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
29LLVMValueRef si_get_sample_id(struct si_shader_context *ctx)
30{
31   return si_unpack_param(ctx, ctx->args.ancillary, 8, 4);
32}
33
34static 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
40static 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->internal_bindings);
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
60static 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->internal_bindings);
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
116static 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 */
142static 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
200static 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      if (LLVMTypeOf(alpha) == ctx->ac.f16)
213         alpha_ref = LLVMBuildFPTrunc(ctx->ac.builder, alpha_ref, ctx->ac.f16, "");
214
215      LLVMValueRef alpha_pass = LLVMBuildFCmp(ctx->ac.builder, cond, alpha, alpha_ref, "");
216      ac_build_kill_if_false(&ctx->ac, alpha_pass);
217   } else {
218      ac_build_kill_if_false(&ctx->ac, ctx->ac.i1false);
219   }
220}
221
222static LLVMValueRef si_scale_alpha_by_sample_mask(struct si_shader_context *ctx, LLVMValueRef alpha,
223                                                  unsigned samplemask_param)
224{
225   LLVMValueRef coverage;
226
227   /* alpha = alpha * popcount(coverage) / SI_NUM_SMOOTH_AA_SAMPLES */
228   coverage = LLVMGetParam(ctx->main_fn, samplemask_param);
229   coverage = ac_build_bit_count(&ctx->ac, ac_to_integer(&ctx->ac, coverage));
230   coverage = LLVMBuildUIToFP(ctx->ac.builder, coverage, ctx->ac.f32, "");
231
232   coverage = LLVMBuildFMul(ctx->ac.builder, coverage,
233                            LLVMConstReal(ctx->ac.f32, 1.0 / SI_NUM_SMOOTH_AA_SAMPLES), "");
234
235   if (LLVMTypeOf(alpha) == ctx->ac.f16)
236      coverage = LLVMBuildFPTrunc(ctx->ac.builder, coverage, ctx->ac.f16, "");
237
238   return LLVMBuildFMul(ctx->ac.builder, alpha, coverage, "");
239}
240
241struct si_ps_exports {
242   unsigned num;
243   struct ac_export_args args[10];
244};
245
246static LLVMValueRef pack_two_16bit(struct ac_llvm_context *ctx, LLVMValueRef args[2])
247{
248   LLVMValueRef tmp = ac_build_gather_values(ctx, args, 2);
249   return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2f16, "");
250}
251
252static LLVMValueRef get_color_32bit(struct si_shader_context *ctx, unsigned color_type,
253                                    LLVMValueRef value)
254{
255   switch (color_type) {
256   case SI_TYPE_FLOAT16:
257      return LLVMBuildFPExt(ctx->ac.builder, value, ctx->ac.f32, "");
258   case SI_TYPE_INT16:
259      value = ac_to_integer(&ctx->ac, value);
260      value = LLVMBuildSExt(ctx->ac.builder, value, ctx->ac.i32, "");
261      return ac_to_float(&ctx->ac, value);
262   case SI_TYPE_UINT16:
263      value = ac_to_integer(&ctx->ac, value);
264      value = LLVMBuildZExt(ctx->ac.builder, value, ctx->ac.i32, "");
265      return ac_to_float(&ctx->ac, value);
266   case SI_TYPE_ANY32:
267      return value;
268   }
269   return NULL;
270}
271
272/* Initialize arguments for the shader export intrinsic */
273static void si_llvm_init_ps_export_args(struct si_shader_context *ctx, LLVMValueRef *values,
274                                        unsigned cbuf, unsigned compacted_mrt_index,
275                                        unsigned color_type, struct ac_export_args *args)
276{
277   const struct si_shader_key *key = &ctx->shader->key;
278   unsigned col_formats = key->part.ps.epilog.spi_shader_col_format;
279   LLVMValueRef f32undef = LLVMGetUndef(ctx->ac.f32);
280   unsigned spi_shader_col_format;
281   unsigned chan;
282   bool is_int8, is_int10;
283
284   assert(cbuf < 8);
285
286   spi_shader_col_format = (col_formats >> (cbuf * 4)) & 0xf;
287   is_int8 = (key->part.ps.epilog.color_is_int8 >> cbuf) & 0x1;
288   is_int10 = (key->part.ps.epilog.color_is_int10 >> cbuf) & 0x1;
289
290   /* Default is 0xf. Adjusted below depending on the format. */
291   args->enabled_channels = 0xf; /* writemask */
292
293   /* Specify whether the EXEC mask represents the valid mask */
294   args->valid_mask = 0;
295
296   /* Specify whether this is the last export */
297   args->done = 0;
298
299   /* Specify the target we are exporting */
300   args->target = V_008DFC_SQ_EXP_MRT + compacted_mrt_index;
301
302   args->compr = false;
303   args->out[0] = f32undef;
304   args->out[1] = f32undef;
305   args->out[2] = f32undef;
306   args->out[3] = f32undef;
307
308   LLVMValueRef (*packf)(struct ac_llvm_context * ctx, LLVMValueRef args[2]) = NULL;
309   LLVMValueRef (*packi)(struct ac_llvm_context * ctx, LLVMValueRef args[2], unsigned bits,
310                         bool hi) = NULL;
311
312   switch (spi_shader_col_format) {
313   case V_028714_SPI_SHADER_ZERO:
314      args->enabled_channels = 0; /* writemask */
315      args->target = V_008DFC_SQ_EXP_NULL;
316      break;
317
318   case V_028714_SPI_SHADER_32_R:
319      args->enabled_channels = 1; /* writemask */
320      args->out[0] = get_color_32bit(ctx, color_type, values[0]);
321      break;
322
323   case V_028714_SPI_SHADER_32_GR:
324      args->enabled_channels = 0x3; /* writemask */
325      args->out[0] = get_color_32bit(ctx, color_type, values[0]);
326      args->out[1] = get_color_32bit(ctx, color_type, values[1]);
327      break;
328
329   case V_028714_SPI_SHADER_32_AR:
330      if (ctx->screen->info.chip_class >= GFX10) {
331         args->enabled_channels = 0x3; /* writemask */
332         args->out[0] = get_color_32bit(ctx, color_type, values[0]);
333         args->out[1] = get_color_32bit(ctx, color_type, values[3]);
334      } else {
335         args->enabled_channels = 0x9; /* writemask */
336         args->out[0] = get_color_32bit(ctx, color_type, values[0]);
337         args->out[3] = get_color_32bit(ctx, color_type, values[3]);
338      }
339      break;
340
341   case V_028714_SPI_SHADER_FP16_ABGR:
342      if (color_type != SI_TYPE_ANY32)
343         packf = pack_two_16bit;
344      else
345         packf = ac_build_cvt_pkrtz_f16;
346      break;
347
348   case V_028714_SPI_SHADER_UNORM16_ABGR:
349      if (color_type != SI_TYPE_ANY32)
350         packf = ac_build_cvt_pknorm_u16_f16;
351      else
352         packf = ac_build_cvt_pknorm_u16;
353      break;
354
355   case V_028714_SPI_SHADER_SNORM16_ABGR:
356      if (color_type != SI_TYPE_ANY32)
357         packf = ac_build_cvt_pknorm_i16_f16;
358      else
359         packf = ac_build_cvt_pknorm_i16;
360      break;
361
362   case V_028714_SPI_SHADER_UINT16_ABGR:
363      if (color_type != SI_TYPE_ANY32)
364         packf = pack_two_16bit;
365      else
366         packi = ac_build_cvt_pk_u16;
367      break;
368
369   case V_028714_SPI_SHADER_SINT16_ABGR:
370      if (color_type != SI_TYPE_ANY32)
371         packf = pack_two_16bit;
372      else
373         packi = ac_build_cvt_pk_i16;
374      break;
375
376   case V_028714_SPI_SHADER_32_ABGR:
377      for (unsigned i = 0; i < 4; i++)
378         args->out[i] = get_color_32bit(ctx, color_type, values[i]);
379      break;
380   }
381
382   /* Pack f16 or norm_i16/u16. */
383   if (packf) {
384      for (chan = 0; chan < 2; chan++) {
385         LLVMValueRef pack_args[2] = {values[2 * chan], values[2 * chan + 1]};
386         LLVMValueRef packed;
387
388         packed = packf(&ctx->ac, pack_args);
389         args->out[chan] = ac_to_float(&ctx->ac, packed);
390      }
391      args->compr = 1; /* COMPR flag */
392   }
393   /* Pack i16/u16. */
394   if (packi) {
395      for (chan = 0; chan < 2; chan++) {
396         LLVMValueRef pack_args[2] = {ac_to_integer(&ctx->ac, values[2 * chan]),
397                                      ac_to_integer(&ctx->ac, values[2 * chan + 1])};
398         LLVMValueRef packed;
399
400         packed = packi(&ctx->ac, pack_args, is_int8 ? 8 : is_int10 ? 10 : 16, chan == 1);
401         args->out[chan] = ac_to_float(&ctx->ac, packed);
402      }
403      args->compr = 1; /* COMPR flag */
404   }
405}
406
407static bool si_export_mrt_color(struct si_shader_context *ctx, LLVMValueRef *color, unsigned index,
408                                unsigned compacted_mrt_index, unsigned samplemask_param,
409                                bool is_last, unsigned color_type, struct si_ps_exports *exp)
410{
411   int i;
412
413   /* Clamp color */
414   if (ctx->shader->key.part.ps.epilog.clamp_color)
415      for (i = 0; i < 4; i++)
416         color[i] = ac_build_clamp(&ctx->ac, color[i]);
417
418   /* Alpha to one */
419   if (ctx->shader->key.part.ps.epilog.alpha_to_one)
420      color[3] = LLVMConstReal(LLVMTypeOf(color[0]), 1);
421
422   /* Alpha test */
423   if (index == 0 && ctx->shader->key.part.ps.epilog.alpha_func != PIPE_FUNC_ALWAYS)
424      si_alpha_test(ctx, color[3]);
425
426   /* Line & polygon smoothing */
427   if (ctx->shader->key.part.ps.epilog.poly_line_smoothing)
428      color[3] = si_scale_alpha_by_sample_mask(ctx, color[3], samplemask_param);
429
430   /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
431   if (ctx->shader->key.part.ps.epilog.last_cbuf > 0) {
432      struct ac_export_args args[8];
433      int c, last = -1;
434
435      assert(compacted_mrt_index == 0);
436
437      /* Get the export arguments, also find out what the last one is. */
438      for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
439         si_llvm_init_ps_export_args(ctx, color, c, compacted_mrt_index,
440                                     color_type, &args[c]);
441         if (args[c].enabled_channels) {
442            compacted_mrt_index++;
443            last = c;
444         }
445      }
446      if (last == -1)
447         return false;
448
449      /* Emit all exports. */
450      for (c = 0; c <= ctx->shader->key.part.ps.epilog.last_cbuf; c++) {
451         if (is_last && last == c) {
452            args[c].valid_mask = 1; /* whether the EXEC mask is valid */
453            args[c].done = 1;       /* DONE bit */
454         } else if (!args[c].enabled_channels)
455            continue; /* unnecessary NULL export */
456
457         memcpy(&exp->args[exp->num++], &args[c], sizeof(args[c]));
458      }
459   } else {
460      struct ac_export_args args;
461
462      /* Export */
463      si_llvm_init_ps_export_args(ctx, color, index, compacted_mrt_index,
464                                  color_type, &args);
465      if (is_last) {
466         args.valid_mask = 1; /* whether the EXEC mask is valid */
467         args.done = 1;       /* DONE bit */
468      } else if (!args.enabled_channels)
469         return false; /* unnecessary NULL export */
470
471      memcpy(&exp->args[exp->num++], &args, sizeof(args));
472   }
473   return true;
474}
475
476/**
477 * Return PS outputs in this order:
478 *
479 * v[0:3] = color0.xyzw
480 * v[4:7] = color1.xyzw
481 * ...
482 * vN+0 = Depth
483 * vN+1 = Stencil
484 * vN+2 = SampleMask
485 * vN+3 = SampleMaskIn (used for OpenGL smoothing)
486 *
487 * The alpha-ref SGPR is returned via its original location.
488 */
489static void si_llvm_return_fs_outputs(struct ac_shader_abi *abi)
490{
491   struct si_shader_context *ctx = si_shader_context_from_abi(abi);
492   struct si_shader *shader = ctx->shader;
493   struct si_shader_info *info = &shader->selector->info;
494   LLVMBuilderRef builder = ctx->ac.builder;
495   unsigned i, j, first_vgpr, vgpr;
496   LLVMValueRef *addrs = abi->outputs;
497
498   LLVMValueRef color[8][4] = {};
499   LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
500   LLVMValueRef ret;
501
502   /* Read the output values. */
503   for (i = 0; i < info->num_outputs; i++) {
504      unsigned semantic = info->output_semantic[i];
505
506      switch (semantic) {
507      case FRAG_RESULT_DEPTH:
508         depth = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
509         break;
510      case FRAG_RESULT_STENCIL:
511         stencil = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
512         break;
513      case FRAG_RESULT_SAMPLE_MASK:
514         samplemask = LLVMBuildLoad(builder, addrs[4 * i + 0], "");
515         break;
516      default:
517         if (semantic >= FRAG_RESULT_DATA0 && semantic <= FRAG_RESULT_DATA7) {
518            unsigned index = semantic - FRAG_RESULT_DATA0;
519
520            for (j = 0; j < 4; j++) {
521               LLVMValueRef ptr = addrs[4 * i + j];
522               LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
523               color[index][j] = result;
524            }
525         } else {
526            fprintf(stderr, "Warning: Unhandled fs output type:%d\n", semantic);
527         }
528         break;
529      }
530   }
531
532   /* Fill the return structure. */
533   ret = ctx->return_value;
534
535   /* Set SGPRs. */
536   ret = LLVMBuildInsertValue(
537      builder, ret, ac_to_integer(&ctx->ac, LLVMGetParam(ctx->main_fn, SI_PARAM_ALPHA_REF)),
538      SI_SGPR_ALPHA_REF, "");
539
540   /* Set VGPRs */
541   first_vgpr = vgpr = SI_SGPR_ALPHA_REF + 1;
542   for (i = 0; i < ARRAY_SIZE(color); i++) {
543      if (!color[i][0])
544         continue;
545
546      if (LLVMTypeOf(color[i][0]) == ctx->ac.f16) {
547         for (j = 0; j < 2; j++) {
548            LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, &color[i][j * 2], 2);
549            tmp = LLVMBuildBitCast(builder, tmp, ctx->ac.f32, "");
550            ret = LLVMBuildInsertValue(builder, ret, tmp, vgpr++, "");
551         }
552         vgpr += 2;
553      } else {
554         for (j = 0; j < 4; j++)
555            ret = LLVMBuildInsertValue(builder, ret, color[i][j], vgpr++, "");
556      }
557   }
558   if (depth)
559      ret = LLVMBuildInsertValue(builder, ret, depth, vgpr++, "");
560   if (stencil)
561      ret = LLVMBuildInsertValue(builder, ret, stencil, vgpr++, "");
562   if (samplemask)
563      ret = LLVMBuildInsertValue(builder, ret, samplemask, vgpr++, "");
564
565   /* Add the input sample mask for smoothing at the end. */
566   if (vgpr < first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC)
567      vgpr = first_vgpr + PS_EPILOG_SAMPLEMASK_MIN_LOC;
568   ret = LLVMBuildInsertValue(builder, ret, LLVMGetParam(ctx->main_fn, SI_PARAM_SAMPLE_COVERAGE),
569                              vgpr++, "");
570
571   ctx->return_value = ret;
572}
573
574static void si_llvm_emit_polygon_stipple(struct si_shader_context *ctx,
575                                         LLVMValueRef param_internal_bindings,
576                                         struct ac_arg param_pos_fixed_pt)
577{
578   LLVMBuilderRef builder = ctx->ac.builder;
579   LLVMValueRef slot, desc, offset, row, bit, address[2];
580
581   /* Use the fixed-point gl_FragCoord input.
582    * Since the stipple pattern is 32x32 and it repeats, just get 5 bits
583    * per coordinate to get the repeating effect.
584    */
585   address[0] = si_unpack_param(ctx, param_pos_fixed_pt, 0, 5);
586   address[1] = si_unpack_param(ctx, param_pos_fixed_pt, 16, 5);
587
588   /* Load the buffer descriptor. */
589   slot = LLVMConstInt(ctx->ac.i32, SI_PS_CONST_POLY_STIPPLE, 0);
590   desc = ac_build_load_to_sgpr(&ctx->ac, param_internal_bindings, slot);
591
592   /* The stipple pattern is 32x32, each row has 32 bits. */
593   offset = LLVMBuildMul(builder, address[1], LLVMConstInt(ctx->ac.i32, 4, 0), "");
594   row = si_buffer_load_const(ctx, desc, offset);
595   row = ac_to_integer(&ctx->ac, row);
596   bit = LLVMBuildLShr(builder, row, address[0], "");
597   bit = LLVMBuildTrunc(builder, bit, ctx->ac.i1, "");
598   ac_build_kill_if_false(&ctx->ac, bit);
599}
600
601/**
602 * Build the pixel shader prolog function. This handles:
603 * - two-side color selection and interpolation
604 * - overriding interpolation parameters for the API PS
605 * - polygon stippling
606 *
607 * All preloaded SGPRs and VGPRs are passed through unmodified unless they are
608 * overriden by other states. (e.g. per-sample interpolation)
609 * Interpolated colors are stored after the preloaded VGPRs.
610 */
611void si_llvm_build_ps_prolog(struct si_shader_context *ctx, union si_shader_part_key *key)
612{
613   LLVMValueRef ret, func;
614   int num_returns, i, num_color_channels;
615
616   memset(&ctx->args, 0, sizeof(ctx->args));
617
618   /* Declare inputs. */
619   LLVMTypeRef return_types[AC_MAX_ARGS];
620   num_returns = 0;
621   num_color_channels = util_bitcount(key->ps_prolog.colors_read);
622   assert(key->ps_prolog.num_input_sgprs + key->ps_prolog.num_input_vgprs + num_color_channels <=
623          AC_MAX_ARGS);
624   for (i = 0; i < key->ps_prolog.num_input_sgprs; i++) {
625      ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, NULL);
626      return_types[num_returns++] = ctx->ac.i32;
627   }
628
629   struct ac_arg pos_fixed_pt;
630   struct ac_arg ancillary;
631   struct ac_arg param_sample_mask;
632   for (i = 0; i < key->ps_prolog.num_input_vgprs; i++) {
633      struct ac_arg *arg = NULL;
634      if (i == key->ps_prolog.ancillary_vgpr_index) {
635         arg = &ancillary;
636      } else if (i == key->ps_prolog.ancillary_vgpr_index + 1) {
637         arg = &param_sample_mask;
638      } else if (i == key->ps_prolog.num_input_vgprs - 1) {
639         /* POS_FIXED_PT is always last. */
640         arg = &pos_fixed_pt;
641      }
642      ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, arg);
643      return_types[num_returns++] = ctx->ac.f32;
644   }
645
646   /* Declare outputs (same as inputs + add colors if needed) */
647   for (i = 0; i < num_color_channels; i++)
648      return_types[num_returns++] = ctx->ac.f32;
649
650   /* Create the function. */
651   si_llvm_create_func(ctx, "ps_prolog", return_types, num_returns, 0);
652   func = ctx->main_fn;
653
654   /* Copy inputs to outputs. This should be no-op, as the registers match,
655    * but it will prevent the compiler from overwriting them unintentionally.
656    */
657   ret = ctx->return_value;
658   for (i = 0; i < ctx->args.arg_count; i++) {
659      LLVMValueRef p = LLVMGetParam(func, i);
660      ret = LLVMBuildInsertValue(ctx->ac.builder, ret, p, i, "");
661   }
662
663   /* Polygon stippling. */
664   if (key->ps_prolog.states.poly_stipple) {
665      LLVMValueRef list = si_prolog_get_internal_bindings(ctx);
666
667      si_llvm_emit_polygon_stipple(ctx, list, pos_fixed_pt);
668   }
669
670   if (key->ps_prolog.states.bc_optimize_for_persp ||
671       key->ps_prolog.states.bc_optimize_for_linear) {
672      unsigned i, base = key->ps_prolog.num_input_sgprs;
673      LLVMValueRef center[2], centroid[2], tmp, bc_optimize;
674
675      /* The shader should do: if (PRIM_MASK[31]) CENTROID = CENTER;
676       * The hw doesn't compute CENTROID if the whole wave only
677       * contains fully-covered quads.
678       *
679       * PRIM_MASK is after user SGPRs.
680       */
681      bc_optimize = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
682      bc_optimize =
683         LLVMBuildLShr(ctx->ac.builder, bc_optimize, LLVMConstInt(ctx->ac.i32, 31, 0), "");
684      bc_optimize = LLVMBuildTrunc(ctx->ac.builder, bc_optimize, ctx->ac.i1, "");
685
686      if (key->ps_prolog.states.bc_optimize_for_persp) {
687         /* Read PERSP_CENTER. */
688         for (i = 0; i < 2; i++)
689            center[i] = LLVMGetParam(func, base + 2 + i);
690         /* Read PERSP_CENTROID. */
691         for (i = 0; i < 2; i++)
692            centroid[i] = LLVMGetParam(func, base + 4 + i);
693         /* Select PERSP_CENTROID. */
694         for (i = 0; i < 2; i++) {
695            tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize, center[i], centroid[i], "");
696            ret = LLVMBuildInsertValue(ctx->ac.builder, ret, tmp, base + 4 + i, "");
697         }
698      }
699      if (key->ps_prolog.states.bc_optimize_for_linear) {
700         /* Read LINEAR_CENTER. */
701         for (i = 0; i < 2; i++)
702            center[i] = LLVMGetParam(func, base + 8 + i);
703         /* Read LINEAR_CENTROID. */
704         for (i = 0; i < 2; i++)
705            centroid[i] = LLVMGetParam(func, base + 10 + i);
706         /* Select LINEAR_CENTROID. */
707         for (i = 0; i < 2; i++) {
708            tmp = LLVMBuildSelect(ctx->ac.builder, bc_optimize, center[i], centroid[i], "");
709            ret = LLVMBuildInsertValue(ctx->ac.builder, ret, tmp, base + 10 + i, "");
710         }
711      }
712   }
713
714   /* Force per-sample interpolation. */
715   if (key->ps_prolog.states.force_persp_sample_interp) {
716      unsigned i, base = key->ps_prolog.num_input_sgprs;
717      LLVMValueRef persp_sample[2];
718
719      /* Read PERSP_SAMPLE. */
720      for (i = 0; i < 2; i++)
721         persp_sample[i] = LLVMGetParam(func, base + i);
722      /* Overwrite PERSP_CENTER. */
723      for (i = 0; i < 2; i++)
724         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_sample[i], base + 2 + i, "");
725      /* Overwrite PERSP_CENTROID. */
726      for (i = 0; i < 2; i++)
727         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_sample[i], base + 4 + i, "");
728   }
729   if (key->ps_prolog.states.force_linear_sample_interp) {
730      unsigned i, base = key->ps_prolog.num_input_sgprs;
731      LLVMValueRef linear_sample[2];
732
733      /* Read LINEAR_SAMPLE. */
734      for (i = 0; i < 2; i++)
735         linear_sample[i] = LLVMGetParam(func, base + 6 + i);
736      /* Overwrite LINEAR_CENTER. */
737      for (i = 0; i < 2; i++)
738         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_sample[i], base + 8 + i, "");
739      /* Overwrite LINEAR_CENTROID. */
740      for (i = 0; i < 2; i++)
741         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_sample[i], base + 10 + i, "");
742   }
743
744   /* Force center interpolation. */
745   if (key->ps_prolog.states.force_persp_center_interp) {
746      unsigned i, base = key->ps_prolog.num_input_sgprs;
747      LLVMValueRef persp_center[2];
748
749      /* Read PERSP_CENTER. */
750      for (i = 0; i < 2; i++)
751         persp_center[i] = LLVMGetParam(func, base + 2 + i);
752      /* Overwrite PERSP_SAMPLE. */
753      for (i = 0; i < 2; i++)
754         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_center[i], base + i, "");
755      /* Overwrite PERSP_CENTROID. */
756      for (i = 0; i < 2; i++)
757         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, persp_center[i], base + 4 + i, "");
758   }
759   if (key->ps_prolog.states.force_linear_center_interp) {
760      unsigned i, base = key->ps_prolog.num_input_sgprs;
761      LLVMValueRef linear_center[2];
762
763      /* Read LINEAR_CENTER. */
764      for (i = 0; i < 2; i++)
765         linear_center[i] = LLVMGetParam(func, base + 8 + i);
766      /* Overwrite LINEAR_SAMPLE. */
767      for (i = 0; i < 2; i++)
768         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_center[i], base + 6 + i, "");
769      /* Overwrite LINEAR_CENTROID. */
770      for (i = 0; i < 2; i++)
771         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, linear_center[i], base + 10 + i, "");
772   }
773
774   /* Interpolate colors. */
775   unsigned color_out_idx = 0;
776   for (i = 0; i < 2; i++) {
777      unsigned writemask = (key->ps_prolog.colors_read >> (i * 4)) & 0xf;
778      unsigned face_vgpr = key->ps_prolog.num_input_sgprs + key->ps_prolog.face_vgpr_index;
779      LLVMValueRef interp[2], color[4];
780      LLVMValueRef interp_ij = NULL, prim_mask = NULL, face = NULL;
781
782      if (!writemask)
783         continue;
784
785      /* If the interpolation qualifier is not CONSTANT (-1). */
786      if (key->ps_prolog.color_interp_vgpr_index[i] != -1) {
787         unsigned interp_vgpr =
788            key->ps_prolog.num_input_sgprs + key->ps_prolog.color_interp_vgpr_index[i];
789
790         /* Get the (i,j) updated by bc_optimize handling. */
791         interp[0] = LLVMBuildExtractValue(ctx->ac.builder, ret, interp_vgpr, "");
792         interp[1] = LLVMBuildExtractValue(ctx->ac.builder, ret, interp_vgpr + 1, "");
793         interp_ij = ac_build_gather_values(&ctx->ac, interp, 2);
794      }
795
796      /* Use the absolute location of the input. */
797      prim_mask = LLVMGetParam(func, SI_PS_NUM_USER_SGPR);
798
799      if (key->ps_prolog.states.color_two_side) {
800         face = LLVMGetParam(func, face_vgpr);
801         face = ac_to_integer(&ctx->ac, face);
802      }
803
804      interp_fs_color(ctx, key->ps_prolog.color_attr_index[i], i, key->ps_prolog.num_interp_inputs,
805                      key->ps_prolog.colors_read, interp_ij, prim_mask, face, color);
806
807      while (writemask) {
808         unsigned chan = u_bit_scan(&writemask);
809         ret = LLVMBuildInsertValue(ctx->ac.builder, ret, color[chan],
810                                    ctx->args.arg_count + color_out_idx++, "");
811      }
812   }
813
814   /* Section 15.2.2 (Shader Inputs) of the OpenGL 4.5 (Core Profile) spec
815    * says:
816    *
817    *    "When per-sample shading is active due to the use of a fragment
818    *     input qualified by sample or due to the use of the gl_SampleID
819    *     or gl_SamplePosition variables, only the bit for the current
820    *     sample is set in gl_SampleMaskIn. When state specifies multiple
821    *     fragment shader invocations for a given fragment, the sample
822    *     mask for any single fragment shader invocation may specify a
823    *     subset of the covered samples for the fragment. In this case,
824    *     the bit corresponding to each covered sample will be set in
825    *     exactly one fragment shader invocation."
826    *
827    * The samplemask loaded by hardware is always the coverage of the
828    * entire pixel/fragment, so mask bits out based on the sample ID.
829    */
830   if (key->ps_prolog.states.samplemask_log_ps_iter) {
831      /* The bit pattern matches that used by fixed function fragment
832       * processing. */
833      static const uint16_t ps_iter_masks[] = {
834         0xffff, /* not used */
835         0x5555, 0x1111, 0x0101, 0x0001,
836      };
837      assert(key->ps_prolog.states.samplemask_log_ps_iter < ARRAY_SIZE(ps_iter_masks));
838
839      uint32_t ps_iter_mask = ps_iter_masks[key->ps_prolog.states.samplemask_log_ps_iter];
840      LLVMValueRef sampleid = si_unpack_param(ctx, ancillary, 8, 4);
841      LLVMValueRef samplemask = ac_get_arg(&ctx->ac, param_sample_mask);
842
843      samplemask = ac_to_integer(&ctx->ac, samplemask);
844      samplemask =
845         LLVMBuildAnd(ctx->ac.builder, samplemask,
846                      LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, ps_iter_mask, false),
847                                   sampleid, ""),
848                      "");
849      samplemask = ac_to_float(&ctx->ac, samplemask);
850
851      ret = LLVMBuildInsertValue(ctx->ac.builder, ret, samplemask, param_sample_mask.arg_index, "");
852   }
853
854   /* Tell LLVM to insert WQM instruction sequence when needed. */
855   if (key->ps_prolog.wqm) {
856      LLVMAddTargetDependentFunctionAttr(func, "amdgpu-ps-wqm-outputs", "");
857   }
858
859   si_llvm_build_ret(ctx, ret);
860}
861
862/**
863 * Build the pixel shader epilog function. This handles everything that must be
864 * emulated for pixel shader exports. (alpha-test, format conversions, etc)
865 */
866void si_llvm_build_ps_epilog(struct si_shader_context *ctx, union si_shader_part_key *key)
867{
868   LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
869   int i;
870   struct si_ps_exports exp = {};
871
872   memset(&ctx->args, 0, sizeof(ctx->args));
873
874   /* Declare input SGPRs. */
875   ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->internal_bindings);
876   ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->bindless_samplers_and_images);
877   ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->const_and_shader_buffers);
878   ac_add_arg(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_INT, &ctx->samplers_and_images);
879   si_add_arg_checked(&ctx->args, AC_ARG_SGPR, 1, AC_ARG_FLOAT, NULL, SI_PARAM_ALPHA_REF);
880
881   /* Declare input VGPRs. */
882   unsigned required_num_params =
883      ctx->args.num_sgprs_used + util_bitcount(key->ps_epilog.colors_written) * 4 +
884      key->ps_epilog.writes_z + key->ps_epilog.writes_stencil + key->ps_epilog.writes_samplemask;
885
886   required_num_params =
887      MAX2(required_num_params, ctx->args.num_sgprs_used + PS_EPILOG_SAMPLEMASK_MIN_LOC + 1);
888
889   while (ctx->args.arg_count < required_num_params)
890      ac_add_arg(&ctx->args, AC_ARG_VGPR, 1, AC_ARG_FLOAT, NULL);
891
892   /* Create the function. */
893   si_llvm_create_func(ctx, "ps_epilog", NULL, 0, 0);
894   /* Disable elimination of unused inputs. */
895   ac_llvm_add_target_dep_function_attr(ctx->main_fn, "InitialPSInputAddr", 0xffffff);
896
897   /* Process colors. */
898   unsigned vgpr = ctx->args.num_sgprs_used;
899   unsigned colors_written = key->ps_epilog.colors_written;
900   int last_color_export = -1;
901
902   /* Find the last color export. */
903   if (!key->ps_epilog.writes_z && !key->ps_epilog.writes_stencil &&
904       !key->ps_epilog.writes_samplemask) {
905      unsigned spi_format = key->ps_epilog.states.spi_shader_col_format;
906
907      /* If last_cbuf > 0, FS_COLOR0_WRITES_ALL_CBUFS is true. */
908      if (colors_written == 0x1 && key->ps_epilog.states.last_cbuf > 0) {
909         /* Just set this if any of the colorbuffers are enabled. */
910         if (spi_format & ((1ull << (4 * (key->ps_epilog.states.last_cbuf + 1))) - 1))
911            last_color_export = 0;
912      } else {
913         for (i = 0; i < 8; i++)
914            if (colors_written & (1 << i) && (spi_format >> (i * 4)) & 0xf)
915               last_color_export = i;
916      }
917   }
918
919   unsigned num_compacted_mrts = 0;
920   while (colors_written) {
921      LLVMValueRef color[4];
922      int output_index = u_bit_scan(&colors_written);
923      unsigned color_type = (key->ps_epilog.color_types >> (output_index * 2)) & 0x3;
924
925      if (color_type != SI_TYPE_ANY32) {
926         for (i = 0; i < 4; i++) {
927            color[i] = LLVMGetParam(ctx->main_fn, vgpr + i / 2);
928            color[i] = LLVMBuildBitCast(ctx->ac.builder, color[i], ctx->ac.v2f16, "");
929            color[i] = ac_llvm_extract_elem(&ctx->ac, color[i], i % 2);
930         }
931         vgpr += 4;
932      } else {
933         for (i = 0; i < 4; i++)
934            color[i] = LLVMGetParam(ctx->main_fn, vgpr++);
935      }
936
937      if (si_export_mrt_color(ctx, color, output_index, num_compacted_mrts,
938                              ctx->args.arg_count - 1,
939                              output_index == last_color_export, color_type, &exp))
940         num_compacted_mrts++;
941   }
942
943   /* Process depth, stencil, samplemask. */
944   if (key->ps_epilog.writes_z)
945      depth = LLVMGetParam(ctx->main_fn, vgpr++);
946   if (key->ps_epilog.writes_stencil)
947      stencil = LLVMGetParam(ctx->main_fn, vgpr++);
948   if (key->ps_epilog.writes_samplemask)
949      samplemask = LLVMGetParam(ctx->main_fn, vgpr++);
950
951   if (depth || stencil || samplemask)
952      ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &exp.args[exp.num++]);
953   else if (last_color_export == -1)
954      ac_build_export_null(&ctx->ac);
955
956   if (exp.num) {
957      for (unsigned i = 0; i < exp.num; i++)
958         ac_build_export(&ctx->ac, &exp.args[i]);
959   }
960
961   /* Compile. */
962   LLVMBuildRetVoid(ctx->ac.builder);
963}
964
965void si_llvm_build_monolithic_ps(struct si_shader_context *ctx, struct si_shader *shader)
966{
967   LLVMValueRef parts[3];
968   unsigned num_parts = 0, main_index;
969   LLVMValueRef main_fn = ctx->main_fn;
970
971   union si_shader_part_key prolog_key;
972   si_get_ps_prolog_key(shader, &prolog_key, false);
973
974   if (si_need_ps_prolog(&prolog_key)) {
975      si_llvm_build_ps_prolog(ctx, &prolog_key);
976      parts[num_parts++] = ctx->main_fn;
977   }
978
979   main_index = num_parts;
980   parts[num_parts++] = main_fn;
981
982   union si_shader_part_key epilog_key;
983   si_get_ps_epilog_key(shader, &epilog_key);
984   si_llvm_build_ps_epilog(ctx, &epilog_key);
985   parts[num_parts++] = ctx->main_fn;
986
987   si_build_wrapper_function(ctx, parts, num_parts, main_index, 0, false);
988}
989
990void si_llvm_init_ps_callbacks(struct si_shader_context *ctx)
991{
992   ctx->abi.emit_outputs = si_llvm_return_fs_outputs;
993   ctx->abi.load_sample_position = load_sample_position;
994   ctx->abi.load_sample_mask_in = load_sample_mask_in;
995   ctx->abi.emit_fbfetch = si_nir_emit_fbfetch;
996}
997