_context
*lp
,
660 struct lp_compute_shader
*shader
,
661 const struct lp_compute_shader_variant_key
*key
)
663 struct lp_compute_shader_variant
*variant
;
664 char module_name
[64];
666 variant
= CALLOC_STRUCT(lp_compute_shader_variant
);
670 snprintf(module_name
, sizeof(module_name
), "cs%u_variant%u",
671 shader
->no
, shader
->variants_created
);
673 variant
->gallivm
= gallivm_create(module_name
, lp
->context
);
674 if (!variant
->gallivm
) {
679 variant
->shader
= shader
;
680 variant
->list_item_global
.base
= variant
;
681 variant
->list_item_local
.base
= variant
;
682 variant
->no
= shader
->variants_created
++;
684 memcpy(&variant
->key
, key
, shader
->variant_key_size
);
686 if ((LP_DEBUG
& DEBUG_CS
) || (gallivm_debug
& GALLIVM_DEBUG_IR
)) {
687 lp_debug_cs_variant(variant
);
690 lp_jit_init_cs_types(variant
);
692 generate_compute(lp
, shader
, variant
);
694 gallivm_compile_module(variant
->gallivm
);
696 variant
->nr_instrs
+= lp_build_count_ir_module(variant
->gallivm
->module
);
698 variant
->jit_function
= (lp_jit_cs_func
)gallivm_jit_function(variant
->gallivm
, variant
->function
);
700 gallivm_free_ir(variant
->gallivm
);
705 lp_cs_ctx_set_cs_variant( struct lp_cs_context
*csctx
,
706 struct lp_compute_shader_variant
*variant
)
708 csctx
->cs
.current
.variant
= variant
;
712 llvmpipe_update_cs(struct llvmpipe_context
*lp
)
714 struct lp_compute_shader
*shader
= lp
->cs
;
716 struct lp_compute_shader_variant_key key
;
717 struct lp_compute_shader_variant
*variant
= NULL
;
718 struct lp_cs_variant_list_item
*li
;
720 make_variant_key(lp
, shader
, &key
);
722 /* Search the variants for one which matches the key */
723 li
= first_elem(&shader
->variants
);
724 while(!at_end(&shader
->variants
, li
)) {
725 if(memcmp(&li
->base
->key
, &key
, shader
->variant_key_size
) == 0) {
733 /* Move this variant to the head of the list to implement LRU
734 * deletion of shader's when we have too many.
736 move_to_head(&lp
->cs_variants_list
, &variant
->list_item_global
);
739 /* variant not found, create it now */
742 unsigned variants_to_cull
;
744 if (LP_DEBUG
& DEBUG_CS
) {
745 debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
748 lp
->nr_cs_variants
? lp
->nr_cs_instrs
/ lp
->nr_cs_variants
: 0);
751 /* First, check if we've exceeded the max number of shader variants.
752 * If so, free 6.25% of them (the least recently used ones).
754 variants_to_cull
= lp
->nr_cs_variants
>= LP_MAX_SHADER_VARIANTS
? LP_MAX_SHADER_VARIANTS
/ 16 : 0;
756 if (variants_to_cull
||
757 lp
->nr_cs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
) {
758 if (gallivm_debug
& GALLIVM_DEBUG_PERF
) {
759 debug_printf("Evicting CS: %u cs variants,\t%u total variants,"
760 "\t%u instrs,\t%u instrs/variant\n",
761 shader
->variants_cached
,
762 lp
->nr_cs_variants
, lp
->nr_cs_instrs
,
763 lp
->nr_cs_instrs
/ lp
->nr_cs_variants
);
767 * We need to re-check lp->nr_cs_variants because an arbitrarliy large
768 * number of shader variants (potentially all of them) could be
769 * pending for destruction on flush.
772 for (i
= 0; i
< variants_to_cull
|| lp
->nr_cs_instrs
>= LP_MAX_SHADER_INSTRUCTIONS
; i
++) {
773 struct lp_cs_variant_list_item
*item
;
774 if (is_empty_list(&lp
->cs_variants_list
)) {
777 item
= last_elem(&lp
->cs_variants_list
);
780 llvmpipe_remove_cs_shader_variant(lp
, item
->base
);
784 * Generate the new variant.
787 variant
= generate_variant(lp
, shader
, &key
);
790 LP_COUNT_ADD(llvm_compile_time
, dt
);
791 LP_COUNT_ADD(nr_llvm_compiles
, 2); /* emit vs. omit in/out test */
793 /* Put the new variant into the list */
795 insert_at_head(&shader
->variants
, &variant
->list_item_local
);
796 insert_at_head(&lp
->cs_variants_list
, &variant
->list_item_global
);
797 lp
->nr_cs_variants
++;
798 lp
->nr_cs_instrs
+= variant
->nr_instrs
;
799 shader
->variants_cached
++;
802 /* Bind this variant */
803 lp_cs_ctx_set_cs_variant(lp
->csctx
, variant
);
807 * Called during state validation when LP_CSNEW_SAMPLER_VIEW is set.
810 lp_csctx_set_sampler_views(struct lp_cs_context
*csctx
,
812 struct pipe_sampler_view
**views
)
814 unsigned i
, max_tex_num
;
816 LP_DBG(DEBUG_SETUP
, "%s\n", __FUNCTION__
);
818 assert(num
<= PIPE_MAX_SHADER_SAMPLER_VIEWS
);
820 max_tex_num
= MAX2(num
, csctx
->cs
.current_tex_num
);
822 for (i
= 0; i
< max_tex_num
; i
++) {
823 struct pipe_sampler_view
*view
= i
< num
? views
[i
] : NULL
;
826 struct pipe_resource
*res
= view
->texture
;
827 struct llvmpipe_resource
*lp_tex
= llvmpipe_resource(res
);
828 struct lp_jit_texture
*jit_tex
;
829 jit_tex
= &csctx
->cs
.current
.jit_context
.textures
[i
];
831 /* We're referencing the texture's internal data, so save a
834 pipe_resource_reference(&csctx
->cs
.current_tex
[i
], res
);
837 /* regular texture - csctx array of mipmap level offsets */
839 unsigned first_level
= 0;
840 unsigned last_level
= 0;
842 if (llvmpipe_resource_is_texture(res
)) {
843 first_level
= view
->u
.tex
.first_level
;
844 last_level
= view
->u
.tex
.last_level
;
845 assert(first_level
<= last_level
);
846 assert(last_level
<= res
->last_level
);
847 jit_tex
->base
= lp_tex
->tex_data
;
850 jit_tex
->base
= lp_tex
->data
;
852 if (LP_PERF
& PERF_TEX_MEM
) {
853 /* use dummy tile memory */
854 jit_tex
->base
= lp_dummy_tile
;
855 jit_tex
->width
= TILE_SIZE
/8;
856 jit_tex
->height
= TILE_SIZE
/8;
858 jit_tex
->first_level
= 0;
859 jit_tex
->last_level
= 0;
860 jit_tex
->mip_offsets
[0] = 0;
861 jit_tex
->row_stride
[0] = 0;
862 jit_tex
->img_stride
[0] = 0;
865 jit_tex
->width
= res
->width0
;
866 jit_tex
->height
= res
->height0
;
867 jit_tex
->depth
= res
->depth0
;
868 jit_tex
->first_level
= first_level
;
869 jit_tex
->last_level
= last_level
;
871 if (llvmpipe_resource_is_texture(res
)) {
872 for (j
= first_level
; j
<= last_level
; j
++) {
873 jit_tex
->mip_offsets
[j
] = lp_tex
->mip_offsets
[j
];
874 jit_tex
->row_stride
[j
] = lp_tex
->row_stride
[j
];
875 jit_tex
->img_stride
[j
] = lp_tex
->img_stride
[j
];
878 if (res
->target
== PIPE_TEXTURE_1D_ARRAY
||
879 res
->target
== PIPE_TEXTURE_2D_ARRAY
||
880 res
->target
== PIPE_TEXTURE_CUBE
||
881 res
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
883 * For array textures, we don't have first_layer, instead
884 * adjust last_layer (stored as depth) plus the mip level offsets
885 * (as we have mip-first layout can't just adjust base ptr).
886 * XXX For mip levels, could do something similar.
888 jit_tex
->depth
= view
->u
.tex
.last_layer
- view
->u
.tex
.first_layer
+ 1;
889 for (j
= first_level
; j
<= last_level
; j
++) {
890 jit_tex
->mip_offsets
[j
] += view
->u
.tex
.first_layer
*
891 lp_tex
->img_stride
[j
];
893 if (view
->target
== PIPE_TEXTURE_CUBE
||
894 view
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
895 assert(jit_tex
->depth
% 6 == 0);
897 assert(view
->u
.tex
.first_layer
<= view
->u
.tex
.last_layer
);
898 assert(view
->u
.tex
.last_layer
< res
->array_size
);
903 * For buffers, we don't have "offset", instead adjust
904 * the size (stored as width) plus the base pointer.
906 unsigned view_blocksize
= util_format_get_blocksize(view
->format
);
907 /* probably don't really need to fill that out */
908 jit_tex
->mip_offsets
[0] = 0;
909 jit_tex
->row_stride
[0] = 0;
910 jit_tex
->img_stride
[0] = 0;
912 /* everything specified in number of elements here. */
913 jit_tex
->width
= view
->u
.buf
.size
/ view_blocksize
;
914 jit_tex
->base
= (uint8_t *)jit_tex
->base
+ view
->u
.buf
.offset
;
915 /* XXX Unsure if we need to sanitize parameters? */
916 assert(view
->u
.buf
.offset
+ view
->u
.buf
.size
<= res
->width0
);
921 /* display target texture/surface */
923 * XXX: Where should this be unmapped?
925 struct llvmpipe_screen
*screen
= llvmpipe_screen(res
->screen
);
926 struct sw_winsys
*winsys
= screen
->winsys
;
927 jit_tex
->base
= winsys
->displaytarget_map(winsys
, lp_tex
->dt
,
929 jit_tex
->row_stride
[0] = lp_tex
->row_stride
[0];
930 jit_tex
->img_stride
[0] = lp_tex
->img_stride
[0];
931 jit_tex
->mip_offsets
[0] = 0;
932 jit_tex
->width
= res
->width0
;
933 jit_tex
->height
= res
->height0
;
934 jit_tex
->depth
= res
->depth0
;
935 jit_tex
->first_level
= jit_tex
->last_level
= 0;
936 assert(jit_tex
->base
);
940 pipe_resource_reference(&csctx
->cs
.current_tex
[i
], NULL
);
943 csctx
->cs
.current_tex_num
= num
;
948 * Called during state validation when LP_NEW_SAMPLER is set.
951 lp_csctx_set_sampler_state(struct lp_cs_context
*csctx
,
953 struct pipe_sampler_state
**samplers
)
957 LP_DBG(DEBUG_SETUP
, "%s\n", __FUNCTION__
);
959 assert(num
<= PIPE_MAX_SAMPLERS
);
961 for (i
= 0; i
< PIPE_MAX_SAMPLERS
; i
++) {
962 const struct pipe_sampler_state
*sampler
= i
< num
? samplers
[i
] : NULL
;
965 struct lp_jit_sampler
*jit_sam
;
966 jit_sam
= &csctx
->cs
.current
.jit_context
.samplers
[i
];
968 jit_sam
->min_lod
= sampler
->min_lod
;
969 jit_sam
->max_lod
= sampler
->max_lod
;
970 jit_sam
->lod_bias
= sampler
->lod_bias
;
971 COPY_4V(jit_sam
->border_color
, sampler
->border_color
.f
);
977 lp_csctx_set_cs_constants(struct lp_cs_context
*csctx
,
979 struct pipe_constant_buffer
*buffers
)
983 LP_DBG(DEBUG_SETUP
, "%s %p\n", __FUNCTION__
, (void *) buffers
);
985 assert(num
<= ARRAY_SIZE(csctx
->constants
));
987 for (i
= 0; i
< num
; ++i
) {
988 util_copy_constant_buffer(&csctx
->constants
[i
].current
, &buffers
[i
]);
990 for (; i
< ARRAY_SIZE(csctx
->constants
); i
++) {
991 util_copy_constant_buffer(&csctx
->constants
[i
].current
, NULL
);
996 lp_csctx_set_cs_ssbos(struct lp_cs_context
*csctx
,
998 struct pipe_shader_buffer
*buffers
)
1001 LP_DBG(DEBUG_SETUP
, "%s %p\n", __FUNCTION__
, (void *)buffers
);
1003 assert (num
<= ARRAY_SIZE(csctx
->ssbos
));
1005 for (i
= 0; i
< num
; ++i
) {
1006 util_copy_shader_buffer(&csctx
->ssbos
[i
].current
, &buffers
[i
]);
1008 for (; i
< ARRAY_SIZE(csctx
->ssbos
); i
++) {
1009 util_copy_shader_buffer(&csctx
->ssbos
[i
].current
, NULL
);
1014 lp_csctx_set_cs_images(struct lp_cs_context
*csctx
,
1016 struct pipe_image_view
*images
)
1020 LP_DBG(DEBUG_SETUP
, "%s %p\n", __FUNCTION__
, (void *) images
);
1022 assert(num
<= ARRAY_SIZE(csctx
->images
));
1024 for (i
= 0; i
< num
; ++i
) {
1025 struct pipe_image_view
*image
= &images
[i
];
1026 util_copy_image_view(&csctx
->images
[i
].current
, &images
[i
]);
1028 struct pipe_resource
*res
= image
->resource
;
1029 struct llvmpipe_resource
*lp_res
= llvmpipe_resource(res
);
1030 struct lp_jit_image
*jit_image
;
1032 jit_image
= &csctx
->cs
.current
.jit_context
.images
[i
];
1036 /* regular texture - csctx array of mipmap level offsets */
1037 if (llvmpipe_resource_is_texture(res
)) {
1038 jit_image
->base
= lp_res
->tex_data
;
1040 jit_image
->base
= lp_res
->data
;
1042 jit_image
->width
= res
->width0
;
1043 jit_image
->height
= res
->height0
;
1044 jit_image
->depth
= res
->depth0
;
1046 if (llvmpipe_resource_is_texture(res
)) {
1047 uint32_t mip_offset
= lp_res
->mip_offsets
[image
->u
.tex
.level
];
1049 jit_image
->width
= u_minify(jit_image
->width
, image
->u
.tex
.level
);
1050 jit_image
->height
= u_minify(jit_image
->height
, image
->u
.tex
.level
);
1052 if (res
->target
== PIPE_TEXTURE_1D_ARRAY
||
1053 res
->target
== PIPE_TEXTURE_2D_ARRAY
||
1054 res
->target
== PIPE_TEXTURE_3D
||
1055 res
->target
== PIPE_TEXTURE_CUBE
||
1056 res
->target
== PIPE_TEXTURE_CUBE_ARRAY
) {
1058 * For array textures, we don't have first_layer, instead
1059 * adjust last_layer (stored as depth) plus the mip level offsets
1060 * (as we have mip-first layout can't just adjust base ptr).
1061 * XXX For mip levels, could do something similar.
1063 jit_image
->depth
= image
->u
.tex
.last_layer
- image
->u
.tex
.first_layer
+ 1;
1064 mip_offset
+= image
->u
.tex
.first_layer
* lp_res
->img_stride
[image
->u
.tex
.level
];
1066 jit_image
->depth
= u_minify(jit_image
->depth
, image
->u
.tex
.level
);
1068 jit_image
->row_stride
= lp_res
->row_stride
[image
->u
.tex
.level
];
1069 jit_image
->img_stride
= lp_res
->img_stride
[image
->u
.tex
.level
];
1070 jit_image
->base
= (uint8_t *)jit_image
->base
+ mip_offset
;
1072 unsigned view_blocksize
= util_format_get_blocksize(image
->format
);
1073 jit_image
->width
= image
->u
.buf
.size
/ view_blocksize
;
1074 jit_image
->base
= (uint8_t *)jit_image
->base
+ image
->u
.buf
.offset
;
1078 for (; i
< ARRAY_SIZE(csctx
->images
); i
++) {
1079 util_copy_image_view(&csctx
->images
[i
].current
, NULL
);
1084 update_csctx_consts(struct llvmpipe_context
*llvmpipe
)
1086 struct lp_cs_context
*csctx
= llvmpipe
->csctx
;
1089 for (i
= 0; i
< ARRAY_SIZE(csctx
->constants
); ++i
) {
1090 struct pipe_resource
*buffer
= csctx
->constants
[i
].current
.buffer
;
1091 const ubyte
*current_data
= NULL
;
1094 /* resource buffer */
1095 current_data
= (ubyte
*) llvmpipe_resource_data(buffer
);
1097 else if (csctx
->constants
[i
].current
.user_buffer
) {
1098 /* user-space buffer */
1099 current_data
= (ubyte
*) csctx
->constants
[i
].current
.user_buffer
;
1103 current_data
+= csctx
->constants
[i
].current
.buffer_offset
;
1105 csctx
->cs
.current
.jit_context
.constants
[i
] = (const float *)current_data
;
1106 csctx
->cs
.current
.jit_context
.num_constants
[i
] = csctx
->constants
[i
].current
.buffer_size
;
1108 csctx
->cs
.current
.jit_context
.constants
[i
] = NULL
;
1109 csctx
->cs
.current
.jit_context
.num_constants
[i
] = 0;
1115 update_csctx_ssbo(struct llvmpipe_context
*llvmpipe
)
1117 struct lp_cs_context
*csctx
= llvmpipe
->csctx
;
1119 for (i
= 0; i
< ARRAY_SIZE(csctx
->ssbos
); ++i
) {
1120 struct pipe_resource
*buffer
= csctx
->ssbos
[i
].current
.buffer
;
1121 const ubyte
*current_data
= NULL
;
1125 /* resource buffer */
1126 current_data
= (ubyte
*) llvmpipe_resource_data(buffer
);
1128 current_data
+= csctx
->ssbos
[i
].current
.buffer_offset
;
1130 csctx
->cs
.current
.jit_context
.ssbos
[i
] = (const uint32_t *)current_data
;
1131 csctx
->cs
.current
.jit_context
.num_ssbos
[i
] = csctx
->ssbos
[i
].current
.buffer_size
;
1133 csctx
->cs
.current
.jit_context
.ssbos
[i
] = NULL
;
1134 csctx
->cs
.current
.jit_context
.num_ssbos
[i
] = 0;
1140 llvmpipe_cs_update_derived(struct llvmpipe_context
*llvmpipe
, void *input
)
1142 if (llvmpipe
->cs_dirty
& (LP_CSNEW_CS
))
1143 llvmpipe_update_cs(llvmpipe
);
1145 if (llvmpipe
->cs_dirty
& LP_CSNEW_CONSTANTS
) {
1146 lp_csctx_set_cs_constants(llvmpipe
->csctx
,
1147 ARRAY_SIZE(llvmpipe
->constants
[PIPE_SHADER_COMPUTE
]),
1148 llvmpipe
->constants
[PIPE_SHADER_COMPUTE
]);
1149 update_csctx_consts(llvmpipe
);
1152 if (llvmpipe
->cs_dirty
& LP_CSNEW_SSBOS
) {
1153 lp_csctx_set_cs_ssbos(llvmpipe
->csctx
,
1154 ARRAY_SIZE(llvmpipe
->ssbos
[PIPE_SHADER_COMPUTE
]),
1155 llvmpipe
->ssbos
[PIPE_SHADER_COMPUTE
]);
1156 update_csctx_ssbo(llvmpipe
);
1159 if (llvmpipe
->cs_dirty
& LP_CSNEW_SAMPLER_VIEW
)
1160 lp_csctx_set_sampler_views(llvmpipe
->csctx
,
1161 llvmpipe
->num_sampler_views
[PIPE_SHADER_COMPUTE
],
1162 llvmpipe
->sampler_views
[PIPE_SHADER_COMPUTE
]);
1164 if (llvmpipe
->cs_dirty
& LP_CSNEW_SAMPLER
)
1165 lp_csctx_set_sampler_state(llvmpipe
->csctx
,
1166 llvmpipe
->num_samplers
[PIPE_SHADER_COMPUTE
],
1167 llvmpipe
->samplers
[PIPE_SHADER_COMPUTE
]);
1169 if (llvmpipe
->cs_dirty
& LP_CSNEW_IMAGES
)
1170 lp_csctx_set_cs_images(llvmpipe
->csctx
,
1171 ARRAY_SIZE(llvmpipe
->images
[PIPE_SHADER_COMPUTE
]),
1172 llvmpipe
->images
[PIPE_SHADER_COMPUTE
]);
1175 struct lp_cs_context
*csctx
= llvmpipe
->csctx
;
1176 csctx
->input
= input
;
1177 csctx
->cs
.current
.jit_context
.kernel_args
= input
;
1180 llvmpipe
->cs_dirty
= 0;
1184 cs_exec_fn(void *init_data
, int iter_idx
, struct lp_cs_local_mem
*lmem
)
1186 struct lp_cs_job_info
*job_info
= init_data
;
1187 struct lp_jit_cs_thread_data thread_data
;
1189 memset(&thread_data
, 0, sizeof(thread_data
));
1191 if (lmem
->local_size
< job_info
->req_local_mem
) {
1192 lmem
->local_mem_ptr
= REALLOC(lmem
->local_mem_ptr
, lmem
->local_size
,
1193 job_info
->req_local_mem
);
1194 lmem
->local_size
= job_info
->req_local_mem
;
1196 thread_data
.shared
= lmem
->local_mem_ptr
;
1198 unsigned grid_z
= iter_idx
/ (job_info
->grid_size
[0] * job_info
->grid_size
[1]);
1199 unsigned grid_y
= (iter_idx
- (grid_z
* (job_info
->grid_size
[0] * job_info
->grid_size
[1]))) / job_info
->grid_size
[0];
1200 unsigned grid_x
= (iter_idx
- (grid_z
* (job_info
->grid_size
[0] * job_info
->grid_size
[1])) - (grid_y
* job_info
->grid_size
[0]));
1201 struct lp_compute_shader_variant
*variant
= job_info
->current
->variant
;
1202 variant
->jit_function(&job_info
->current
->jit_context
,
1203 job_info
->block_size
[0], job_info
->block_size
[1], job_info
->block_size
[2],
1204 grid_x
, grid_y
, grid_z
,
1205 job_info
->grid_size
[0], job_info
->grid_size
[1], job_info
->grid_size
[2], job_info
->work_dim
,
1210 fill_grid_size(struct pipe_context
*pipe
,
1211 const struct pipe_grid_info
*info
,
1212 uint32_t grid_size
[3])
1214 struct pipe_transfer
*transfer
;
1216 if (!info
->indirect
) {
1217 grid_size
[0] = info
->grid
[0];
1218 grid_size
[1] = info
->grid
[1];
1219 grid_size
[2] = info
->grid
[2];
1222 params
= pipe_buffer_map_range(pipe
, info
->indirect
,
1223 info
->indirect_offset
,
1224 3 * sizeof(uint32_t),
1231 grid_size
[0] = params
[0];
1232 grid_size
[1] = params
[1];
1233 grid_size
[2] = params
[2];
1234 pipe_buffer_unmap(pipe
, transfer
);
1237 static void llvmpipe_launch_grid(struct pipe_context
*pipe
,
1238 const struct pipe_grid_info
*info
)
1240 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1241 struct llvmpipe_screen
*screen
= llvmpipe_screen(pipe
->screen
);
1242 struct lp_cs_job_info job_info
;
1244 memset(&job_info
, 0, sizeof(job_info
));
1246 llvmpipe_cs_update_derived(llvmpipe
, info
->input
);
1248 fill_grid_size(pipe
, info
, job_info
.grid_size
);
1250 job_info
.block_size
[0] = info
->block
[0];
1251 job_info
.block_size
[1] = info
->block
[1];
1252 job_info
.block_size
[2] = info
->block
[2];
1253 job_info
.work_dim
= info
->work_dim
;
1254 job_info
.req_local_mem
= llvmpipe
->cs
->req_local_mem
;
1255 job_info
.current
= &llvmpipe
->csctx
->cs
.current
;
1257 int num_tasks
= job_info
.grid_size
[2] * job_info
.grid_size
[1] * job_info
.grid_size
[0];
1259 struct lp_cs_tpool_task
*task
;
1260 mtx_lock(&screen
->cs_mutex
);
1261 task
= lp_cs_tpool_queue_task(screen
->cs_tpool
, cs_exec_fn
, &job_info
, num_tasks
);
1263 lp_cs_tpool_wait_for_task(screen
->cs_tpool
, &task
);
1264 mtx_unlock(&screen
->cs_mutex
);
1266 llvmpipe
->pipeline_statistics
.cs_invocations
+= num_tasks
* info
->block
[0] * info
->block
[1] * info
->block
[2];
1270 llvmpipe_set_compute_resources(struct pipe_context
*pipe
,
1271 unsigned start
, unsigned count
,
1272 struct pipe_surface
**resources
)
1279 llvmpipe_set_global_binding(struct pipe_context
*pipe
,
1280 unsigned first
, unsigned count
,
1281 struct pipe_resource
**resources
,
1284 struct llvmpipe_context
*llvmpipe
= llvmpipe_context(pipe
);
1285 struct lp_compute_shader
*cs
= llvmpipe
->cs
;
1288 if (first
+ count
> cs
->max_global_buffers
) {
1289 unsigned old_max
= cs
->max_global_buffers
;
1290 cs
->max_global_buffers
= first
+ count
;
1291 cs
->global_buffers
= realloc(cs
->global_buffers
,
1292 cs
->max_global_buffers
* sizeof(cs
->global_buffers
[0]));
1293 if (!cs
->global_buffers
) {
1297 memset(&cs
->global_buffers
[old_max
], 0, (cs
->max_global_buffers
- old_max
) * sizeof(cs
->global_buffers
[0]));
1301 for (i
= 0; i
< count
; i
++)
1302 pipe_resource_reference(&cs
->global_buffers
[first
+ i
], NULL
);
1306 for (i
= 0; i
< count
; i
++) {
1309 pipe_resource_reference(&cs
->global_buffers
[first
+ i
], resources
[i
]);
1310 struct llvmpipe_resource
*lp_res
= llvmpipe_resource(resources
[i
]);
1311 offset
= *handles
[i
];
1312 va
= (uint64_t)((char *)lp_res
->data
+ offset
);
1313 memcpy(handles
[i
], &va
, sizeof(va
));
1318 llvmpipe_init_compute_funcs(struct llvmpipe_context
*llvmpipe
)
1320 llvmpipe
->pipe
.create_compute_state
= llvmpipe_create_compute_state
;
1321 llvmpipe
->pipe
.bind_compute_state
= llvmpipe_bind_compute_state
;
1322 llvmpipe
->pipe
.delete_compute_state
= llvmpipe_delete_compute_state
;
1323 llvmpipe
->pipe
.set_compute_resources
= llvmpipe_set_compute_resources
;
1324 llvmpipe
->pipe
.set_global_binding
= llvmpipe_set_global_binding
;
1325 llvmpipe
->pipe
.launch_grid
= llvmpipe_launch_grid
;
1329 lp_csctx_destroy(struct lp_cs_context
*csctx
)
1332 for (i
= 0; i
< ARRAY_SIZE(csctx
->cs
.current_tex
); i
++) {
1333 pipe_resource_reference(&csctx
->cs
.current_tex
[i
], NULL
);
1335 for (i
= 0; i
< ARRAY_SIZE(csctx
->constants
); i
++) {
1336 pipe_resource_reference(&csctx
->constants
[i
].current
.buffer
, NULL
);
1338 for (i
= 0; i
< ARRAY_SIZE(csctx
->ssbos
); i
++) {
1339 pipe_resource_reference(&csctx
->ssbos
[i
].current
.buffer
, NULL
);
1344 struct lp_cs_context
*lp_csctx_create(struct pipe_context
*pipe
)
1346 struct lp_cs_context
*csctx
;
1348 csctx
= CALLOC_STRUCT(lp_cs_context
);