projects / mesa.git / blob
? search:


828cf5e7028ae169a0022515c7ab1d0b0cd3605e
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_nir_soa.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
26 #include "lp_bld_nir.h"
27 #include "lp_bld_init.h"
28 #include "lp_bld_flow.h"
29 #include "lp_bld_logic.h"
30 #include "lp_bld_gather.h"
31 #include "lp_bld_const.h"
32 #include "lp_bld_struct.h"
33 #include "lp_bld_arit.h"
34 #include "lp_bld_bitarit.h"
35 #include "lp_bld_coro.h"
36 #include "lp_bld_printf.h"
37 #include "util/u_math.h"
38 /*
39 * combine the execution mask if there is one with the current mask.
40 */
41 static LLVMValueRef
42 mask_vec(struct lp_build_nir_context *bld_base)
43 {
44 struct lp_build_nir_soa_context * bld = (struct lp_build_nir_soa_context *)bld_base;
45 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
46 struct lp_exec_mask *exec_mask = &bld->exec_mask;
47 LLVMValueRef bld_mask = bld->mask ? lp_build_mask_value(bld->mask) : NULL;
48 if (!exec_mask->has_mask) {
49 return bld_mask;
50 }
51 if (!bld_mask)
52 return exec_mask->exec_mask;
53 return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
54 exec_mask->exec_mask, "");
55 }
56
57 static LLVMValueRef
58 emit_fetch_64bit(
59 struct lp_build_nir_context * bld_base,
60 LLVMValueRef input,
61 LLVMValueRef input2)
62 {
63 struct gallivm_state *gallivm = bld_base->base.gallivm;
64 LLVMBuilderRef builder = gallivm->builder;
65 LLVMValueRef res;
66 int i;
67 LLVMValueRef shuffles[2 * (LP_MAX_VECTOR_WIDTH/32)];
68 int len = bld_base->base.type.length * 2;
69 assert(len <= (2 * (LP_MAX_VECTOR_WIDTH/32)));
70
71 for (i = 0; i < bld_base->base.type.length * 2; i+=2) {
72 shuffles[i] = lp_build_const_int32(gallivm, i / 2);
73 shuffles[i + 1] = lp_build_const_int32(gallivm, i / 2 + bld_base->base.type.length);
74 }
75 res = LLVMBuildShuffleVector(builder, input, input2, LLVMConstVector(shuffles, len), "");
76
77 return LLVMBuildBitCast(builder, res, bld_base->dbl_bld.vec_type, "");
78 }
79
80 static void
81 emit_store_64bit_chan(struct lp_build_nir_context *bld_base,
82 LLVMValueRef chan_ptr,
83 LLVMValueRef chan_ptr2,
84 LLVMValueRef value)
85 {
86 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
87 struct gallivm_state *gallivm = bld_base->base.gallivm;
88 LLVMBuilderRef builder = gallivm->builder;
89 struct lp_build_context *float_bld = &bld_base->base;
90 unsigned i;
91 LLVMValueRef temp, temp2;
92 LLVMValueRef shuffles[LP_MAX_VECTOR_WIDTH/32];
93 LLVMValueRef shuffles2[LP_MAX_VECTOR_WIDTH/32];
94 int len = bld_base->base.type.length * 2;
95
96 value = LLVMBuildBitCast(gallivm->builder, value, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), len), "");
97 for (i = 0; i < bld_base->base.type.length; i++) {
98 shuffles[i] = lp_build_const_int32(gallivm, i * 2);
99 shuffles2[i] = lp_build_const_int32(gallivm, (i * 2) + 1);
100 }
101
102 temp = LLVMBuildShuffleVector(builder, value,
103 LLVMGetUndef(LLVMTypeOf(value)),
104 LLVMConstVector(shuffles,
105 bld_base->base.type.length),
106 "");
107 temp2 = LLVMBuildShuffleVector(builder, value,
108 LLVMGetUndef(LLVMTypeOf(value)),
109 LLVMConstVector(shuffles2,
110 bld_base->base.type.length),
111 "");
112
113 lp_exec_mask_store(&bld->exec_mask, float_bld, temp, chan_ptr);
114 lp_exec_mask_store(&bld->exec_mask, float_bld, temp2, chan_ptr2);
115 }
116
117 static LLVMValueRef
118 get_soa_array_offsets(struct lp_build_context *uint_bld,
119 LLVMValueRef indirect_index,
120 int num_components,
121 unsigned chan_index,
122 bool need_perelement_offset)
123 {
124 struct gallivm_state *gallivm = uint_bld->gallivm;
125 LLVMValueRef chan_vec =
126 lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, chan_index);
127 LLVMValueRef length_vec =
128 lp_build_const_int_vec(gallivm, uint_bld->type, uint_bld->type.length);
129 LLVMValueRef index_vec;
130
131 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
132 index_vec = lp_build_mul(uint_bld, indirect_index, lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, num_components));
133 index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
134 index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
135
136 if (need_perelement_offset) {
137 LLVMValueRef pixel_offsets;
138 unsigned i;
139 /* build pixel offset vector: {0, 1, 2, 3, ...} */
140 pixel_offsets = uint_bld->undef;
141 for (i = 0; i < uint_bld->type.length; i++) {
142 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
143 pixel_offsets = LLVMBuildInsertElement(gallivm->builder, pixel_offsets,
144 ii, ii, "");
145 }
146 index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
147 }
148 return index_vec;
149 }
150
151 static LLVMValueRef
152 build_gather(struct lp_build_nir_context *bld_base,
153 struct lp_build_context *bld,
154 LLVMValueRef base_ptr,
155 LLVMValueRef indexes,
156 LLVMValueRef overflow_mask,
157 LLVMValueRef indexes2)
158 {
159 struct gallivm_state *gallivm = bld_base->base.gallivm;
160 LLVMBuilderRef builder = gallivm->builder;
161 struct lp_build_context *uint_bld = &bld_base->uint_bld;
162 LLVMValueRef res;
163 unsigned i;
164
165 if (indexes2)
166 res = LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), bld_base->base.type.length * 2));
167 else
168 res = bld->undef;
169 /*
170 * overflow_mask is a vector telling us which channels
171 * in the vector overflowed. We use the overflow behavior for
172 * constant buffers which is defined as:
173 * Out of bounds access to constant buffer returns 0 in all
174 * components. Out of bounds behavior is always with respect
175 * to the size of the buffer bound at that slot.
176 */
177
178 if (overflow_mask) {
179 /*
180 * We avoid per-element control flow here (also due to llvm going crazy,
181 * though I suspect it's better anyway since overflow is likely rare).
182 * Note that since we still fetch from buffers even if num_elements was
183 * zero (in this case we'll fetch from index zero) the jit func callers
184 * MUST provide valid fake constant buffers of size 4x32 (the values do
185 * not matter), otherwise we'd still need (not per element though)
186 * control flow.
187 */
188 indexes = lp_build_select(uint_bld, overflow_mask, uint_bld->zero, indexes);
189 if (indexes2)
190 indexes2 = lp_build_select(uint_bld, overflow_mask, uint_bld->zero, indexes2);
191 }
192
193 /*
194 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
195 */
196 for (i = 0; i < bld->type.length * (indexes2 ? 2 : 1); i++) {
197 LLVMValueRef si, di;
198 LLVMValueRef index;
199 LLVMValueRef scalar_ptr, scalar;
200
201 di = lp_build_const_int32(gallivm, i);
202 if (indexes2)
203 si = lp_build_const_int32(gallivm, i >> 1);
204 else
205 si = di;
206
207 if (indexes2 && (i & 1)) {
208 index = LLVMBuildExtractElement(builder,
209 indexes2, si, "");
210 } else {
211 index = LLVMBuildExtractElement(builder,
212 indexes, si, "");
213 }
214 scalar_ptr = LLVMBuildGEP(builder, base_ptr,
215 &index, 1, "gather_ptr");
216 scalar = LLVMBuildLoad(builder, scalar_ptr, "");
217
218 res = LLVMBuildInsertElement(builder, res, scalar, di, "");
219 }
220
221 if (overflow_mask) {
222 if (indexes2) {
223 res = LLVMBuildBitCast(builder, res, bld_base->dbl_bld.vec_type, "");
224 overflow_mask = LLVMBuildSExt(builder, overflow_mask,
225 bld_base->dbl_bld.int_vec_type, "");
226 res = lp_build_select(&bld_base->dbl_bld, overflow_mask,
227 bld_base->dbl_bld.zero, res);
228 } else
229 res = lp_build_select(bld, overflow_mask, bld->zero, res);
230 }
231
232 return res;
233 }
234
235 /**
236 * Scatter/store vector.
237 */
238 static void
239 emit_mask_scatter(struct lp_build_nir_soa_context *bld,
240 LLVMValueRef base_ptr,
241 LLVMValueRef indexes,
242 LLVMValueRef values,
243 struct lp_exec_mask *mask)
244 {
245 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
246 LLVMBuilderRef builder = gallivm->builder;
247 unsigned i;
248 LLVMValueRef pred = mask->has_mask ? mask->exec_mask : NULL;
249
250 /*
251 * Loop over elements of index_vec, store scalar value.
252 */
253 for (i = 0; i < bld->bld_base.base.type.length; i++) {
254 LLVMValueRef ii = lp_build_const_int32(gallivm, i);
255 LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
256 LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
257 LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
258 LLVMValueRef scalar_pred = pred ?
259 LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
260
261 if (0)
262 lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
263 ii, val, index, scalar_ptr);
264
265 if (scalar_pred) {
266 LLVMValueRef real_val, dst_val;
267 dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
268 real_val = lp_build_select(&bld->uint_elem_bld, scalar_pred, val, dst_val);
269 LLVMBuildStore(builder, real_val, scalar_ptr);
270 }
271 else {
272 LLVMBuildStore(builder, val, scalar_ptr);
273 }
274 }
275 }
276
277 static void emit_load_var(struct lp_build_nir_context *bld_base,
278 nir_variable_mode deref_mode,
279 unsigned num_components,
280 unsigned bit_size,
281 nir_variable *var,
282 unsigned vertex_index,
283 unsigned const_index,
284 LLVMValueRef indir_index,
285 LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
286 {
287 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
288 struct gallivm_state *gallivm = bld_base->base.gallivm;
289 int dmul = bit_size == 64 ? 2 : 1;
290 switch (deref_mode) {
291 case nir_var_shader_in: {
292 for (unsigned i = 0; i < num_components; i++) {
293 int idx = (i * dmul) + var->data.location_frac;
294 if (bld->gs_iface) {
295 LLVMValueRef vertex_index_val = lp_build_const_int32(gallivm, vertex_index);
296 LLVMValueRef attrib_index_val = lp_build_const_int32(gallivm, const_index + var->data.driver_location);
297 LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx);
298 LLVMValueRef result2;
299 result[i] = bld->gs_iface->fetch_input(bld->gs_iface, &bld_base->base,
300 false, vertex_index_val, 0, attrib_index_val, swizzle_index_val);
301 if (bit_size == 64) {
302 LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
303 result2 = bld->gs_iface->fetch_input(bld->gs_iface, &bld_base->base,
304 false, vertex_index_val, 0, attrib_index_val, swizzle_index_val);
305 result[i] = emit_fetch_64bit(bld_base, result[i], result2);
306 }
307 } else {
308 if (indir_index) {
309 LLVMValueRef attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, var->data.driver_location));
310 LLVMValueRef index_vec = get_soa_array_offsets(&bld_base->uint_bld,
311 attrib_index_val, 4, idx,
312 TRUE);
313 LLVMValueRef index_vec2 = NULL;
314 LLVMTypeRef fptr_type;
315 LLVMValueRef inputs_array;
316 fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
317 inputs_array = LLVMBuildBitCast(gallivm->builder, bld->inputs_array, fptr_type, "");
318
319 if (bit_size == 64)
320 index_vec2 = get_soa_array_offsets(&bld_base->uint_bld,
321 indir_index, 4, idx + 1, TRUE);
322
323 /* Gather values from the input register array */
324 result[i] = build_gather(bld_base, &bld_base->base, inputs_array, index_vec, NULL, index_vec2);
325 } else {
326 if (bld->indirects & nir_var_shader_in) {
327 LLVMValueRef lindex = lp_build_const_int32(gallivm,
328 var->data.driver_location * 4 + idx);
329 LLVMValueRef input_ptr = lp_build_pointer_get(gallivm->builder,
330 bld->inputs_array, lindex);
331 if (bit_size == 64) {
332 LLVMValueRef lindex2 = lp_build_const_int32(gallivm,
333 var->data.driver_location * 4 + (idx + 1));
334 LLVMValueRef input_ptr2 = lp_build_pointer_get(gallivm->builder,
335 bld->inputs_array, lindex2);
336 result[i] = emit_fetch_64bit(bld_base, input_ptr, input_ptr2);
337 } else {
338 result[i] = input_ptr;
339 }
340 } else {
341 if (bit_size == 64) {
342 LLVMValueRef tmp[2];
343 tmp[0] = bld->inputs[var->data.driver_location + const_index][idx];
344 tmp[1] = bld->inputs[var->data.driver_location + const_index][idx + 1];
345 result[i] = emit_fetch_64bit(bld_base, tmp[0], tmp[1]);
346 } else {
347 result[i] = bld->inputs[var->data.driver_location + const_index][idx];
348 }
349 }
350 }
351 }
352 }
353 }
354 default:
355 break;
356 }
357 }
358
359 static void emit_store_chan(struct lp_build_nir_context *bld_base,
360 nir_variable_mode deref_mode,
361 unsigned bit_size,
362 unsigned location, unsigned comp,
363 unsigned chan,
364 LLVMValueRef dst)
365 {
366 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
367 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
368 struct lp_build_context *float_bld = &bld_base->base;
369
370 if (bit_size == 64) {
371 chan *= 2;
372 chan += comp;
373 if (chan >= 4) {
374 chan -= 4;
375 location++;
376 }
377 emit_store_64bit_chan(bld_base, bld->outputs[location][chan],
378 bld->outputs[location][chan + 1], dst);
379 } else {
380 dst = LLVMBuildBitCast(builder, dst, float_bld->vec_type, "");
381 lp_exec_mask_store(&bld->exec_mask, float_bld, dst,
382 bld->outputs[location][chan + comp]);
383 }
384 }
385
386 static void emit_store_var(struct lp_build_nir_context *bld_base,
387 nir_variable_mode deref_mode,
388 unsigned bit_size,
389 unsigned num_components,
390 unsigned writemask,
391 unsigned const_index,
392 nir_variable *var, LLVMValueRef dst)
393 {
394 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
395 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
396 switch (deref_mode) {
397 case nir_var_shader_out: {
398 unsigned location = var->data.driver_location + const_index;
399 unsigned comp = var->data.location_frac;
400 if (bld_base->shader->info.stage == MESA_SHADER_FRAGMENT) {
401 if (var->data.location == FRAG_RESULT_STENCIL)
402 comp = 1;
403 else if (var->data.location == FRAG_RESULT_DEPTH)
404 comp = 2;
405 }
406 for (unsigned chan = 0; chan < num_components; chan++) {
407 if (writemask & (1u << chan)) {
408 LLVMValueRef chan_val = (num_components == 1) ? dst : LLVMBuildExtractValue(builder, dst, chan, "");
409 emit_store_chan(bld_base, deref_mode, bit_size, location, comp, chan, chan_val);
410 }
411 }
412 break;
413 }
414 default:
415 break;
416 }
417 }
418
419 static LLVMValueRef emit_load_reg(struct lp_build_nir_context *bld_base,
420 struct lp_build_context *reg_bld,
421 const nir_reg_src *reg,
422 LLVMValueRef indir_src,
423 LLVMValueRef reg_storage)
424 {
425 struct gallivm_state *gallivm = bld_base->base.gallivm;
426 LLVMBuilderRef builder = gallivm->builder;
427 int nc = reg->reg->num_components;
428 LLVMValueRef vals[NIR_MAX_VEC_COMPONENTS];
429 struct lp_build_context *uint_bld = &bld_base->uint_bld;
430 if (reg->reg->num_array_elems) {
431 LLVMValueRef indirect_val = lp_build_const_int_vec(gallivm, uint_bld->type, reg->base_offset);
432 if (reg->indirect) {
433 LLVMValueRef max_index = lp_build_const_int_vec(gallivm, uint_bld->type, reg->reg->num_array_elems - 1);
434 indirect_val = LLVMBuildAdd(builder, indirect_val, indir_src, "");
435 indirect_val = lp_build_min(uint_bld, indirect_val, max_index);
436 }
437 reg_storage = LLVMBuildBitCast(builder, reg_storage, LLVMPointerType(reg_bld->elem_type, 0), "");
438 for (unsigned i = 0; i < nc; i++) {
439 LLVMValueRef indirect_offset = get_soa_array_offsets(uint_bld, indirect_val, nc, i, TRUE);
440 vals[i] = build_gather(bld_base, reg_bld, reg_storage, indirect_offset, NULL, NULL);
441 }
442 } else {
443 for (unsigned i = 0; i < nc; i++) {
444 LLVMValueRef this_storage = nc == 1 ? reg_storage : lp_build_array_get_ptr(gallivm, reg_storage,
445 lp_build_const_int32(gallivm, i));
446 vals[i] = LLVMBuildLoad(builder, this_storage, "");
447 }
448 }
449 return nc == 1 ? vals[0] : lp_nir_array_build_gather_values(builder, vals, nc);
450 }
451
452 static void emit_store_reg(struct lp_build_nir_context *bld_base,
453 struct lp_build_context *reg_bld,
454 const nir_reg_dest *reg,
455 unsigned writemask,
456 LLVMValueRef indir_src,
457 LLVMValueRef reg_storage,
458 LLVMValueRef dst[NIR_MAX_VEC_COMPONENTS])
459 {
460 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
461 struct gallivm_state *gallivm = bld_base->base.gallivm;
462 LLVMBuilderRef builder = gallivm->builder;
463 struct lp_build_context *uint_bld = &bld_base->uint_bld;
464 int nc = reg->reg->num_components;
465 if (reg->reg->num_array_elems > 0) {
466 LLVMValueRef indirect_val = lp_build_const_int_vec(gallivm, uint_bld->type, reg->base_offset);
467 if (reg->indirect) {
468 LLVMValueRef max_index = lp_build_const_int_vec(gallivm, uint_bld->type, reg->reg->num_array_elems - 1);
469 indirect_val = LLVMBuildAdd(builder, indirect_val, indir_src, "");
470 indirect_val = lp_build_min(uint_bld, indirect_val, max_index);
471 }
472 reg_storage = LLVMBuildBitCast(builder, reg_storage, LLVMPointerType(reg_bld->elem_type, 0), "");
473 for (unsigned i = 0; i < nc; i++) {
474 if (!(writemask & (1 << i)))
475 continue;
476 LLVMValueRef indirect_offset = get_soa_array_offsets(uint_bld, indirect_val, nc, i, TRUE);
477 dst[i] = LLVMBuildBitCast(builder, dst[i], reg_bld->vec_type, "");
478 emit_mask_scatter(bld, reg_storage, indirect_offset, dst[i], &bld->exec_mask);
479 }
480 return;
481 }
482
483 for (unsigned i = 0; i < nc; i++) {
484 LLVMValueRef this_storage = nc == 1 ? reg_storage : lp_build_array_get_ptr(gallivm, reg_storage,
485 lp_build_const_int32(gallivm, i));
486 dst[i] = LLVMBuildBitCast(builder, dst[i], reg_bld->vec_type, "");
487 lp_exec_mask_store(&bld->exec_mask, reg_bld, dst[i], this_storage);
488 }
489 }
490
491 static void emit_load_kernel_arg(struct lp_build_nir_context *bld_base,
492 unsigned nc,
493 unsigned bit_size,
494 unsigned offset_bit_size,
495 bool offset_is_uniform,
496 LLVMValueRef offset,
497 LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
498 {
499 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
500 struct gallivm_state *gallivm = bld_base->base.gallivm;
501 LLVMBuilderRef builder = gallivm->builder;
502 struct lp_build_context *bld_broad = get_int_bld(bld_base, true, bit_size);
503 LLVMValueRef kernel_args_ptr = bld->kernel_args_ptr;
504 unsigned size_shift = 0;
505 struct lp_build_context *bld_offset = get_int_bld(bld_base, true, offset_bit_size);
506 if (bit_size == 16)
507 size_shift = 1;
508 else if (bit_size == 32)
509 size_shift = 2;
510 else if (bit_size == 64)
511 size_shift = 3;
512 if (size_shift)
513 offset = lp_build_shr(bld_offset, offset, lp_build_const_int_vec(gallivm, bld_offset->type, size_shift));
514
515 LLVMTypeRef ptr_type = LLVMPointerType(bld_broad->elem_type, 0);
516 kernel_args_ptr = LLVMBuildBitCast(builder, kernel_args_ptr, ptr_type, "");
517
518 if (offset_is_uniform) {
519 offset = LLVMBuildExtractElement(builder, offset, lp_build_const_int32(gallivm, 0), "");
520
521 for (unsigned c = 0; c < nc; c++) {
522 LLVMValueRef this_offset = LLVMBuildAdd(builder, offset, offset_bit_size == 64 ? lp_build_const_int64(gallivm, c) : lp_build_const_int32(gallivm, c), "");
523
524 LLVMValueRef scalar = lp_build_pointer_get(builder, kernel_args_ptr, this_offset);
525 result[c] = lp_build_broadcast_scalar(bld_broad, scalar);
526 }
527 }
528 }
529
530 static LLVMValueRef global_addr_to_ptr(struct gallivm_state *gallivm, LLVMValueRef addr_ptr, unsigned bit_size)
531 {
532 LLVMBuilderRef builder = gallivm->builder;
533 switch (bit_size) {
534 case 8:
535 addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
536 break;
537 case 16:
538 addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt16TypeInContext(gallivm->context), 0), "");
539 break;
540 case 32:
541 default:
542 addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0), "");
543 break;
544 case 64:
545 addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt64TypeInContext(gallivm->context), 0), "");
546 break;
547 }
548 return addr_ptr;
549 }
550
551 static void emit_load_global(struct lp_build_nir_context *bld_base,
552 unsigned nc,
553 unsigned bit_size,
554 unsigned addr_bit_size,
555 LLVMValueRef addr,
556 LLVMValueRef outval[NIR_MAX_VEC_COMPONENTS])
557 {
558 struct gallivm_state *gallivm = bld_base->base.gallivm;
559 LLVMBuilderRef builder = gallivm->builder;
560 struct lp_build_context *uint_bld = &bld_base->uint_bld;
561 struct lp_build_context *res_bld;
562
563 res_bld = get_int_bld(bld_base, true, bit_size);
564
565 for (unsigned c = 0; c < nc; c++) {
566 LLVMValueRef result = lp_build_alloca(gallivm, res_bld->vec_type, "");
567
568 struct lp_build_loop_state loop_state;
569 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
570
571 LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
572 loop_state.counter, "");
573 addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, bit_size);
574
575 LLVMValueRef value_ptr = lp_build_pointer_get(builder, addr_ptr, lp_build_const_int32(gallivm, c));
576
577 LLVMValueRef temp_res;
578 temp_res = LLVMBuildLoad(builder, result, "");
579 temp_res = LLVMBuildInsertElement(builder, temp_res, value_ptr, loop_state.counter, "");
580 LLVMBuildStore(builder, temp_res, result);
581 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
582 NULL, LLVMIntUGE);
583 outval[c] = LLVMBuildLoad(builder, result, "");
584 }
585 }
586
587 static void emit_store_global(struct lp_build_nir_context *bld_base,
588 unsigned writemask,
589 unsigned nc, unsigned bit_size,
590 unsigned addr_bit_size,
591 LLVMValueRef addr,
592 LLVMValueRef dst)
593 {
594 struct gallivm_state *gallivm = bld_base->base.gallivm;
595 LLVMBuilderRef builder = gallivm->builder;
596 struct lp_build_context *uint_bld = &bld_base->uint_bld;
597
598 for (unsigned c = 0; c < nc; c++) {
599 if (!(writemask & (1u << c)))
600 continue;
601 LLVMValueRef val = (nc == 1) ? dst : LLVMBuildExtractValue(builder, dst, c, "");
602
603 LLVMValueRef exec_mask = mask_vec(bld_base);
604 struct lp_build_loop_state loop_state;
605 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
606 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
607 loop_state.counter, "");
608
609 LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
610 loop_state.counter, "");
611 addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, bit_size);
612 switch (bit_size) {
613 case 32:
614 value_ptr = LLVMBuildBitCast(builder, value_ptr, LLVMInt32TypeInContext(gallivm->context), "");
615 break;
616 case 64:
617 value_ptr = LLVMBuildBitCast(builder, value_ptr, LLVMInt64TypeInContext(gallivm->context), "");
618 break;
619 default:
620 break;
621 }
622 struct lp_build_if_state ifthen;
623
624 LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
625 cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
626 lp_build_if(&ifthen, gallivm, cond);
627 lp_build_pointer_set(builder, addr_ptr, lp_build_const_int32(gallivm, c), value_ptr);
628 lp_build_endif(&ifthen);
629 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
630 NULL, LLVMIntUGE);
631 }
632 }
633
634 static void emit_atomic_global(struct lp_build_nir_context *bld_base,
635 nir_intrinsic_op nir_op,
636 unsigned addr_bit_size,
637 LLVMValueRef addr,
638 LLVMValueRef val, LLVMValueRef val2,
639 LLVMValueRef *result)
640 {
641 struct gallivm_state *gallivm = bld_base->base.gallivm;
642 LLVMBuilderRef builder = gallivm->builder;
643 struct lp_build_context *uint_bld = &bld_base->uint_bld;
644 LLVMAtomicRMWBinOp op;
645 switch (nir_op) {
646 case nir_intrinsic_global_atomic_add:
647 op = LLVMAtomicRMWBinOpAdd;
648 break;
649 case nir_intrinsic_global_atomic_exchange:
650 op = LLVMAtomicRMWBinOpXchg;
651 break;
652 case nir_intrinsic_global_atomic_and:
653 op = LLVMAtomicRMWBinOpAnd;
654 break;
655 case nir_intrinsic_global_atomic_or:
656 op = LLVMAtomicRMWBinOpOr;
657 break;
658 case nir_intrinsic_global_atomic_xor:
659 op = LLVMAtomicRMWBinOpXor;
660 break;
661 case nir_intrinsic_global_atomic_umin:
662 op = LLVMAtomicRMWBinOpUMin;
663 break;
664 case nir_intrinsic_global_atomic_umax:
665 op = LLVMAtomicRMWBinOpUMax;
666 break;
667 case nir_intrinsic_global_atomic_imin:
668 op = LLVMAtomicRMWBinOpMin;
669 break;
670 case nir_intrinsic_global_atomic_imax:
671 op = LLVMAtomicRMWBinOpMax;
672 break;
673 default:
674 break;
675 }
676
677 LLVMValueRef atom_res = lp_build_alloca(gallivm,
678 uint_bld->vec_type, "");
679 LLVMValueRef exec_mask = mask_vec(bld_base);
680 struct lp_build_loop_state loop_state;
681 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
682
683 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
684 loop_state.counter, "");
685
686 LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
687 loop_state.counter, "");
688 addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, 32);
689 struct lp_build_if_state ifthen;
690 LLVMValueRef cond, temp_res;
691 LLVMValueRef scalar;
692 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
693 cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
694 lp_build_if(&ifthen, gallivm, cond);
695
696 if (nir_op == nir_intrinsic_global_atomic_comp_swap) {
697 LLVMValueRef cas_src_ptr = LLVMBuildExtractElement(gallivm->builder, val2,
698 loop_state.counter, "");
699 cas_src_ptr = LLVMBuildBitCast(gallivm->builder, cas_src_ptr, uint_bld->elem_type, "");
700 scalar = LLVMBuildAtomicCmpXchg(builder, addr_ptr, value_ptr,
701 cas_src_ptr,
702 LLVMAtomicOrderingSequentiallyConsistent,
703 LLVMAtomicOrderingSequentiallyConsistent,
704 false);
705 scalar = LLVMBuildExtractValue(gallivm->builder, scalar, 0, "");
706 } else {
707 scalar = LLVMBuildAtomicRMW(builder, op,
708 addr_ptr, value_ptr,
709 LLVMAtomicOrderingSequentiallyConsistent,
710 false);
711 }
712 temp_res = LLVMBuildLoad(builder, atom_res, "");
713 temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
714 LLVMBuildStore(builder, temp_res, atom_res);
715 lp_build_else(&ifthen);
716 temp_res = LLVMBuildLoad(builder, atom_res, "");
717 temp_res = LLVMBuildInsertElement(builder, temp_res, lp_build_const_int32(gallivm, 0), loop_state.counter, "");
718 LLVMBuildStore(builder, temp_res, atom_res);
719 lp_build_endif(&ifthen);
720 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
721 NULL, LLVMIntUGE);
722 *result = LLVMBuildLoad(builder, atom_res, "");
723 }
724
725 static void emit_load_ubo(struct lp_build_nir_context *bld_base,
726 unsigned nc,
727 unsigned bit_size,
728 bool offset_is_uniform,
729 LLVMValueRef index,
730 LLVMValueRef offset,
731 LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
732 {
733 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
734 struct gallivm_state *gallivm = bld_base->base.gallivm;
735 LLVMBuilderRef builder = gallivm->builder;
736 struct lp_build_context *uint_bld = &bld_base->uint_bld;
737 struct lp_build_context *bld_broad = bit_size == 64 ? &bld_base->dbl_bld : &bld_base->base;
738 LLVMValueRef consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, index);
739 unsigned size_shift = 0;
740 if (bit_size == 32)
741 size_shift = 2;
742 else if (bit_size == 64)
743 size_shift = 3;
744 if (size_shift)
745 offset = lp_build_shr(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, size_shift));
746 if (bit_size == 64) {
747 LLVMTypeRef dptr_type = LLVMPointerType(bld_base->dbl_bld.elem_type, 0);
748 consts_ptr = LLVMBuildBitCast(builder, consts_ptr, dptr_type, "");
749 }
750
751 if (offset_is_uniform) {
752 offset = LLVMBuildExtractElement(builder, offset, lp_build_const_int32(gallivm, 0), "");
753
754 for (unsigned c = 0; c < nc; c++) {
755 LLVMValueRef this_offset = LLVMBuildAdd(builder, offset, lp_build_const_int32(gallivm, c), "");
756
757 LLVMValueRef scalar = lp_build_pointer_get(builder, consts_ptr, this_offset);
758 result[c] = lp_build_broadcast_scalar(bld_broad, scalar);
759 }
760 } else {
761 LLVMValueRef overflow_mask;
762 LLVMValueRef num_consts = lp_build_array_get(gallivm, bld->const_sizes_ptr, index);
763
764 num_consts = LLVMBuildShl(gallivm->builder, num_consts, lp_build_const_int32(gallivm, 4), "");
765 num_consts = lp_build_broadcast_scalar(uint_bld, num_consts);
766 for (unsigned c = 0; c < nc; c++) {
767 LLVMValueRef this_offset = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
768 overflow_mask = lp_build_compare(gallivm, uint_bld->type, PIPE_FUNC_GEQUAL,
769 this_offset, num_consts);
770
771 result[c] = build_gather(bld_base, bld_broad, consts_ptr, this_offset, overflow_mask, NULL);
772 }
773 }
774 }
775
776
777 static void emit_load_mem(struct lp_build_nir_context *bld_base,
778 unsigned nc,
779 unsigned bit_size,
780 LLVMValueRef index,
781 LLVMValueRef offset,
782 LLVMValueRef outval[NIR_MAX_VEC_COMPONENTS])
783 {
784 struct gallivm_state *gallivm = bld_base->base.gallivm;
785 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
786 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
787 LLVMValueRef ssbo_ptr = NULL;
788 struct lp_build_context *uint_bld = &bld_base->uint_bld;
789 struct lp_build_context *uint64_bld = &bld_base->uint64_bld;
790 LLVMValueRef ssbo_limit = NULL;
791
792 if (index) {
793 LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
794 ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, bit_size == 64 ? 3 : 2), "");
795 ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
796
797 ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
798 } else
799 ssbo_ptr = bld->shared_ptr;
800
801 offset = LLVMBuildAShr(gallivm->builder, offset, lp_build_const_int_vec(gallivm, uint_bld->type, bit_size == 64 ? 3 : 2), "");
802 for (unsigned c = 0; c < nc; c++) {
803 LLVMValueRef loop_index = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
804 LLVMValueRef exec_mask = mask_vec(bld_base);
805
806 if (ssbo_limit) {
807 LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, loop_index, ssbo_limit);
808 exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
809 }
810
811 LLVMValueRef result = lp_build_alloca(gallivm, bit_size == 64 ? uint64_bld->vec_type : uint_bld->vec_type, "");
812 struct lp_build_loop_state loop_state;
813 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
814
815 struct lp_build_if_state ifthen;
816 LLVMValueRef cond, temp_res;
817
818 loop_index = LLVMBuildExtractElement(gallivm->builder, loop_index,
819 loop_state.counter, "");
820
821 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
822 cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
823
824 lp_build_if(&ifthen, gallivm, cond);
825 LLVMValueRef scalar;
826 if (bit_size == 64) {
827 LLVMValueRef ssbo_ptr2 = LLVMBuildBitCast(builder, ssbo_ptr, LLVMPointerType(uint64_bld->elem_type, 0), "");
828 scalar = lp_build_pointer_get(builder, ssbo_ptr2, loop_index);
829 } else
830 scalar = lp_build_pointer_get(builder, ssbo_ptr, loop_index);
831
832 temp_res = LLVMBuildLoad(builder, result, "");
833 temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
834 LLVMBuildStore(builder, temp_res, result);
835 lp_build_else(&ifthen);
836 temp_res = LLVMBuildLoad(builder, result, "");
837 LLVMValueRef zero;
838 if (bit_size == 64)
839 zero = LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 0, 0);
840 else
841 zero = lp_build_const_int32(gallivm, 0);
842 temp_res = LLVMBuildInsertElement(builder, temp_res, zero, loop_state.counter, "");
843 LLVMBuildStore(builder, temp_res, result);
844 lp_build_endif(&ifthen);
845 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
846 NULL, LLVMIntUGE);
847 outval[c] = LLVMBuildLoad(gallivm->builder, result, "");
848 }
849 }
850
851 static void emit_store_mem(struct lp_build_nir_context *bld_base,
852 unsigned writemask,
853 unsigned nc,
854 unsigned bit_size,
855 LLVMValueRef index,
856 LLVMValueRef offset,
857 LLVMValueRef dst)
858 {
859 struct gallivm_state *gallivm = bld_base->base.gallivm;
860 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
861 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
862 LLVMValueRef ssbo_ptr;
863 struct lp_build_context *uint_bld = &bld_base->uint_bld;
864 LLVMValueRef ssbo_limit = NULL;
865
866 if (index) {
867 LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
868 ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, bit_size == 64 ? 3 : 2), "");
869 ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
870 ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
871 } else
872 ssbo_ptr = bld->shared_ptr;
873
874 offset = lp_build_shr_imm(uint_bld, offset, bit_size == 64 ? 3 : 2);
875 for (unsigned c = 0; c < nc; c++) {
876 if (!(writemask & (1u << c)))
877 continue;
878 LLVMValueRef loop_index = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
879 LLVMValueRef val = (nc == 1) ? dst : LLVMBuildExtractValue(builder, dst, c, "");
880
881 LLVMValueRef exec_mask = mask_vec(bld_base);
882 if (ssbo_limit) {
883 LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, loop_index, ssbo_limit);
884 exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
885 }
886
887 struct lp_build_loop_state loop_state;
888 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
889 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
890 loop_state.counter, "");
891 if (bit_size == 64)
892 value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, bld_base->uint64_bld.elem_type, "");
893 else
894 value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, uint_bld->elem_type, "");
895 struct lp_build_if_state ifthen;
896 LLVMValueRef cond;
897
898 loop_index = LLVMBuildExtractElement(gallivm->builder, loop_index,
899 loop_state.counter, "");
900 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
901 cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
902 lp_build_if(&ifthen, gallivm, cond);
903 if (bit_size == 64) {
904 LLVMValueRef ssbo_ptr2 = LLVMBuildBitCast(builder, ssbo_ptr, LLVMPointerType(bld_base->uint64_bld.elem_type, 0), "");
905 lp_build_pointer_set(builder, ssbo_ptr2, loop_index, value_ptr);
906 } else
907 lp_build_pointer_set(builder, ssbo_ptr, loop_index, value_ptr);
908 lp_build_endif(&ifthen);
909 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
910 NULL, LLVMIntUGE);
911 }
912 }
913
914 static void emit_atomic_mem(struct lp_build_nir_context *bld_base,
915 nir_intrinsic_op nir_op,
916 LLVMValueRef index, LLVMValueRef offset,
917 LLVMValueRef val, LLVMValueRef val2,
918 LLVMValueRef *result)
919 {
920 struct gallivm_state *gallivm = bld_base->base.gallivm;
921 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
922 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
923 LLVMValueRef ssbo_ptr;
924 struct lp_build_context *uint_bld = &bld_base->uint_bld;
925 LLVMAtomicRMWBinOp op;
926 LLVMValueRef ssbo_limit = NULL;
927
928 if (index) {
929 LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
930 ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, 2), "");
931 ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
932 ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
933 } else
934 ssbo_ptr = bld->shared_ptr;
935
936 switch (nir_op) {
937 case nir_intrinsic_shared_atomic_add:
938 case nir_intrinsic_ssbo_atomic_add:
939 op = LLVMAtomicRMWBinOpAdd;
940 break;
941 case nir_intrinsic_shared_atomic_exchange:
942 case nir_intrinsic_ssbo_atomic_exchange:
943 op = LLVMAtomicRMWBinOpXchg;
944 break;
945 case nir_intrinsic_shared_atomic_and:
946 case nir_intrinsic_ssbo_atomic_and:
947 op = LLVMAtomicRMWBinOpAnd;
948 break;
949 case nir_intrinsic_shared_atomic_or:
950 case nir_intrinsic_ssbo_atomic_or:
951 op = LLVMAtomicRMWBinOpOr;
952 break;
953 case nir_intrinsic_shared_atomic_xor:
954 case nir_intrinsic_ssbo_atomic_xor:
955 op = LLVMAtomicRMWBinOpXor;
956 break;
957 case nir_intrinsic_shared_atomic_umin:
958 case nir_intrinsic_ssbo_atomic_umin:
959 op = LLVMAtomicRMWBinOpUMin;
960 break;
961 case nir_intrinsic_shared_atomic_umax:
962 case nir_intrinsic_ssbo_atomic_umax:
963 op = LLVMAtomicRMWBinOpUMax;
964 break;
965 case nir_intrinsic_ssbo_atomic_imin:
966 case nir_intrinsic_shared_atomic_imin:
967 op = LLVMAtomicRMWBinOpMin;
968 break;
969 case nir_intrinsic_ssbo_atomic_imax:
970 case nir_intrinsic_shared_atomic_imax:
971 op = LLVMAtomicRMWBinOpMax;
972 break;
973 default:
974 break;
975 }
976
977 offset = lp_build_shr_imm(uint_bld, offset, 2);
978 LLVMValueRef atom_res = lp_build_alloca(gallivm,
979 uint_bld->vec_type, "");
980
981 LLVMValueRef exec_mask = mask_vec(bld_base);
982 if (ssbo_limit) {
983 LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, offset, ssbo_limit);
984 exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
985 }
986
987 struct lp_build_loop_state loop_state;
988 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
989
990 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
991 loop_state.counter, "");
992 value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, uint_bld->elem_type, "");
993
994 offset = LLVMBuildExtractElement(gallivm->builder, offset,
995 loop_state.counter, "");
996
997 LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, ssbo_ptr,
998 &offset, 1, "");
999
1000 struct lp_build_if_state ifthen;
1001 LLVMValueRef cond, temp_res;
1002 LLVMValueRef scalar;
1003 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
1004 cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
1005 lp_build_if(&ifthen, gallivm, cond);
1006
1007 if (nir_op == nir_intrinsic_ssbo_atomic_comp_swap || nir_op == nir_intrinsic_shared_atomic_comp_swap) {
1008 LLVMValueRef cas_src_ptr = LLVMBuildExtractElement(gallivm->builder, val2,
1009 loop_state.counter, "");
1010 cas_src_ptr = LLVMBuildBitCast(gallivm->builder, cas_src_ptr, uint_bld->elem_type, "");
1011 scalar = LLVMBuildAtomicCmpXchg(builder, scalar_ptr, value_ptr,
1012 cas_src_ptr,
1013 LLVMAtomicOrderingSequentiallyConsistent,
1014 LLVMAtomicOrderingSequentiallyConsistent,
1015 false);
1016 scalar = LLVMBuildExtractValue(gallivm->builder, scalar, 0, "");
1017 } else {
1018 scalar = LLVMBuildAtomicRMW(builder, op,
1019 scalar_ptr, value_ptr,
1020 LLVMAtomicOrderingSequentiallyConsistent,
1021 false);
1022 }
1023 temp_res = LLVMBuildLoad(builder, atom_res, "");
1024 temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
1025 LLVMBuildStore(builder, temp_res, atom_res);
1026 lp_build_else(&ifthen);
1027 temp_res = LLVMBuildLoad(builder, atom_res, "");
1028 temp_res = LLVMBuildInsertElement(builder, temp_res, lp_build_const_int32(gallivm, 0), loop_state.counter, "");
1029 LLVMBuildStore(builder, temp_res, atom_res);
1030 lp_build_endif(&ifthen);
1031
1032 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
1033 NULL, LLVMIntUGE);
1034 *result = LLVMBuildLoad(builder, atom_res, "");
1035 }
1036
1037 static void emit_barrier(struct lp_build_nir_context *bld_base)
1038 {
1039 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1040 struct gallivm_state * gallivm = bld_base->base.gallivm;
1041
1042 LLVMBasicBlockRef resume = lp_build_insert_new_block(gallivm, "resume");
1043
1044 lp_build_coro_suspend_switch(gallivm, bld->coro, resume, false);
1045 LLVMPositionBuilderAtEnd(gallivm->builder, resume);
1046 }
1047
1048 static LLVMValueRef emit_get_buffer_size(struct lp_build_nir_context *bld_base,
1049 LLVMValueRef index)
1050 {
1051 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1052 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1053 struct lp_build_context *bld_broad = &bld_base->uint_bld;
1054 LLVMValueRef size_ptr = lp_build_array_get(bld_base->base.gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, bld_broad->zero, ""));
1055 return lp_build_broadcast_scalar(bld_broad, size_ptr);
1056 }
1057
1058 static void emit_image_op(struct lp_build_nir_context *bld_base,
1059 struct lp_img_params *params)
1060 {
1061 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1062 params->type = bld_base->base.type;
1063 params->context_ptr = bld->context_ptr;
1064 params->thread_data_ptr = bld->thread_data_ptr;
1065 params->exec_mask = mask_vec(bld_base);
1066 bld->image->emit_op(bld->image,
1067 bld->bld_base.base.gallivm,
1068 params);
1069
1070 }
1071
1072 static void emit_image_size(struct lp_build_nir_context *bld_base,
1073 struct lp_sampler_size_query_params *params)
1074 {
1075 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1076
1077 params->int_type = bld_base->int_bld.type;
1078 params->context_ptr = bld->context_ptr;
1079
1080 bld->image->emit_size_query(bld->image,
1081 bld->bld_base.base.gallivm,
1082 params);
1083
1084 }
1085
1086 static void init_var_slots(struct lp_build_nir_context *bld_base,
1087 nir_variable *var, unsigned sc)
1088 {
1089 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1090 unsigned slots = glsl_count_attribute_slots(var->type, false) * 4;
1091
1092 for (unsigned comp = sc; comp < slots + sc; comp++) {
1093 unsigned this_loc = var->data.driver_location + (comp / 4);
1094 unsigned this_chan = comp % 4;
1095
1096 if (!bld->outputs[this_loc][this_chan])
1097 bld->outputs[this_loc][this_chan] = lp_build_alloca(bld_base->base.gallivm,
1098 bld_base->base.vec_type, "output");
1099 }
1100 }
1101
1102 static void emit_var_decl(struct lp_build_nir_context *bld_base,
1103 nir_variable *var)
1104 {
1105 unsigned sc = var->data.location_frac;
1106 switch (var->data.mode) {
1107 case nir_var_shader_out: {
1108 if (bld_base->shader->info.stage == MESA_SHADER_FRAGMENT) {
1109 if (var->data.location == FRAG_RESULT_STENCIL)
1110 sc = 1;
1111 else if (var->data.location == FRAG_RESULT_DEPTH)
1112 sc = 2;
1113 }
1114 init_var_slots(bld_base, var, sc);
1115 break;
1116 }
1117 default:
1118 break;
1119 }
1120 }
1121
1122 static void emit_tex(struct lp_build_nir_context *bld_base,
1123 struct lp_sampler_params *params)
1124 {
1125 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1126
1127 params->type = bld_base->base.type;
1128 params->context_ptr = bld->context_ptr;
1129 params->thread_data_ptr = bld->thread_data_ptr;
1130
1131 bld->sampler->emit_tex_sample(bld->sampler,
1132 bld->bld_base.base.gallivm,
1133 params);
1134 }
1135
1136 static void emit_tex_size(struct lp_build_nir_context *bld_base,
1137 struct lp_sampler_size_query_params *params)
1138 {
1139 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1140
1141 params->int_type = bld_base->int_bld.type;
1142 params->context_ptr = bld->context_ptr;
1143
1144 bld->sampler->emit_size_query(bld->sampler,
1145 bld->bld_base.base.gallivm,
1146 params);
1147 }
1148
1149 static void emit_sysval_intrin(struct lp_build_nir_context *bld_base,
1150 nir_intrinsic_instr *instr,
1151 LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
1152 {
1153 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1154 struct gallivm_state *gallivm = bld_base->base.gallivm;
1155 switch (instr->intrinsic) {
1156 case nir_intrinsic_load_instance_id:
1157 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
1158 break;
1159 case nir_intrinsic_load_base_instance:
1160 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.base_instance);
1161 break;
1162 case nir_intrinsic_load_base_vertex:
1163 result[0] = bld->system_values.basevertex;
1164 break;
1165 case nir_intrinsic_load_vertex_id:
1166 result[0] = bld->system_values.vertex_id;
1167 break;
1168 case nir_intrinsic_load_primitive_id:
1169 result[0] = bld->system_values.prim_id;
1170 break;
1171 case nir_intrinsic_load_work_group_id:
1172 for (unsigned i = 0; i < 3; i++)
1173 result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.block_id, lp_build_const_int32(gallivm, i), ""));
1174 break;
1175 case nir_intrinsic_load_local_invocation_id:
1176 for (unsigned i = 0; i < 3; i++)
1177 result[i] = LLVMBuildExtractValue(gallivm->builder, bld->system_values.thread_id, i, "");
1178 break;
1179 case nir_intrinsic_load_num_work_groups:
1180 for (unsigned i = 0; i < 3; i++)
1181 result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.grid_size, lp_build_const_int32(gallivm, i), ""));
1182 break;
1183 case nir_intrinsic_load_invocation_id:
1184 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.invocation_id);
1185 break;
1186 case nir_intrinsic_load_front_face:
1187 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.front_facing);
1188 break;
1189 case nir_intrinsic_load_draw_id:
1190 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.draw_id);
1191 break;
1192 default:
1193 break;
1194 case nir_intrinsic_load_local_group_size:
1195 for (unsigned i = 0; i < 3; i++)
1196 result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.block_size, lp_build_const_int32(gallivm, i), ""));
1197 break;
1198 case nir_intrinsic_load_work_dim:
1199 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.work_dim);
1200 break;
1201 }
1202 }
1203
1204 static void bgnloop(struct lp_build_nir_context *bld_base)
1205 {
1206 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1207 lp_exec_bgnloop(&bld->exec_mask, true);
1208 }
1209
1210 static void endloop(struct lp_build_nir_context *bld_base)
1211 {
1212 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1213 lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1214 }
1215
1216 static void if_cond(struct lp_build_nir_context *bld_base, LLVMValueRef cond)
1217 {
1218 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1219 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1220 lp_exec_mask_cond_push(&bld->exec_mask, LLVMBuildBitCast(builder, cond, bld_base->base.int_vec_type, ""));
1221 }
1222
1223 static void else_stmt(struct lp_build_nir_context *bld_base)
1224 {
1225 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1226 lp_exec_mask_cond_invert(&bld->exec_mask);
1227 }
1228
1229 static void endif_stmt(struct lp_build_nir_context *bld_base)
1230 {
1231 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1232 lp_exec_mask_cond_pop(&bld->exec_mask);
1233 }
1234
1235 static void break_stmt(struct lp_build_nir_context *bld_base)
1236 {
1237 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1238
1239 lp_exec_break(&bld->exec_mask, NULL, false);
1240 }
1241
1242 static void continue_stmt(struct lp_build_nir_context *bld_base)
1243 {
1244 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1245 lp_exec_continue(&bld->exec_mask);
1246 }
1247
1248 static void discard(struct lp_build_nir_context *bld_base, LLVMValueRef cond)
1249 {
1250 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1251 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1252 LLVMValueRef mask;
1253
1254 if (!cond) {
1255 if (bld->exec_mask.has_mask) {
1256 mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1257 } else {
1258 mask = LLVMConstNull(bld->bld_base.base.int_vec_type);
1259 }
1260 } else {
1261 mask = LLVMBuildNot(builder, cond, "");
1262 if (bld->exec_mask.has_mask) {
1263 LLVMValueRef invmask;
1264 invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1265 mask = LLVMBuildOr(builder, mask, invmask, "");
1266 }
1267 }
1268 lp_build_mask_update(bld->mask, mask);
1269 }
1270
1271 static void
1272 increment_vec_ptr_by_mask(struct lp_build_nir_context * bld_base,
1273 LLVMValueRef ptr,
1274 LLVMValueRef mask)
1275 {
1276 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1277 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
1278
1279 current_vec = LLVMBuildSub(builder, current_vec, mask, "");
1280
1281 LLVMBuildStore(builder, current_vec, ptr);
1282 }
1283
1284 static void
1285 clear_uint_vec_ptr_from_mask(struct lp_build_nir_context * bld_base,
1286 LLVMValueRef ptr,
1287 LLVMValueRef mask)
1288 {
1289 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1290 LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
1291
1292 current_vec = lp_build_select(&bld_base->uint_bld,
1293 mask,
1294 bld_base->uint_bld.zero,
1295 current_vec);
1296
1297 LLVMBuildStore(builder, current_vec, ptr);
1298 }
1299
1300 static LLVMValueRef
1301 clamp_mask_to_max_output_vertices(struct lp_build_nir_soa_context * bld,
1302 LLVMValueRef current_mask_vec,
1303 LLVMValueRef total_emitted_vertices_vec)
1304 {
1305 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1306 struct lp_build_context *int_bld = &bld->bld_base.int_bld;
1307 LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
1308 total_emitted_vertices_vec,
1309 bld->max_output_vertices_vec);
1310
1311 return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
1312 }
1313
1314 static void emit_vertex(struct lp_build_nir_context *bld_base, uint32_t stream_id)
1315 {
1316 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1317 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1318
1319 assert(bld->gs_iface->emit_vertex);
1320 LLVMValueRef total_emitted_vertices_vec =
1321 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
1322 LLVMValueRef mask = mask_vec(bld_base);
1323 mask = clamp_mask_to_max_output_vertices(bld, mask,
1324 total_emitted_vertices_vec);
1325 bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base.base,
1326 bld->outputs,
1327 total_emitted_vertices_vec,
1328 lp_build_const_int_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, stream_id));
1329
1330 increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr,
1331 mask);
1332 increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr,
1333 mask);
1334 }
1335
1336 static void
1337 end_primitive_masked(struct lp_build_nir_context * bld_base,
1338 LLVMValueRef mask)
1339 {
1340 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1341 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1342
1343 struct lp_build_context *uint_bld = &bld_base->uint_bld;
1344 LLVMValueRef emitted_vertices_vec =
1345 LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr, "");
1346 LLVMValueRef emitted_prims_vec =
1347 LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr, "");
1348 LLVMValueRef total_emitted_vertices_vec =
1349 LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr, "");
1350
1351 LLVMValueRef emitted_mask = lp_build_cmp(uint_bld,
1352 PIPE_FUNC_NOTEQUAL,
1353 emitted_vertices_vec,
1354 uint_bld->zero);
1355 mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
1356 bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base.base,
1357 total_emitted_vertices_vec,
1358 emitted_vertices_vec, emitted_prims_vec, mask_vec(bld_base));
1359 increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr,
1360 mask);
1361 clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr,
1362 mask);
1363 }
1364
1365 static void end_primitive(struct lp_build_nir_context *bld_base, uint32_t stream_id)
1366 {
1367 struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1368
1369 assert(bld->gs_iface->end_primitive);
1370
1371 LLVMValueRef mask = mask_vec(bld_base);
1372 end_primitive_masked(bld_base, mask);
1373 }
1374
1375 static void
1376 emit_prologue(struct lp_build_nir_soa_context *bld)
1377 {
1378 struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
1379 if (bld->indirects & nir_var_shader_in && !bld->gs_iface) {
1380 uint32_t num_inputs = util_bitcount64(bld->bld_base.shader->info.inputs_read);
1381 unsigned index, chan;
1382 LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1383 LLVMValueRef array_size = lp_build_const_int32(gallivm, num_inputs * 4);
1384 bld->inputs_array = lp_build_array_alloca(gallivm,
1385 vec_type, array_size,
1386 "input_array");
1387
1388 for (index = 0; index < num_inputs; ++index) {
1389 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1390 LLVMValueRef lindex =
1391 lp_build_const_int32(gallivm, index * 4 + chan);
1392 LLVMValueRef input_ptr =
1393 LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1394 &lindex, 1, "");
1395 LLVMValueRef value = bld->inputs[index][chan];
1396 if (value)
1397 LLVMBuildStore(gallivm->builder, value, input_ptr);
1398 }
1399 }
1400 }
1401 }
1402
1403 static void emit_vote(struct lp_build_nir_context *bld_base, LLVMValueRef src, nir_intrinsic_instr *instr, LLVMValueRef result[4])
1404 {
1405 struct gallivm_state * gallivm = bld_base->base.gallivm;
1406 LLVMBuilderRef builder = gallivm->builder;
1407
1408 LLVMValueRef exec_mask = mask_vec(bld_base);
1409 struct lp_build_loop_state loop_state;
1410
1411 LLVMValueRef outer_cond = LLVMBuildICmp(builder, LLVMIntNE, exec_mask, bld_base->uint_bld.zero, "");
1412
1413 LLVMValueRef res_store = lp_build_alloca(gallivm, bld_base->int_bld.elem_type, "");
1414 LLVMValueRef init_val;
1415 if (instr->intrinsic == nir_intrinsic_vote_ieq) {
1416 /* for equal we unfortunately have to loop and find the first valid one. */
1417 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1418 LLVMValueRef if_cond = LLVMBuildExtractElement(gallivm->builder, outer_cond, loop_state.counter, "");
1419
1420 struct lp_build_if_state ifthen;
1421 lp_build_if(&ifthen, gallivm, if_cond);
1422 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, src,
1423 loop_state.counter, "");
1424 LLVMBuildStore(builder, value_ptr, res_store);
1425 lp_build_endif(&ifthen);
1426 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, bld_base->uint_bld.type.length),
1427 NULL, LLVMIntUGE);
1428 lp_build_print_value(gallivm, "init_val is ", LLVMBuildLoad(builder, res_store, ""));
1429 init_val = LLVMBuildLoad(builder, res_store, "");
1430 } else {
1431 LLVMBuildStore(builder, lp_build_const_int32(gallivm, instr->intrinsic == nir_intrinsic_vote_any ? 0 : -1), res_store);
1432 }
1433
1434 LLVMValueRef res;
1435 lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1436 LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, src,
1437 loop_state.counter, "");
1438 struct lp_build_if_state ifthen;
1439 LLVMValueRef if_cond;
1440 if_cond = LLVMBuildExtractElement(gallivm->builder, outer_cond, loop_state.counter, "");
1441
1442 lp_build_if(&ifthen, gallivm, if_cond);
1443 res = LLVMBuildLoad(builder, res_store, "");
1444
1445 if (instr->intrinsic == nir_intrinsic_vote_ieq) {
1446 LLVMValueRef tmp = LLVMBuildICmp(builder, LLVMIntEQ, init_val, value_ptr, "");
1447 tmp = LLVMBuildSExt(builder, tmp, bld_base->uint_bld.elem_type, "");
1448 res = LLVMBuildOr(builder, res, tmp, "");
1449 } else if (instr->intrinsic == nir_intrinsic_vote_any)
1450 res = LLVMBuildOr(builder, res, value_ptr, "");
1451 else
1452 res = LLVMBuildAnd(builder, res, value_ptr, "");
1453 LLVMBuildStore(builder, res, res_store);
1454 lp_build_endif(&ifthen);
1455 lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, bld_base->uint_bld.type.length),
1456 NULL, LLVMIntUGE);
1457 result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildLoad(builder, res_store, ""));
1458 }
1459
1460 void lp_build_nir_soa(struct gallivm_state *gallivm,
1461 struct nir_shader *shader,
1462 const struct lp_build_tgsi_params *params,
1463 LLVMValueRef (*outputs)[4])
1464 {
1465 struct lp_build_nir_soa_context bld;
1466 struct lp_type type = params->type;
1467 struct lp_type res_type;
1468
1469 assert(type.length <= LP_MAX_VECTOR_LENGTH);
1470 memset(&res_type, 0, sizeof res_type);
1471 res_type.width = type.width;
1472 res_type.length = type.length;
1473 res_type.sign = 1;
1474
1475 /* Setup build context */
1476 memset(&bld, 0, sizeof bld);
1477 lp_build_context_init(&bld.bld_base.base, gallivm, type);
1478 lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1479 lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1480 lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1481 lp_build_context_init(&bld.uint_elem_bld, gallivm, lp_elem_type(lp_uint_type(type)));
1482 {
1483 struct lp_type dbl_type;
1484 dbl_type = type;
1485 dbl_type.width *= 2;
1486 lp_build_context_init(&bld.bld_base.dbl_bld, gallivm, dbl_type);
1487 }
1488 {
1489 struct lp_type uint64_type;
1490 uint64_type = lp_uint_type(type);
1491 uint64_type.width *= 2;
1492 lp_build_context_init(&bld.bld_base.uint64_bld, gallivm, uint64_type);
1493 }
1494 {
1495 struct lp_type int64_type;
1496 int64_type = lp_int_type(type);
1497 int64_type.width *= 2;
1498 lp_build_context_init(&bld.bld_base.int64_bld, gallivm, int64_type);
1499 }
1500 {
1501 struct lp_type uint16_type;
1502 uint16_type = lp_uint_type(type);
1503 uint16_type.width /= 2;
1504 lp_build_context_init(&bld.bld_base.uint16_bld, gallivm, uint16_type);
1505 }
1506 {
1507 struct lp_type int16_type;
1508 int16_type = lp_int_type(type);
1509 int16_type.width /= 2;
1510 lp_build_context_init(&bld.bld_base.int16_bld, gallivm, int16_type);
1511 }
1512 {
1513 struct lp_type uint8_type;
1514 uint8_type = lp_uint_type(type);
1515 uint8_type.width /= 4;
1516 lp_build_context_init(&bld.bld_base.uint8_bld, gallivm, uint8_type);
1517 }
1518 {
1519 struct lp_type int8_type;
1520 int8_type = lp_int_type(type);
1521 int8_type.width /= 4;
1522 lp_build_context_init(&bld.bld_base.int8_bld, gallivm, int8_type);
1523 }
1524 bld.bld_base.load_var = emit_load_var;
1525 bld.bld_base.store_var = emit_store_var;
1526 bld.bld_base.load_reg = emit_load_reg;
1527 bld.bld_base.store_reg = emit_store_reg;
1528 bld.bld_base.emit_var_decl = emit_var_decl;
1529 bld.bld_base.load_ubo = emit_load_ubo;
1530 bld.bld_base.load_kernel_arg = emit_load_kernel_arg;
1531 bld.bld_base.load_global = emit_load_global;
1532 bld.bld_base.store_global = emit_store_global;
1533 bld.bld_base.atomic_global = emit_atomic_global;
1534 bld.bld_base.tex = emit_tex;
1535 bld.bld_base.tex_size = emit_tex_size;
1536 bld.bld_base.bgnloop = bgnloop;
1537 bld.bld_base.endloop = endloop;
1538 bld.bld_base.if_cond = if_cond;
1539 bld.bld_base.else_stmt = else_stmt;
1540 bld.bld_base.endif_stmt = endif_stmt;
1541 bld.bld_base.break_stmt = break_stmt;
1542 bld.bld_base.continue_stmt = continue_stmt;
1543 bld.bld_base.sysval_intrin = emit_sysval_intrin;
1544 bld.bld_base.discard = discard;
1545 bld.bld_base.emit_vertex = emit_vertex;
1546 bld.bld_base.end_primitive = end_primitive;
1547 bld.bld_base.load_mem = emit_load_mem;
1548 bld.bld_base.store_mem = emit_store_mem;
1549 bld.bld_base.get_buffer_size = emit_get_buffer_size;
1550 bld.bld_base.atomic_mem = emit_atomic_mem;
1551 bld.bld_base.barrier = emit_barrier;
1552 bld.bld_base.image_op = emit_image_op;
1553 bld.bld_base.image_size = emit_image_size;
1554 bld.bld_base.vote = emit_vote;
1555
1556 bld.mask = params->mask;
1557 bld.inputs = params->inputs;
1558 bld.outputs = outputs;
1559 bld.consts_ptr = params->consts_ptr;
1560 bld.const_sizes_ptr = params->const_sizes_ptr;
1561 bld.ssbo_ptr = params->ssbo_ptr;
1562 bld.ssbo_sizes_ptr = params->ssbo_sizes_ptr;
1563 bld.sampler = params->sampler;
1564 // bld.bld_base.info = params->info;
1565
1566 bld.context_ptr = params->context_ptr;
1567 bld.thread_data_ptr = params->thread_data_ptr;
1568 bld.image = params->image;
1569 bld.shared_ptr = params->shared_ptr;
1570 bld.coro = params->coro;
1571 bld.kernel_args_ptr = params->kernel_args;
1572 bld.indirects = 0;
1573 if (params->info->indirect_files & (1 << TGSI_FILE_INPUT))
1574 bld.indirects |= nir_var_shader_in;
1575
1576 bld.gs_iface = params->gs_iface;
1577 if (bld.gs_iface) {
1578 struct lp_build_context *uint_bld = &bld.bld_base.uint_bld;
1579
1580 bld.max_output_vertices_vec = lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
1581 shader->info.gs.vertices_out);
1582 bld.emitted_prims_vec_ptr =
1583 lp_build_alloca(gallivm, uint_bld->vec_type, "emitted_prims_ptr");
1584 bld.emitted_vertices_vec_ptr =
1585 lp_build_alloca(gallivm, uint_bld->vec_type, "emitted_vertices_ptr");
1586 bld.total_emitted_vertices_vec_ptr =
1587 lp_build_alloca(gallivm, uint_bld->vec_type, "total_emitted_vertices_ptr");
1588 }
1589 lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
1590
1591 bld.system_values = *params->system_values;
1592
1593 bld.bld_base.shader = shader;
1594
1595 emit_prologue(&bld);
1596 lp_build_nir_llvm(&bld.bld_base, shader);
1597
1598 if (bld.gs_iface) {
1599 LLVMBuilderRef builder = bld.bld_base.base.gallivm->builder;
1600 LLVMValueRef total_emitted_vertices_vec;
1601 LLVMValueRef emitted_prims_vec;
1602 end_primitive_masked(&bld.bld_base, lp_build_mask_value(bld.mask));
1603 total_emitted_vertices_vec =
1604 LLVMBuildLoad(builder, bld.total_emitted_vertices_vec_ptr, "");
1605 emitted_prims_vec =
1606 LLVMBuildLoad(builder, bld.emitted_prims_vec_ptr, "");
1607
1608 bld.gs_iface->gs_epilogue(bld.gs_iface,
1609 total_emitted_vertices_vec,
1610 emitted_prims_vec, 0);
1611 }
1612 lp_exec_mask_fini(&bld.exec_mask);
1613 }