radeonsi: replace TGSI_SEMANTIC with VARYING_SLOT and FRAG_RESULT
[mesa.git] / src / gallium / drivers / radeonsi / si_shader_llvm_ps.c
1 /*
2 * Copyright 2020 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 #include "si_pipe.h"
26 #include "si_shader_internal.h"
27 #include "sid.h"
28
29 LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
30 {
31 return si_unpack_param(ctx, ctx->args.ancillary, 8, 4);
32 }
33
34 static LLVMValueRef load_sample_mask_in(struct ac_shader_abi *abi)
35 {
36 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
37 return ac_to_integer(&ctx->ac, ac_get_arg(&ctx->ac, ctx->args.sample_coverage));
38 }
39
40 static LLVMValueRef load_sample_position(struct ac_shader_abi *abi, LLVMValueRef sample_id)
41 {
42 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
43 LLVMValueRef desc = ac_get_arg(&ctx->ac, ctx->rw_buffers);
44 LLVMValueRef buf_index = LLVMConstInt(ctx->ac.i32, SI_PS_CONST_SAMPLE_POSITIONS, 0);
45 LLVMValueRef resource = ac_build_load_to_sgpr(&ctx->ac, desc, buf_index);
46
47 /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
48 LLVMValueRef offset0 =
49 LLVMBuildMul(ctx->ac.builder, sample_id, LLVMConstInt(ctx->ac.i32, 8, 0), "");
50 LLVMValueRef offset1 =
51 LLVMBuildAdd(ctx->ac.builder, offset0, LLVMConstInt(ctx->ac.i32, 4, 0), "");
52
53 LLVMValueRef pos[4] = {si_buffer_load_const(ctx, resource, offset0),
54 si_buffer_load_const(ctx, resource, offset1),
55 LLVMConstReal(ctx->ac.f32, 0), LLVMConstReal(ctx->ac.f32, 0)};
56
57 return ac_build_gather_values(&ctx->ac, pos, 4);
58 }
59
60 static LLVMValueRef si_nir_emit_fbfetch(struct ac_shader_abi *abi)
61 {
62 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
63 struct ac_image_args args = {};
64 LLVMValueRef ptr, image, fmask;
65
66 /* Ignore src0, because KHR_blend_func_extended disallows multiple render
67 * targets.
68 */
69
70 /* Load the image descriptor. */
71 STATIC_ASSERT(SI_PS_IMAGE_COLORBUF0 % 2 == 0);
72 ptr = ac_get_arg(&ctx->ac, ctx->rw_buffers);
73 ptr =
74 LLVMBuildPointerCast(ctx->ac.builder, ptr, ac_array_in_const32_addr_space(ctx->ac.v8i32), "");
75 image =
76 ac_build_load_to_sgpr(&ctx->ac, ptr, LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0 / 2, 0));
77
78 unsigned chan = 0;
79
80 args.coords[chan++] = si_unpack_param(ctx, ctx->pos_fixed_pt, 0, 16);
81
82 if (!ctx->shader->key.mono.u.ps.fbfetch_is_1D)
83 args.coords[chan++] = si_unpack_param(ctx, ctx->pos_fixed_pt, 16, 16);
84
85 /* Get the current render target layer index. */
86 if (ctx->shader->key.mono.u.ps.fbfetch_layered)
87 args.coords[chan++] = si_unpack_param(ctx, ctx->args.ancillary, 16, 11);
88
89 if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
90 args.coords[chan++] = si_get_sample_id(ctx);
91
92 if (ctx->shader->key.mono.u.ps.fbfetch_msaa && !(ctx->screen->debug_flags & DBG(NO_FMASK))) {
93 fmask = ac_build_load_to_sgpr(&ctx->ac, ptr,
94 LLVMConstInt(ctx->ac.i32, SI_PS_IMAGE_COLORBUF0_FMASK / 2, 0));
95
96 ac_apply_fmask_to_sample(&ctx->ac, fmask, args.coords,
97 ctx->shader->key.mono.u.ps.fbfetch_layered);
98 }
99
100 args.opcode = ac_image_load;
101 args.resource = image;
102 args.dmask = 0xf;
103 args.attributes = AC_FUNC_ATTR_READNONE;
104
105 if (ctx->shader->key.mono.u.ps.fbfetch_msaa)
106 args.dim =
107 ctx->shader->key.mono.u.ps.fbfetch_layered ? ac_image_2darraymsaa : ac_image_2dmsaa;
108 else if (ctx->shader->key.mono.u.ps.fbfetch_is_1D)
109 args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ? ac_image_1darray : ac_image_1d;
110 else
111 args.dim = ctx->shader->key.mono.u.ps.fbfetch_layered ? ac_image_2darray : ac_image_2d;
112
113 return ac_build_image_opcode(&ctx->ac, &args);
114 }
115
116 static LLVMValueRef si_build_fs_interp(struct si_shader_context *ctx, unsigned attr_index,
117 unsigned chan, LLVMValueRef prim_mask, LLVMValueRef i,
118 LLVMValueRef j)
119 {
120 if (i || j) {
121 return ac_build_fs_interp(&ctx->ac, LLVMConstInt(ctx->ac.i32, chan, 0),
122 LLVMConstInt(ctx->ac.i32, attr_index, 0), prim_mask, i, j);
123 }
124 return ac_build_fs_interp_mov(&ctx->ac, LLVMConstInt(ctx->ac.i32, 2, 0), /* P0 */
125 LLVMConstInt(ctx->ac.i32, chan, 0),
126 LLVMConstInt(ctx->ac.i32, attr_index, 0), prim_mask);
127 }
128
129 /**
130 * Interpolate a fragment shader input.
131 *
132 * @param ctx context
133 * @param input_index index of the input in hardware
134 * @param semantic_index semantic index
135 * @param num_interp_inputs number of all interpolated inputs (= BCOLOR offset)
136 * @param colors_read_mask color components read (4 bits for each color, 8 bits in total)
137 * @param interp_param interpolation weights (i,j)
138 * @param prim_mask SI_PARAM_PRIM_MASK
139 * @param face SI_PARAM_FRONT_FACE
140 * @param result the return value (4 components)
141 */
142 static void interp_fs_color(struct si_shader_context *ctx, unsigned input_index,
143 unsigned semantic_index, unsigned num_interp_inputs,
144 unsigned colors_read_mask, LLVMValueRef interp_param,
145 LLVMValueRef prim_mask, LLVMValueRef face, LLVMValueRef result[4])
146 {
147 LLVMValueRef i = NULL, j = NULL;
148 unsigned chan;
149
150 /* fs.constant returns the param from the middle vertex, so it's not
151 * really useful for flat shading. It's meant to be used for custom
152 * interpolation (but the intrinsic can't fetch from the other two
153 * vertices).
154 *
155 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
156 * to do the right thing. The only reason we use fs.constant is that
157 * fs.interp cannot be used on integers, because they can be equal
158 * to NaN.
159 *
160 * When interp is false we will use fs.constant or for newer llvm,
161 * amdgcn.interp.mov.
162 */
163 bool interp = interp_param != NULL;
164
165 if (interp) {
166 interp_param =
167 LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2f32, "");
168
169 i = LLVMBuildExtractElement(ctx->ac.builder, interp_param, ctx->ac.i32_0, "");
170 j = LLVMBuildExtractElement(ctx->ac.builder, interp_param, ctx->ac.i32_1, "");
171 }
172
173 if (ctx->shader->key.part.ps.prolog.color_two_side) {
174 LLVMValueRef is_face_positive;
175
176 /* If BCOLOR0 is used, BCOLOR1 is at offset "num_inputs + 1",
177 * otherwise it's at offset "num_inputs".
178 */
179 unsigned back_attr_offset = num_interp_inputs;
180 if (semantic_index == 1 && colors_read_mask & 0xf)
181 back_attr_offset += 1;
182
183 is_face_positive = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, face, ctx->ac.i32_0, "");
184
185 for (chan = 0; chan < 4; chan++) {
186 LLVMValueRef front, back;
187
188 front = si_build_fs_interp(ctx, input_index, chan, prim_mask, i, j);
189 back = si_build_fs_interp(ctx, back_attr_offset, chan, prim_mask, i, j);
190
191 result[chan] = LLVMBuildSelect(ctx->ac.builder, is_face_positive, front, back, "");
192 }
193 } else {
194 for (chan = 0; chan < 4; chan++) {
195 result[chan] = si_build_fs_interp(ctx, input_index, chan, prim_mask, i, j);
196 }
197 }
198 }
199
200 static void si_alpha_test(struct si_shader_context *ctx, LLVMValueRef alpha)
201 {
202 if (ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_NEVER) {
203 static LLVMRealPredicate cond_map[PIPE_FUNC_ALWAYS + 1] = {
204 [PIPE_FUNC_LESS] = LLVMRealOLT, [PIPE_FUNC_EQUAL] = LLVMRealOEQ,
205 [PIPE_FUNC_LEQUAL] = LLVMRealOLE, [PIPE_FUNC_GREATER] = LLVMRealOGT,
206 [PIPE_FUNC_NOTEQUAL] = LLVMRealONE, [PIPE_FUNC_GEQUAL] = LLVMRealOGE,
207 };
208 LLVMRealPredicate cond = cond_map[ctx->shader->key.part.ps.epilog.alpha_func];
209 assert(cond);
210
211 LLVMValueRef alpha_ref = LLVMGetParam(ctx->main_fn, SI_PARAM_ALPHA_REF);
212 LLVMValueRef alpha_pass = LLVMBuildFCmp(ctx->ac.builder, cond, alpha, alpha_ref, "");
213 ac_build_kill_if_false(&ctx->ac, alpha_pass);
214 } else {
215 ac_build_kill_if_false(&ctx->ac, ctx->ac.i1false);
216 }
217 }
218
219 static LLVMValueRef si_scale_alpha_by_sample_mask(struct si_shader_context *ctx, LLVMValueRef alpha,
220 unsigned samplemask_param)
221 {
222 LLVMValueRef coverage;
223
224 /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
225 coverage = LLVMGetParam(ctx->main_fn, samplemask_param);
226 coverage = ac_to_integer(&ctx->ac, coverage);
227
228 coverage = ac_build_intrinsic(&ctx->ac, "llvm.ctpop.i32", ctx->ac.i32, &coverage, 1,
229 AC_FUNC_ATTR_READNONE);
230
231 coverage = LLVMBuildUIToFP(ctx->ac.builder, coverage, ctx->ac.f32, "");
232
233 coverage = LLVMBuildFMul(ctx->ac.builder, coverage,
234 LLVMConstReal(ctx->ac.f32, 1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
235
236 return LLVMBuildFMul(ctx->ac.builder, alpha, coverage, "");
237 }
238
239 struct si_ps_exports {
240 unsigned num;
241 struct ac_export_args args[10];
242 };
243
244 static void si_export_mrt_z(struct si_shader_context *ctx, LLVMValueRef depth, LLVMValueRef stencil,
245 LLVMValueRef samplemask, struct si_ps_exports *exp)
246 {
247 struct ac_export_args args;
248
249 ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args);
250
251 memcpy(&exp->args[exp->num++], &args, sizeof(args));
252 }
253
254 /* Initialize arguments for the shader export intrinsic */
255 static void si_llvm_init_ps_export_args(struct si_shader_context *ctx, LLVMValueRef *values,
256 unsigned cbuf, unsigned compacted_mrt_index,
257 struct ac_export_args *args)
258 {
259 const struct si_shader_key *key = &ctx->shader->key;
260 unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
261 LLVMValueRef f32undef = LLVMGetUndef(ctx->ac.f32);
262 unsigned spi_shader_col_format;
263 unsigned chan;
264 bool is_int8, is_int10;
265
266 assert(cbuf >= 0 && cbuf < 8);
267
268 spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
269 is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
270 is_int10 = (key->part.ps.epilog.color_is_int10 >> cbuf) & 0x1;
271
272 /* Default is 0xf. Adjusted below depending on the format. */
273 args->enabled_channels = 0xf; /* writemask */
274
275 /* Specify whether the EXEC mask represents the valid mask */
276 args->valid_mask = 0;
277
278 /* Specify whether this is the last export */
279 args->done = 0;
280
281 /* Specify the target we are exporting */
282 args->target = V_008DFC_SQ_EXP_MRT + compacted_mrt_index;
283
284 args->compr = false;
285 args->out[0] = f32undef;
286 args->out[1] = f32undef;
287 args->out[2] = f32undef;
288 args->out[3] = f32undef;
289
290 LLVMValueRef (*packf)(struct ac_llvm_context * ctx, LLVMValueRef args[2]) = NULL;
291 LLVMValueRef (*packi)(struct ac_llvm_context * ctx, LLVMValueRef args[2], unsigned bits,
292 bool hi) = NULL;
293
294 switch (spi_shader_col_format) {
295 case V_028714_SPI_SHADER_ZERO:
296 args->enabled_channels = 0; /* writemask */
297 args->target = V_008DFC_SQ_EXP_NULL;
298 break;
299
300 case V_028714_SPI_SHADER_32_R:
301 args->enabled_channels = 1; /* writemask */
302 args->out[0] = values[0];
303 break;
304
305 case V_028714_SPI_SHADER_32_GR:
306 args->enabled_channels = 0x3; /* writemask */
307 args->out[0] = values[0];
308 args->out[1] = values[1];
309 break;
310
311 case V_028714_SPI_SHADER_32_AR:
312 if (ctx->screen->info.chip_class >= GFX10) {
313 args->enabled_channels = 0x3; /* writemask */
314 args->out[0] = values[0];
315 args->out[1] = values[3];
316 } else {
317 args->enabled_channels = 0x9; /* writemask */
318 args->out[0] = values[0];
319 args->out[3] = values[3];
320 }
321 break;
322
323 case V_028714_SPI_SHADER_FP16_ABGR:
324 packf = ac_build_cvt_pkrtz_f16;
325 break;
326
327 case V_028714_SPI_SHADER_UNORM16_ABGR:
328 packf = ac_build_cvt_pknorm_u16;
329 break;
330
331 case V_028714_SPI_SHADER_SNORM16_ABGR:
332 packf = ac_build_cvt_pknorm_i16;
333 break;
334
335 case V_028714_SPI_SHADER_UINT16_ABGR:
336 packi = ac_build_cvt_pk_u16;
337 break;
338
339 case V_028714_SPI_SHADER_SINT16_ABGR:
340 packi = ac_build_cvt_pk_i16;
341 break;
342
343 case V_028714_SPI_SHADER_32_ABGR:
344 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
345 break;
346 }
347
348 /* Pack f16 or norm_i16/u16. */
349 if (packf) {
350 for (chan = 0; chan < 2; chan++) {
351 LLVMValueRef pack_args[2] = {values[2 * chan], values[2 * chan + 1]};
352 LLVMValueRef packed;
353
354 packed = packf(&ctx->ac, pack_args);
355 args->out[chan] = ac_to_float(&ctx->ac, packed);
356 }
357 args->compr = 1; /* COMPR flag */
358 }
359 /* Pack i16/u16. */
360 if (packi) {
361 for (chan = 0; chan < 2; chan++) {
362 LLVMValueRef pack_args[2] = {ac_to_integer(&ctx->ac, values[2 * chan]),
363 ac_to_integer(&ctx->ac, values[2 * chan + 1])};
364 LLVMValueRef packed;
365
366 packed = packi(&ctx->ac, pack_args, is_int8 ? 8 : is_int10 ? 10 : 16, chan == 1);
367 args->out[chan] = ac_to_float(&ctx->ac, packed);
368 }
369 args->compr = 1; /* COMPR flag */
370 }
371 }
372
373 static bool si_export_mrt_color(struct si_shader_context *ctx, LLVMValueRef *color, unsigned index,
374 unsigned compacted_mrt_index, unsigned samplemask_param,
375 bool is_last, struct si_ps_exports *exp)
376 {
377 int i;
378
379 /* Clamp color */
380 if (ctx->shader->key.part.ps.epilog.clamp_color)
381 for (i = 0; i < 4; i++)
382 color[i] = ac_build_clamp(&ctx->ac, color[i]);
383
384 /* Alpha to one */
385 if (ctx->shader->key.part.ps.epilog.alpha_to_one)
386 color[3] = ctx->ac.f32_1;
387
388 /* Alpha test */
389 if (index == 0 && ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
390 si_alpha_test(ctx, color[3]);
391
392 /* Line & polygon smoothing */
393 if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
394 color[3] = si_scale_alpha_by_sample_mask(ctx, color[3], samplemask_param);
395
396 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
397 if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
398 struct ac_export_args args[8];
399 int c, last = -1;
400
401 assert(compacted_mrt_index == 0);
402
403 /* Get the export arguments, also find out what the last one is. */
404 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
405 si_llvm_init_ps_export_args(ctx, color, c, compacted_mrt_index, &args[c]);
406 if (args[c].enabled_channels) {
407 compacted_mrt_index++;
408 last = c;
409 }
410 }
411 if (last == -1)
412 return false;
413
414 /* Emit all exports. */
415 for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
416 if (is_last && last == c) {
417 args[c].valid_mask = 1; /* whether the EXEC mask is valid */
418 args[c].done = 1; /* DONE bit */
419 } else if (!args[c].enabled_channels)
420 continue; /* unnecessary NULL export */
421
422 memcpy(&exp->args[exp->num++], &args[c], sizeof(args[c]));
423 }
424 } else {
425 struct ac_export_args args;
426
427 /* Export */
428 si_llvm_init_ps_export_args(ctx, color, index, compacted_mrt_index, &args);
429 if (is_last) {
430 args.valid_mask = 1; /* whether the EXEC mask is valid */
431 args.done = 1; /* DONE bit */
432 } else if (!args.enabled_channels)
433 return false; /* unnecessary NULL export */
434
435 memcpy(&exp->args[exp->num++], &args, sizeof(args));
436 }
437 return true;
438 }
439
440 static void si_emit_ps_exports(struct si_shader_context *ctx, struct si_ps_exports *exp)
441 {
442 for (unsigned i = 0; i < exp->num; i++)
443 ac_build_export(&ctx->ac, &exp->args[i]);
444 }
445
446 /**
447 * Return PS outputs in this order:
448 *
449 * v[0:3] = color0.xyzw
450 * v[4:7] = color1.xyzw
451 * ...
452 * vN+0 = Depth
453 * vN+1 = Stencil
454 * vN+2 = SampleMask
455 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
456 *
457 * The alpha-ref SGPR is returned via its original location.
458 */
459 static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi, unsigned max_outputs,
460 LLVMValueRef *addrs)
461 {
462 struct si_shader_context *ctx = si_shader_context_from_abi(abi);
463 struct si_shader *shader = ctx->shader;
464 struct si_shader_info *info = &shader->selector->info;
465 LLVMBuilderRef builder = ctx->ac.builder;
466 unsigned i, j, first_vgpr, vgpr;
467
468 LLVMValueRef color[8][4] = {};
469 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
470 LLVMValueRef ret;
471
472 if (ctx->postponed_kill)
473 ac_build_kill_if_false(&ctx->ac, LLVMBuildLoad(builder, ctx->postponed_kill, ""));
474
475 /* Read the output values. */
476 for (i = 0; i < info->num_outputs; i++) {
477 unsigned semantic = info->output_semantic[i];
478
479 switch (semantic) {
480 case FRAG_RESULT_DEPTH:
481 depth = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
482 break;
483 case FRAG_RESULT_STENCIL:
484 stencil = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
485 break;
486 case FRAG_RESULT_SAMPLE_MASK:
487 samplemask = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
488 break;
489 default:
490 if (semantic >= FRAG_RESULT_DATA0 && semantic <= FRAG_RESULT_DATA7) {
491 unsigned index = semantic - FRAG_RESULT_DATA0;
492
493 for (j = 0; j < 4; j++) {
494 LLVMValueRef ptr = addrs[4 * i + j];
495 LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
496 color[index][j] = result;
497 }
498 } else {
499 fprintf(stderr, "Warning: Unhandled fs output type:%d\n", semantic);
500 }
501 break;
502 }
503 }
504
505 /* Fill the return structure. */
506 ret = ctx->return_value;
507
508 /* Set SGPRs. */
509 ret = LLVMBuildInsertValue(
510 builder, ret, ac_to_integer(&ctx->ac, LLVMGetParam(ctx->main_fn, SI_PARAM_ALPHA_REF)),
511 SI_SGPR_ALPHA_REF, "");
512
513 /* Set VGPRs */
514 first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
515 for (i = 0; i < ARRAY_SIZE(color); i++) {
516 if (!color[i][0])
517 continue;
518
519 for (j = 0; j < 4; j++)
520 ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
521 }
522 if (depth)
523 ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
524 if (stencil)
525 ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
526 if (samplemask)
527 ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");
528
529 /* Add the input sample mask for smoothing at the end. */
530 if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
531 vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
532 ret = LLVMBuildInsertValue(builder, ret, LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE),
533 vgpr++, "");
534
535 ctx->return_value = ret;
536 }
537
538 static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
539 LLVMValueRef param_rw_buffers,
540 struct ac_arg param_pos_fixed_pt)
541 {
542 LLVMBuilderRef builder = ctx->ac.builder;
543 LLVMValueRef slot, desc, offset, row, bit, address[2];
544
545 /* Use the fixed-point gl_FragCoord input.
546 * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
547 * per coordinate to get the repeating effect.
548 */
549 address[0] = si_unpack_param(ctx, param_pos_fixed_pt, 0, 5);
550 address[1] = si_unpack_param(ctx, param_pos_fixed_pt, 16, 5);
551
552 /* Load the buffer descriptor. */
553 slot = LLVMConstInt(ctx->ac.i32, SI_PS_CONST_POLY_STIPPLE, 0);
554 desc = ac_build_load_to_sgpr(&ctx->ac, param_rw_buffers, slot);
555
556 /* The stipple pattern is 32x32, each row has 32 bits. */
557 offset = LLVMBuildMul(builder, address[1], LLVMConstInt(ctx->ac.i32, 4, 0), "");
558 row = si_buffer_load_const(ctx, desc, offset);
559 row = ac_to_integer(&ctx->ac, row);
560 bit = LLVMBuildLShr(builder, row, address[0], "");
561 bit = LLVMBuildTrunc(builder, bit, ctx->ac.i1, "");
562 ac_build_kill_if_false(&ctx->ac, bit);
563 }
564
565 /**
566 * Build the pixel shader prolog function. This handles:
567 * - two-side color selection and interpolation
568 * - overriding interpolation parameters for the API PS
569 * - polygon stippling
570 *
571 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
572 * overriden by other states. (e.g. per-sample interpolation)
573 * Interpolated colors are stored after the preloaded VGPRs.
574 */
575 void si_llvm_build_ps_prolog(struct si_shader_context *ctx, union si_shader_part_key *key)
576 {
577 LLVMValueRef ret, func;
578 int num_returns, i, num_color_channels;
579
580 memset(&ctx->args, 0, sizeof(ctx->args));
581
582 /* Declare inputs. */
583 LLVMTypeRef return_types[AC_MAX_ARGS];
584 num_returns = 0;
585 num_color_channels = util_bitcount(key->ps_prolog.colors_read);
586 assert(key->ps_prolog.num_input_sgprs + key->ps_prolog.num_input_vgprs + num_color_channels <=
587 AC_MAX_ARGS);
588 for (i = 0; i < key->ps_prolog.num_input_sgprs; i++) {
589 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
590 return_types[num_returns++] = ctx->ac.i32;
591 }
592
593 struct ac_arg pos_fixed_pt;
594 struct ac_arg ancillary;
595 struct ac_arg param_sample_mask;
596 for (i = 0; i < key->ps_prolog.num_input_vgprs; i++) {
597 struct ac_arg *arg = NULL;
598 if (i == key->ps_prolog.ancillary_vgpr_index) {
599 arg = &ancillary;
600 } else if (i == key->ps_prolog.ancillary_vgpr_index + 1) {
601 arg = &param_sample_mask;
602 } else if (i == key->ps_prolog.num_input_vgprs - 1) {
603 /* POS_FIXED_PT is always last. */
604 arg = &pos_fixed_pt;
605 }
606 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, arg);
607 return_types[num_returns++] = ctx->ac.f32;
608 }
609
610 /* Declare outputs (same as inputs + add colors if needed) */
611 for (i = 0; i < num_color_channels; i++)
612 return_types[num_returns++] = ctx->ac.f32;
613
614 /* Create the function. */
615 si_llvm_create_func(ctx, "ps_prolog", return_types, num_returns, 0);
616 func = ctx->main_fn;
617
618 /* Copy inputs to outputs. This should be no-op, as the registers match,
619 * but it will prevent the compiler from overwriting them unintentionally.
620 */
621 ret = ctx->return_value;
622 for (i = 0; i < ctx->args.arg_count; i++) {
623 LLVMValueRef p = LLVMGetParam(func, i);
624 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
625 }
626
627 /* Polygon stippling. */
628 if (key->ps_prolog.states.poly_stipple) {
629 LLVMValueRef list = si_prolog_get_rw_buffers(ctx);
630
631 si_llvm_emit_polygon_stipple(ctx, list, pos_fixed_pt);
632 }
633
634 if (key->ps_prolog.states.bc_optimize_for_persp ||
635 key->ps_prolog.states.bc_optimize_for_linear) {
636 unsigned i, base = key->ps_prolog.num_input_sgprs;
637 LLVMValueRef center[2], centroid[2], tmp, bc_optimize;
638
639 /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
640 * The hw doesn't compute CENTROID if the whole wave only
641 * contains fully-covered quads.
642 *
643 * PRIM_MASK is after user SGPRs.
644 */
645 bc_optimize = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
646 bc_optimize =
647 LLVMBuildLShr(ctx->ac.builder, bc_optimize, LLVMConstInt(ctx->ac.i32, 31, 0), "");
648 bc_optimize = LLVMBuildTrunc(ctx->ac.builder, bc_optimize, ctx->ac.i1, "");
649
650 if (key->ps_prolog.states.bc_optimize_for_persp) {
651 /* Read PERSP_CENTER. */
652 for (i = 0; i < 2; i++)
653 center[i] = LLVMGetParam(func, base + 2 + i);
654 /* Read PERSP_CENTROID. */
655 for (i = 0; i < 2; i++)
656 centroid[i] = LLVMGetParam(func, base + 4 + i);
657 /* Select PERSP_CENTROID. */
658 for (i = 0; i < 2; i++) {
659 tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize, center[i], centroid[i], "");
660 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, tmp, base + 4 + i, "");
661 }
662 }
663 if (key->ps_prolog.states.bc_optimize_for_linear) {
664 /* Read LINEAR_CENTER. */
665 for (i = 0; i < 2; i++)
666 center[i] = LLVMGetParam(func, base + 8 + i);
667 /* Read LINEAR_CENTROID. */
668 for (i = 0; i < 2; i++)
669 centroid[i] = LLVMGetParam(func, base + 10 + i);
670 /* Select LINEAR_CENTROID. */
671 for (i = 0; i < 2; i++) {
672 tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize, center[i], centroid[i], "");
673 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, tmp, base + 10 + i, "");
674 }
675 }
676 }
677
678 /* Force per-sample interpolation. */
679 if (key->ps_prolog.states.force_persp_sample_interp) {
680 unsigned i, base = key->ps_prolog.num_input_sgprs;
681 LLVMValueRef persp_sample[2];
682
683 /* Read PERSP_SAMPLE. */
684 for (i = 0; i < 2; i++)
685 persp_sample[i] = LLVMGetParam(func, base + i);
686 /* Overwrite PERSP_CENTER. */
687 for (i = 0; i < 2; i++)
688 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_sample[i], base + 2 + i, "");
689 /* Overwrite PERSP_CENTROID. */
690 for (i = 0; i < 2; i++)
691 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_sample[i], base + 4 + i, "");
692 }
693 if (key->ps_prolog.states.force_linear_sample_interp) {
694 unsigned i, base = key->ps_prolog.num_input_sgprs;
695 LLVMValueRef linear_sample[2];
696
697 /* Read LINEAR_SAMPLE. */
698 for (i = 0; i < 2; i++)
699 linear_sample[i] = LLVMGetParam(func, base + 6 + i);
700 /* Overwrite LINEAR_CENTER. */
701 for (i = 0; i < 2; i++)
702 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_sample[i], base + 8 + i, "");
703 /* Overwrite LINEAR_CENTROID. */
704 for (i = 0; i < 2; i++)
705 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_sample[i], base + 10 + i, "");
706 }
707
708 /* Force center interpolation. */
709 if (key->ps_prolog.states.force_persp_center_interp) {
710 unsigned i, base = key->ps_prolog.num_input_sgprs;
711 LLVMValueRef persp_center[2];
712
713 /* Read PERSP_CENTER. */
714 for (i = 0; i < 2; i++)
715 persp_center[i] = LLVMGetParam(func, base + 2 + i);
716 /* Overwrite PERSP_SAMPLE. */
717 for (i = 0; i < 2; i++)
718 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_center[i], base + i, "");
719 /* Overwrite PERSP_CENTROID. */
720 for (i = 0; i < 2; i++)
721 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_center[i], base + 4 + i, "");
722 }
723 if (key->ps_prolog.states.force_linear_center_interp) {
724 unsigned i, base = key->ps_prolog.num_input_sgprs;
725 LLVMValueRef linear_center[2];
726
727 /* Read LINEAR_CENTER. */
728 for (i = 0; i < 2; i++)
729 linear_center[i] = LLVMGetParam(func, base + 8 + i);
730 /* Overwrite LINEAR_SAMPLE. */
731 for (i = 0; i < 2; i++)
732 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_center[i], base + 6 + i, "");
733 /* Overwrite LINEAR_CENTROID. */
734 for (i = 0; i < 2; i++)
735 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_center[i], base + 10 + i, "");
736 }
737
738 /* Interpolate colors. */
739 unsigned color_out_idx = 0;
740 for (i = 0; i < 2; i++) {
741 unsigned writemask = (key->ps_prolog.colors_read >> (i * 4)) & 0xf;
742 unsigned face_vgpr = key->ps_prolog.num_input_sgprs + key->ps_prolog.face_vgpr_index;
743 LLVMValueRef interp[2], color[4];
744 LLVMValueRef interp_ij = NULL, prim_mask = NULL, face = NULL;
745
746 if (!writemask)
747 continue;
748
749 /* If the interpolation qualifier is not CONSTANT (-1). */
750 if (key->ps_prolog.color_interp_vgpr_index[i] != -1) {
751 unsigned interp_vgpr =
752 key->ps_prolog.num_input_sgprs + key->ps_prolog.color_interp_vgpr_index[i];
753
754 /* Get the (i,j) updated by bc_optimize handling. */
755 interp[0] = LLVMBuildExtractValue(ctx->ac.builder, ret, interp_vgpr, "");
756 interp[1] = LLVMBuildExtractValue(ctx->ac.builder, ret, interp_vgpr + 1, "");
757 interp_ij = ac_build_gather_values(&ctx->ac, interp, 2);
758 }
759
760 /* Use the absolute location of the input. */
761 prim_mask = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
762
763 if (key->ps_prolog.states.color_two_side) {
764 face = LLVMGetParam(func, face_vgpr);
765 face = ac_to_integer(&ctx->ac, face);
766 }
767
768 interp_fs_color(ctx, key->ps_prolog.color_attr_index[i], i, key->ps_prolog.num_interp_inputs,
769 key->ps_prolog.colors_read, interp_ij, prim_mask, face, color);
770
771 while (writemask) {
772 unsigned chan = u_bit_scan(&writemask);
773 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, color[chan],
774 ctx->args.arg_count + color_out_idx++, "");
775 }
776 }
777
778 /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
779 * says:
780 *
781 * "When per-sample shading is active due to the use of a fragment
782 * input qualified by sample or due to the use of the gl_SampleID
783 * or gl_SamplePosition variables, only the bit for the current
784 * sample is set in gl_SampleMaskIn. When state specifies multiple
785 * fragment shader invocations for a given fragment, the sample
786 * mask for any single fragment shader invocation may specify a
787 * subset of the covered samples for the fragment. In this case,
788 * the bit corresponding to each covered sample will be set in
789 * exactly one fragment shader invocation."
790 *
791 * The samplemask loaded by hardware is always the coverage of the
792 * entire pixel/fragment, so mask bits out based on the sample ID.
793 */
794 if (key->ps_prolog.states.samplemask_log_ps_iter) {
795 /* The bit pattern matches that used by fixed function fragment
796 * processing. */
797 static const uint16_t ps_iter_masks[] = {
798 0xffff, /* not used */
799 0x5555, 0x1111, 0x0101, 0x0001,
800 };
801 assert(key->ps_prolog.states.samplemask_log_ps_iter < ARRAY_SIZE(ps_iter_masks));
802
803 uint32_t ps_iter_mask = ps_iter_masks[key->ps_prolog.states.samplemask_log_ps_iter];
804 LLVMValueRef sampleid = si_unpack_param(ctx, ancillary, 8, 4);
805 LLVMValueRef samplemask = ac_get_arg(&ctx->ac, param_sample_mask);
806
807 samplemask = ac_to_integer(&ctx->ac, samplemask);
808 samplemask =
809 LLVMBuildAnd(ctx->ac.builder, samplemask,
810 LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, ps_iter_mask, false),
811 sampleid, ""),
812 "");
813 samplemask = ac_to_float(&ctx->ac, samplemask);
814
815 ret = LLVMBuildInsertValue(ctx->ac.builder, ret, samplemask, param_sample_mask.arg_index, "");
816 }
817
818 /* Tell LLVM to insert WQM instruction sequence when needed. */
819 if (key->ps_prolog.wqm) {
820 LLVMAddTargetDependentFunctionAttr(func, "amdgpu-ps-wqm-outputs", "");
821 }
822
823 si_llvm_build_ret(ctx, ret);
824 }
825
826 /**
827 * Build the pixel shader epilog function. This handles everything that must be
828 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
829 */
830 void si_llvm_build_ps_epilog(struct si_shader_context *ctx, union si_shader_part_key *key)
831 {
832 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
833 int i;
834 struct si_ps_exports exp = {};
835
836 memset(&ctx->args, 0, sizeof(ctx->args));
837
838 /* Declare input SGPRs. */
839 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->rw_buffers);
840 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->bindless_samplers_and_images);
841 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->const_and_shader_buffers);
842 ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->samplers_and_images);
843 si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL, SI_PARAM_ALPHA_REF);
844
845 /* Declare input VGPRs. */
846 unsigned required_num_params =
847 ctx->args.num_sgprs_used + util_bitcount(key->ps_epilog.colors_written) * 4 +
848 key->ps_epilog.writes_z + key->ps_epilog.writes_stencil + key->ps_epilog.writes_samplemask;
849
850 required_num_params =
851 MAX2(required_num_params, ctx->args.num_sgprs_used + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
852
853 while (ctx->args.arg_count < required_num_params)
854 ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
855
856 /* Create the function. */
857 si_llvm_create_func(ctx, "ps_epilog", NULL, 0, 0);
858 /* Disable elimination of unused inputs. */
859 ac_llvm_add_target_dep_function_attr(ctx->main_fn, "InitialPSInputAddr", 0xffffff);
860
861 /* Process colors. */
862 unsigned vgpr = ctx->args.num_sgprs_used;
863 unsigned colors_written = key->ps_epilog.colors_written;
864 int last_color_export = -1;
865
866 /* Find the last color export. */
867 if (!key->ps_epilog.writes_z && !key->ps_epilog.writes_stencil &&
868 !key->ps_epilog.writes_samplemask) {
869 unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;
870
871 /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
872 if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
873 /* Just set this if any of the colorbuffers are enabled. */
874 if (spi_format & ((1ull << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
875 last_color_export = 0;
876 } else {
877 for (i = 0; i < 8; i++)
878 if (colors_written & (1 << i) && (spi_format >> (i * 4)) & 0xf)
879 last_color_export = i;
880 }
881 }
882
883 unsigned num_compacted_mrts = 0;
884 while (colors_written) {
885 LLVMValueRef color[4];
886 int output_index = u_bit_scan(&colors_written);
887
888 for (i = 0; i < 4; i++)
889 color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
890
891 if (si_export_mrt_color(ctx, color, output_index, num_compacted_mrts,
892 ctx->args.arg_count - 1,
893 output_index == last_color_export, &exp))
894 num_compacted_mrts++;
895 }
896
897 /* Process depth, stencil, samplemask. */
898 if (key->ps_epilog.writes_z)
899 depth = LLVMGetParam(ctx->main_fn, vgpr++);
900 if (key->ps_epilog.writes_stencil)
901 stencil = LLVMGetParam(ctx->main_fn, vgpr++);
902 if (key->ps_epilog.writes_samplemask)
903 samplemask = LLVMGetParam(ctx->main_fn, vgpr++);
904
905 if (depth || stencil || samplemask)
906 si_export_mrt_z(ctx, depth, stencil, samplemask, &exp);
907 else if (last_color_export == -1)
908 ac_build_export_null(&ctx->ac);
909
910 if (exp.num)
911 si_emit_ps_exports(ctx, &exp);
912
913 /* Compile. */
914 LLVMBuildRetVoid(ctx->ac.builder);
915 }
916
917 void si_llvm_build_monolithic_ps(struct si_shader_context *ctx, struct si_shader *shader)
918 {
919 LLVMValueRef parts[3];
920 unsigned num_parts = 0, main_index;
921 LLVMValueRef main_fn = ctx->main_fn;
922
923 union si_shader_part_key prolog_key;
924 si_get_ps_prolog_key(shader, &prolog_key, false);
925
926 if (si_need_ps_prolog(&prolog_key)) {
927 si_llvm_build_ps_prolog(ctx, &prolog_key);
928 parts[num_parts++] = ctx->main_fn;
929 }
930
931 main_index = num_parts;
932 parts[num_parts++] = main_fn;
933
934 union si_shader_part_key epilog_key;
935 si_get_ps_epilog_key(shader, &epilog_key);
936 si_llvm_build_ps_epilog(ctx, &epilog_key);
937 parts[num_parts++] = ctx->main_fn;
938
939 si_build_wrapper_function(ctx, parts, num_parts, main_index, 0);
940 }
941
942 void si_llvm_init_ps_callbacks(struct si_shader_context *ctx)
943 {
944 ctx->abi.emit_outputs = si_llvm_return_fs_outputs;
945 ctx->abi.load_sample_position = load_sample_position;
946 ctx->abi.load_sample_mask_in = load_sample_mask_in;
947 ctx->abi.emit_fbfetch = si_nir_emit_fbfetch;
948 }