1/*
2 * Copyright 2018 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 "tgsi/tgsi_text.h"
27#include "tgsi/tgsi_ureg.h"
28
29void *si_get_blitter_vs(struct si_context *sctx, enum blitter_attrib_type type, unsigned num_layers)
30{
31   unsigned vs_blit_property;
32   void **vs;
33
34   switch (type) {
35   case UTIL_BLITTER_ATTRIB_NONE:
36      vs = num_layers > 1 ? &sctx->vs_blit_pos_layered : &sctx->vs_blit_pos;
37      vs_blit_property = SI_VS_BLIT_SGPRS_POS;
38      break;
39   case UTIL_BLITTER_ATTRIB_COLOR:
40      vs = num_layers > 1 ? &sctx->vs_blit_color_layered : &sctx->vs_blit_color;
41      vs_blit_property = SI_VS_BLIT_SGPRS_POS_COLOR;
42      break;
43   case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
44   case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
45      assert(num_layers == 1);
46      vs = &sctx->vs_blit_texcoord;
47      vs_blit_property = SI_VS_BLIT_SGPRS_POS_TEXCOORD;
48      break;
49   default:
50      assert(0);
51      return NULL;
52   }
53   if (*vs)
54      return *vs;
55
56   struct ureg_program *ureg = ureg_create(PIPE_SHADER_VERTEX);
57   if (!ureg)
58      return NULL;
59
60   /* Tell the shader to load VS inputs from SGPRs: */
61   ureg_property(ureg, TGSI_PROPERTY_VS_BLIT_SGPRS_AMD, vs_blit_property);
62   ureg_property(ureg, TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION, true);
63
64   /* This is just a pass-through shader with 1-3 MOV instructions. */
65   ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0), ureg_DECL_vs_input(ureg, 0));
66
67   if (type != UTIL_BLITTER_ATTRIB_NONE) {
68      ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, 0), ureg_DECL_vs_input(ureg, 1));
69   }
70
71   if (num_layers > 1) {
72      struct ureg_src instance_id = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_INSTANCEID, 0);
73      struct ureg_dst layer = ureg_DECL_output(ureg, TGSI_SEMANTIC_LAYER, 0);
74
75      ureg_MOV(ureg, ureg_writemask(layer, TGSI_WRITEMASK_X),
76               ureg_scalar(instance_id, TGSI_SWIZZLE_X));
77   }
78   ureg_END(ureg);
79
80   *vs = ureg_create_shader_and_destroy(ureg, &sctx->b);
81   return *vs;
82}
83
84/**
85 * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
86 * VS passes its outputs to TES directly, so the fixed-function shader only
87 * has to write TESSOUTER and TESSINNER.
88 */
89void *si_create_fixed_func_tcs(struct si_context *sctx)
90{
91   struct ureg_src outer, inner;
92   struct ureg_dst tessouter, tessinner;
93   struct ureg_program *ureg = ureg_create(PIPE_SHADER_TESS_CTRL);
94
95   if (!ureg)
96      return NULL;
97
98   outer = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_TESS_DEFAULT_OUTER_LEVEL, 0);
99   inner = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_TESS_DEFAULT_INNER_LEVEL, 0);
100
101   tessouter = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSOUTER, 0);
102   tessinner = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSINNER, 0);
103
104   ureg_MOV(ureg, tessouter, outer);
105   ureg_MOV(ureg, tessinner, inner);
106   ureg_END(ureg);
107
108   return ureg_create_shader_and_destroy(ureg, &sctx->b);
109}
110
111/* Create a compute shader implementing clear_buffer or copy_buffer. */
112void *si_create_dma_compute_shader(struct pipe_context *ctx, unsigned num_dwords_per_thread,
113                                   bool dst_stream_cache_policy, bool is_copy)
114{
115   struct si_screen *sscreen = (struct si_screen *)ctx->screen;
116   assert(util_is_power_of_two_nonzero(num_dwords_per_thread));
117
118   unsigned store_qualifier = TGSI_MEMORY_COHERENT | TGSI_MEMORY_RESTRICT;
119   if (dst_stream_cache_policy)
120      store_qualifier |= TGSI_MEMORY_STREAM_CACHE_POLICY;
121
122   /* Don't cache loads, because there is no reuse. */
123   unsigned load_qualifier = store_qualifier | TGSI_MEMORY_STREAM_CACHE_POLICY;
124
125   unsigned num_mem_ops = MAX2(1, num_dwords_per_thread / 4);
126   unsigned *inst_dwords = alloca(num_mem_ops * sizeof(unsigned));
127
128   for (unsigned i = 0; i < num_mem_ops; i++) {
129      if (i * 4 < num_dwords_per_thread)
130         inst_dwords[i] = MIN2(4, num_dwords_per_thread - i * 4);
131   }
132
133   struct ureg_program *ureg = ureg_create(PIPE_SHADER_COMPUTE);
134   if (!ureg)
135      return NULL;
136
137   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH, sscreen->compute_wave_size);
138   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT, 1);
139   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH, 1);
140
141   struct ureg_src value;
142   if (!is_copy) {
143      ureg_property(ureg, TGSI_PROPERTY_CS_USER_DATA_COMPONENTS_AMD, inst_dwords[0]);
144      value = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_CS_USER_DATA_AMD, 0);
145   }
146
147   struct ureg_src tid = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_THREAD_ID, 0);
148   struct ureg_src blk = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_BLOCK_ID, 0);
149   struct ureg_dst store_addr = ureg_writemask(ureg_DECL_temporary(ureg), TGSI_WRITEMASK_X);
150   struct ureg_dst load_addr = ureg_writemask(ureg_DECL_temporary(ureg), TGSI_WRITEMASK_X);
151   struct ureg_dst dstbuf = ureg_dst(ureg_DECL_buffer(ureg, 0, false));
152   struct ureg_src srcbuf;
153   struct ureg_src *values = NULL;
154
155   if (is_copy) {
156      srcbuf = ureg_DECL_buffer(ureg, 1, false);
157      values = malloc(num_mem_ops * sizeof(struct ureg_src));
158   }
159
160   /* If there are multiple stores, the first store writes into 0*wavesize+tid,
161    * the 2nd store writes into 1*wavesize+tid, the 3rd store writes into 2*wavesize+tid, etc.
162    */
163   ureg_UMAD(ureg, store_addr, blk, ureg_imm1u(ureg, sscreen->compute_wave_size * num_mem_ops),
164             tid);
165   /* Convert from a "store size unit" into bytes. */
166   ureg_UMUL(ureg, store_addr, ureg_src(store_addr), ureg_imm1u(ureg, 4 * inst_dwords[0]));
167   ureg_MOV(ureg, load_addr, ureg_src(store_addr));
168
169   /* Distance between a load and a store for latency hiding. */
170   unsigned load_store_distance = is_copy ? 8 : 0;
171
172   for (unsigned i = 0; i < num_mem_ops + load_store_distance; i++) {
173      int d = i - load_store_distance;
174
175      if (is_copy && i < num_mem_ops) {
176         if (i) {
177            ureg_UADD(ureg, load_addr, ureg_src(load_addr),
178                      ureg_imm1u(ureg, 4 * inst_dwords[i] * sscreen->compute_wave_size));
179         }
180
181         values[i] = ureg_src(ureg_DECL_temporary(ureg));
182         struct ureg_dst dst =
183            ureg_writemask(ureg_dst(values[i]), u_bit_consecutive(0, inst_dwords[i]));
184         struct ureg_src srcs[] = {srcbuf, ureg_src(load_addr)};
185         ureg_memory_insn(ureg, TGSI_OPCODE_LOAD, &dst, 1, srcs, 2, load_qualifier,
186                          TGSI_TEXTURE_BUFFER, 0);
187      }
188
189      if (d >= 0) {
190         if (d) {
191            ureg_UADD(ureg, store_addr, ureg_src(store_addr),
192                      ureg_imm1u(ureg, 4 * inst_dwords[d] * sscreen->compute_wave_size));
193         }
194
195         struct ureg_dst dst = ureg_writemask(dstbuf, u_bit_consecutive(0, inst_dwords[d]));
196         struct ureg_src srcs[] = {ureg_src(store_addr), is_copy ? values[d] : value};
197         ureg_memory_insn(ureg, TGSI_OPCODE_STORE, &dst, 1, srcs, 2, store_qualifier,
198                          TGSI_TEXTURE_BUFFER, 0);
199      }
200   }
201   ureg_END(ureg);
202
203   struct pipe_compute_state state = {};
204   state.ir_type = PIPE_SHADER_IR_TGSI;
205   state.prog = ureg_get_tokens(ureg, NULL);
206
207   void *cs = ctx->create_compute_state(ctx, &state);
208   ureg_destroy(ureg);
209   ureg_free_tokens(state.prog);
210
211   free(values);
212   return cs;
213}
214
215/* Create a compute shader implementing clear_buffer or copy_buffer. */
216void *si_create_clear_buffer_rmw_cs(struct pipe_context *ctx)
217{
218   const char *text = "COMP\n"
219                      "PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
220                      "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
221                      "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
222                      "PROPERTY CS_USER_DATA_COMPONENTS_AMD 2\n"
223                      "DCL SV[0], THREAD_ID\n"
224                      "DCL SV[1], BLOCK_ID\n"
225                      "DCL SV[2], CS_USER_DATA_AMD\n"
226                      "DCL BUFFER[0]\n"
227                      "DCL TEMP[0..1]\n"
228                      "IMM[0] UINT32 {64, 16, 0, 0}\n"
229                      /* ADDRESS = BLOCK_ID * 64 + THREAD_ID; */
230                      "UMAD TEMP[0].x, SV[1].xxxx, IMM[0].xxxx, SV[0].xxxx\n"
231                      /* ADDRESS = ADDRESS * 16; (byte offset, loading one vec4 per thread) */
232                      "UMUL TEMP[0].x, TEMP[0].xxxx, IMM[0].yyyy\n"
233                      "LOAD TEMP[1], BUFFER[0], TEMP[0].xxxx\n"
234                      /* DATA &= inverted_writemask; */
235                      "AND TEMP[1], TEMP[1], SV[2].yyyy\n"
236                      /* DATA |= clear_value_masked; */
237                      "OR TEMP[1], TEMP[1], SV[2].xxxx\n"
238                      "STORE BUFFER[0].xyzw, TEMP[0], TEMP[1]%s\n"
239                      "END\n";
240   char final_text[2048];
241   struct tgsi_token tokens[1024];
242   struct pipe_compute_state state = {0};
243
244   snprintf(final_text, sizeof(final_text), text,
245            SI_COMPUTE_DST_CACHE_POLICY != L2_LRU ? ", STREAM_CACHE_POLICY" : "");
246
247   if (!tgsi_text_translate(final_text, tokens, ARRAY_SIZE(tokens))) {
248      assert(false);
249      return NULL;
250   }
251
252   state.ir_type = PIPE_SHADER_IR_TGSI;
253   state.prog = tokens;
254
255   return ctx->create_compute_state(ctx, &state);
256}
257
258/* Create the compute shader that is used to collect the results.
259 *
260 * One compute grid with a single thread is launched for every query result
261 * buffer. The thread (optionally) reads a previous summary buffer, then
262 * accumulates data from the query result buffer, and writes the result either
263 * to a summary buffer to be consumed by the next grid invocation or to the
264 * user-supplied buffer.
265 *
266 * Data layout:
267 *
268 * CONST
269 *  0.x = end_offset
270 *  0.y = result_stride
271 *  0.z = result_count
272 *  0.w = bit field:
273 *          1: read previously accumulated values
274 *          2: write accumulated values for chaining
275 *          4: write result available
276 *          8: convert result to boolean (0/1)
277 *         16: only read one dword and use that as result
278 *         32: apply timestamp conversion
279 *         64: store full 64 bits result
280 *        128: store signed 32 bits result
281 *        256: SO_OVERFLOW mode: take the difference of two successive half-pairs
282 *  1.x = fence_offset
283 *  1.y = pair_stride
284 *  1.z = pair_count
285 *
286 * BUFFER[0] = query result buffer
287 * BUFFER[1] = previous summary buffer
288 * BUFFER[2] = next summary buffer or user-supplied buffer
289 */
290void *si_create_query_result_cs(struct si_context *sctx)
291{
292   /* TEMP[0].xy = accumulated result so far
293    * TEMP[0].z = result not available
294    *
295    * TEMP[1].x = current result index
296    * TEMP[1].y = current pair index
297    */
298   static const char text_tmpl[] =
299      "COMP\n"
300      "PROPERTY CS_FIXED_BLOCK_WIDTH 1\n"
301      "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
302      "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
303      "DCL BUFFER[0]\n"
304      "DCL BUFFER[1]\n"
305      "DCL BUFFER[2]\n"
306      "DCL CONST[0][0..1]\n"
307      "DCL TEMP[0..5]\n"
308      "IMM[0] UINT32 {0, 31, 2147483647, 4294967295}\n"
309      "IMM[1] UINT32 {1, 2, 4, 8}\n"
310      "IMM[2] UINT32 {16, 32, 64, 128}\n"
311      "IMM[3] UINT32 {1000000, 0, %u, 0}\n" /* for timestamp conversion */
312      "IMM[4] UINT32 {256, 0, 0, 0}\n"
313
314      "AND TEMP[5], CONST[0][0].wwww, IMM[2].xxxx\n"
315      "UIF TEMP[5]\n"
316      /* Check result availability. */
317      "LOAD TEMP[1].x, BUFFER[0], CONST[0][1].xxxx\n"
318      "ISHR TEMP[0].z, TEMP[1].xxxx, IMM[0].yyyy\n"
319      "MOV TEMP[1], TEMP[0].zzzz\n"
320      "NOT TEMP[0].z, TEMP[0].zzzz\n"
321
322      /* Load result if available. */
323      "UIF TEMP[1]\n"
324      "LOAD TEMP[0].xy, BUFFER[0], IMM[0].xxxx\n"
325      "ENDIF\n"
326      "ELSE\n"
327      /* Load previously accumulated result if requested. */
328      "MOV TEMP[0], IMM[0].xxxx\n"
329      "AND TEMP[4], CONST[0][0].wwww, IMM[1].xxxx\n"
330      "UIF TEMP[4]\n"
331      "LOAD TEMP[0].xyz, BUFFER[1], IMM[0].xxxx\n"
332      "ENDIF\n"
333
334      "MOV TEMP[1].x, IMM[0].xxxx\n"
335      "BGNLOOP\n"
336      /* Break if accumulated result so far is not available. */
337      "UIF TEMP[0].zzzz\n"
338      "BRK\n"
339      "ENDIF\n"
340
341      /* Break if result_index >= result_count. */
342      "USGE TEMP[5], TEMP[1].xxxx, CONST[0][0].zzzz\n"
343      "UIF TEMP[5]\n"
344      "BRK\n"
345      "ENDIF\n"
346
347      /* Load fence and check result availability */
348      "UMAD TEMP[5].x, TEMP[1].xxxx, CONST[0][0].yyyy, CONST[0][1].xxxx\n"
349      "LOAD TEMP[5].x, BUFFER[0], TEMP[5].xxxx\n"
350      "ISHR TEMP[0].z, TEMP[5].xxxx, IMM[0].yyyy\n"
351      "NOT TEMP[0].z, TEMP[0].zzzz\n"
352      "UIF TEMP[0].zzzz\n"
353      "BRK\n"
354      "ENDIF\n"
355
356      "MOV TEMP[1].y, IMM[0].xxxx\n"
357      "BGNLOOP\n"
358      /* Load start and end. */
359      "UMUL TEMP[5].x, TEMP[1].xxxx, CONST[0][0].yyyy\n"
360      "UMAD TEMP[5].x, TEMP[1].yyyy, CONST[0][1].yyyy, TEMP[5].xxxx\n"
361      "LOAD TEMP[2].xy, BUFFER[0], TEMP[5].xxxx\n"
362
363      "UADD TEMP[5].y, TEMP[5].xxxx, CONST[0][0].xxxx\n"
364      "LOAD TEMP[3].xy, BUFFER[0], TEMP[5].yyyy\n"
365
366      "U64ADD TEMP[4].xy, TEMP[3], -TEMP[2]\n"
367
368      "AND TEMP[5].z, CONST[0][0].wwww, IMM[4].xxxx\n"
369      "UIF TEMP[5].zzzz\n"
370      /* Load second start/end half-pair and
371       * take the difference
372       */
373      "UADD TEMP[5].xy, TEMP[5], IMM[1].wwww\n"
374      "LOAD TEMP[2].xy, BUFFER[0], TEMP[5].xxxx\n"
375      "LOAD TEMP[3].xy, BUFFER[0], TEMP[5].yyyy\n"
376
377      "U64ADD TEMP[3].xy, TEMP[3], -TEMP[2]\n"
378      "U64ADD TEMP[4].xy, TEMP[4], -TEMP[3]\n"
379      "ENDIF\n"
380
381      "U64ADD TEMP[0].xy, TEMP[0], TEMP[4]\n"
382
383      /* Increment pair index */
384      "UADD TEMP[1].y, TEMP[1].yyyy, IMM[1].xxxx\n"
385      "USGE TEMP[5], TEMP[1].yyyy, CONST[0][1].zzzz\n"
386      "UIF TEMP[5]\n"
387      "BRK\n"
388      "ENDIF\n"
389      "ENDLOOP\n"
390
391      /* Increment result index */
392      "UADD TEMP[1].x, TEMP[1].xxxx, IMM[1].xxxx\n"
393      "ENDLOOP\n"
394      "ENDIF\n"
395
396      "AND TEMP[4], CONST[0][0].wwww, IMM[1].yyyy\n"
397      "UIF TEMP[4]\n"
398      /* Store accumulated data for chaining. */
399      "STORE BUFFER[2].xyz, IMM[0].xxxx, TEMP[0]\n"
400      "ELSE\n"
401      "AND TEMP[4], CONST[0][0].wwww, IMM[1].zzzz\n"
402      "UIF TEMP[4]\n"
403      /* Store result availability. */
404      "NOT TEMP[0].z, TEMP[0]\n"
405      "AND TEMP[0].z, TEMP[0].zzzz, IMM[1].xxxx\n"
406      "STORE BUFFER[2].x, IMM[0].xxxx, TEMP[0].zzzz\n"
407
408      "AND TEMP[4], CONST[0][0].wwww, IMM[2].zzzz\n"
409      "UIF TEMP[4]\n"
410      "STORE BUFFER[2].y, IMM[0].xxxx, IMM[0].xxxx\n"
411      "ENDIF\n"
412      "ELSE\n"
413      /* Store result if it is available. */
414      "NOT TEMP[4], TEMP[0].zzzz\n"
415      "UIF TEMP[4]\n"
416      /* Apply timestamp conversion */
417      "AND TEMP[4], CONST[0][0].wwww, IMM[2].yyyy\n"
418      "UIF TEMP[4]\n"
419      "U64MUL TEMP[0].xy, TEMP[0], IMM[3].xyxy\n"
420      "U64DIV TEMP[0].xy, TEMP[0], IMM[3].zwzw\n"
421      "ENDIF\n"
422
423      /* Convert to boolean */
424      "AND TEMP[4], CONST[0][0].wwww, IMM[1].wwww\n"
425      "UIF TEMP[4]\n"
426      "U64SNE TEMP[0].x, TEMP[0].xyxy, IMM[4].zwzw\n"
427      "AND TEMP[0].x, TEMP[0].xxxx, IMM[1].xxxx\n"
428      "MOV TEMP[0].y, IMM[0].xxxx\n"
429      "ENDIF\n"
430
431      "AND TEMP[4], CONST[0][0].wwww, IMM[2].zzzz\n"
432      "UIF TEMP[4]\n"
433      "STORE BUFFER[2].xy, IMM[0].xxxx, TEMP[0].xyxy\n"
434      "ELSE\n"
435      /* Clamping */
436      "UIF TEMP[0].yyyy\n"
437      "MOV TEMP[0].x, IMM[0].wwww\n"
438      "ENDIF\n"
439
440      "AND TEMP[4], CONST[0][0].wwww, IMM[2].wwww\n"
441      "UIF TEMP[4]\n"
442      "UMIN TEMP[0].x, TEMP[0].xxxx, IMM[0].zzzz\n"
443      "ENDIF\n"
444
445      "STORE BUFFER[2].x, IMM[0].xxxx, TEMP[0].xxxx\n"
446      "ENDIF\n"
447      "ENDIF\n"
448      "ENDIF\n"
449      "ENDIF\n"
450
451      "END\n";
452
453   char text[sizeof(text_tmpl) + 32];
454   struct tgsi_token tokens[1024];
455   struct pipe_compute_state state = {};
456
457   /* Hard code the frequency into the shader so that the backend can
458    * use the full range of optimizations for divide-by-constant.
459    */
460   snprintf(text, sizeof(text), text_tmpl, sctx->screen->info.clock_crystal_freq);
461
462   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
463      assert(false);
464      return NULL;
465   }
466
467   state.ir_type = PIPE_SHADER_IR_TGSI;
468   state.prog = tokens;
469
470   return sctx->b.create_compute_state(&sctx->b, &state);
471}
472
473/* Create a compute shader implementing copy_image.
474 * Luckily, this works with all texture targets except 1D_ARRAY.
475 */
476void *si_create_copy_image_compute_shader(struct pipe_context *ctx)
477{
478   static const char text[] =
479      "COMP\n"
480      "PROPERTY CS_USER_DATA_COMPONENTS_AMD 3\n"
481      "DCL SV[0], THREAD_ID\n"
482      "DCL SV[1], BLOCK_ID\n"
483      "DCL SV[2], BLOCK_SIZE\n"
484      "DCL SV[3], CS_USER_DATA_AMD\n"
485      "DCL IMAGE[0], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
486      "DCL IMAGE[1], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
487      "DCL TEMP[0..3], LOCAL\n"
488      "IMM[0] UINT32 {65535, 16, 0, 0}\n"
489
490      "UMAD TEMP[0].xyz, SV[1], SV[2], SV[0]\n" /* threadID.xyz */
491      "AND TEMP[1].xyz, SV[3], IMM[0].xxxx\n"    /* src.xyz */
492      "UADD TEMP[1].xyz, TEMP[1], TEMP[0]\n" /* src.xyz + threadID.xyz */
493      "LOAD TEMP[3], IMAGE[0], TEMP[1], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
494      "USHR TEMP[2].xyz, SV[3], IMM[0].yyyy\n"   /* dst.xyz */
495      "UADD TEMP[2].xyz, TEMP[2], TEMP[0]\n" /* dst.xyz + threadID.xyz */
496      "STORE IMAGE[1], TEMP[2], TEMP[3], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
497      "END\n";
498
499   struct tgsi_token tokens[1024];
500   struct pipe_compute_state state = {0};
501
502   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
503      assert(false);
504      return NULL;
505   }
506
507   state.ir_type = PIPE_SHADER_IR_TGSI;
508   state.prog = tokens;
509
510   return ctx->create_compute_state(ctx, &state);
511}
512
513void *si_create_copy_image_compute_shader_1d_array(struct pipe_context *ctx)
514{
515   static const char text[] =
516      "COMP\n"
517      "PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
518      "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
519      "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
520      "PROPERTY CS_USER_DATA_COMPONENTS_AMD 3\n"
521      "DCL SV[0], THREAD_ID\n"
522      "DCL SV[1], BLOCK_ID\n"
523      "DCL SV[2], CS_USER_DATA_AMD\n"
524      "DCL IMAGE[0], 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
525      "DCL IMAGE[1], 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
526      "DCL TEMP[0..4], LOCAL\n"
527      "IMM[0] UINT32 {64, 1, 65535, 16}\n"
528
529      "UMAD TEMP[0].xz, SV[1].xyyy, IMM[0].xyyy, SV[0].xyyy\n" /* threadID.xz */
530      "AND TEMP[1].xz, SV[2], IMM[0].zzzz\n"    /* src.xz */
531      "UADD TEMP[1].xz, TEMP[1], TEMP[0]\n"     /* src.xz + threadID.xz */
532      "LOAD TEMP[3], IMAGE[0], TEMP[1].xzzz, 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
533      "USHR TEMP[2].xz, SV[2], IMM[0].wwww\n"   /* dst.xz */
534      "UADD TEMP[2].xz, TEMP[2], TEMP[0]\n" /* dst.xz + threadID.xz */
535      "STORE IMAGE[1], TEMP[2].xzzz, TEMP[3], 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
536      "END\n";
537
538   struct tgsi_token tokens[1024];
539   struct pipe_compute_state state = {0};
540
541   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
542      assert(false);
543      return NULL;
544   }
545
546   state.ir_type = PIPE_SHADER_IR_TGSI;
547   state.prog = tokens;
548
549   return ctx->create_compute_state(ctx, &state);
550}
551
552/* Create a compute shader implementing DCC decompression via a blit.
553 * This is a trivial copy_image shader except that it has a variable block
554 * size and a barrier.
555 */
556void *si_create_dcc_decompress_cs(struct pipe_context *ctx)
557{
558   static const char text[] =
559      "COMP\n"
560      "DCL SV[0], THREAD_ID\n"
561      "DCL SV[1], BLOCK_ID\n"
562      "DCL SV[2], BLOCK_SIZE\n"
563      "DCL IMAGE[0], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
564      "DCL IMAGE[1], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
565      "DCL TEMP[0..1]\n"
566
567      "UMAD TEMP[0].xyz, SV[1].xyzz, SV[2].xyzz, SV[0].xyzz\n"
568      "LOAD TEMP[1], IMAGE[0], TEMP[0].xyzz, 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
569      /* Wait for the whole threadgroup (= DCC block) to load texels before
570       * overwriting them, because overwriting any pixel within a DCC block
571       * can break compression for the whole block.
572       */
573      "BARRIER\n"
574      "STORE IMAGE[1], TEMP[0].xyzz, TEMP[1], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
575      "END\n";
576
577   struct tgsi_token tokens[1024];
578   struct pipe_compute_state state = {0};
579
580   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
581      assert(false);
582      return NULL;
583   }
584
585   state.ir_type = PIPE_SHADER_IR_TGSI;
586   state.prog = tokens;
587
588   return ctx->create_compute_state(ctx, &state);
589}
590
591void *si_clear_render_target_shader(struct pipe_context *ctx)
592{
593   static const char text[] =
594      "COMP\n"
595      "PROPERTY CS_FIXED_BLOCK_WIDTH 8\n"
596      "PROPERTY CS_FIXED_BLOCK_HEIGHT 8\n"
597      "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
598      "DCL SV[0], THREAD_ID\n"
599      "DCL SV[1], BLOCK_ID\n"
600      "DCL IMAGE[0], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
601      "DCL CONST[0][0..1]\n" // 0:xyzw 1:xyzw
602      "DCL TEMP[0..3], LOCAL\n"
603      "IMM[0] UINT32 {8, 1, 0, 0}\n"
604      "MOV TEMP[0].xyz, CONST[0][0].xyzw\n"
605      "UMAD TEMP[1].xyz, SV[1].xyzz, IMM[0].xxyy, SV[0].xyzz\n"
606      "UADD TEMP[2].xyz, TEMP[1].xyzx, TEMP[0].xyzx\n"
607      "MOV TEMP[3].xyzw, CONST[0][1].xyzw\n"
608      "STORE IMAGE[0], TEMP[2].xyzz, TEMP[3], 2D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
609      "END\n";
610
611   struct tgsi_token tokens[1024];
612   struct pipe_compute_state state = {0};
613
614   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
615      assert(false);
616      return NULL;
617   }
618
619   state.ir_type = PIPE_SHADER_IR_TGSI;
620   state.prog = tokens;
621
622   return ctx->create_compute_state(ctx, &state);
623}
624
625/* TODO: Didn't really test 1D_ARRAY */
626void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx)
627{
628   static const char text[] =
629      "COMP\n"
630      "PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
631      "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
632      "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
633      "DCL SV[0], THREAD_ID\n"
634      "DCL SV[1], BLOCK_ID\n"
635      "DCL IMAGE[0], 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT, WR\n"
636      "DCL CONST[0][0..1]\n" // 0:xyzw 1:xyzw
637      "DCL TEMP[0..3], LOCAL\n"
638      "IMM[0] UINT32 {64, 1, 0, 0}\n"
639      "MOV TEMP[0].xy, CONST[0][0].xzzw\n"
640      "UMAD TEMP[1].xy, SV[1].xyzz, IMM[0].xyyy, SV[0].xyzz\n"
641      "UADD TEMP[2].xy, TEMP[1].xyzx, TEMP[0].xyzx\n"
642      "MOV TEMP[3].xyzw, CONST[0][1].xyzw\n"
643      "STORE IMAGE[0], TEMP[2].xyzz, TEMP[3], 1D_ARRAY, PIPE_FORMAT_R32G32B32A32_FLOAT\n"
644      "END\n";
645
646   struct tgsi_token tokens[1024];
647   struct pipe_compute_state state = {0};
648
649   if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
650      assert(false);
651      return NULL;
652   }
653
654   state.ir_type = PIPE_SHADER_IR_TGSI;
655   state.prog = tokens;
656
657   return ctx->create_compute_state(ctx, &state);
658}
659
660void *si_clear_12bytes_buffer_shader(struct pipe_context *ctx)
661{
662   static const char text[] = "COMP\n"
663                              "PROPERTY CS_FIXED_BLOCK_WIDTH 64\n"
664                              "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
665                              "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
666                              "PROPERTY CS_USER_DATA_COMPONENTS_AMD 3\n"
667                              "DCL SV[0], THREAD_ID\n"
668                              "DCL SV[1], BLOCK_ID\n"
669                              "DCL SV[2], CS_USER_DATA_AMD\n"
670                              "DCL BUFFER[0]\n"
671                              "DCL TEMP[0..0]\n"
672                              "IMM[0] UINT32 {64, 1, 12, 0}\n"
673                              "UMAD TEMP[0].x, SV[1].xyzz, IMM[0].xyyy, SV[0].xyzz\n"
674                              "UMUL TEMP[0].x, TEMP[0].xyzz, IMM[0].zzzz\n" // 12 bytes
675                              "STORE BUFFER[0].xyz, TEMP[0].xxxx, SV[2].xyzz%s\n"
676                              "END\n";
677   char final_text[2048];
678   struct tgsi_token tokens[1024];
679   struct pipe_compute_state state = {0};
680
681   snprintf(final_text, sizeof(final_text), text,
682            SI_COMPUTE_DST_CACHE_POLICY != L2_LRU ? ", STREAM_CACHE_POLICY" : "");
683
684   if (!tgsi_text_translate(final_text, tokens, ARRAY_SIZE(tokens))) {
685      assert(false);
686      return NULL;
687   }
688
689   state.ir_type = PIPE_SHADER_IR_TGSI;
690   state.prog = tokens;
691
692   return ctx->create_compute_state(ctx, &state);
693}
694
695/* Load samples from the image, and copy them to the same image. This looks like
696 * a no-op, but it's not. Loads use FMASK, while stores don't, so samples are
697 * reordered to match expanded FMASK.
698 *
699 * After the shader finishes, FMASK should be cleared to identity.
700 */
701void *si_create_fmask_expand_cs(struct pipe_context *ctx, unsigned num_samples, bool is_array)
702{
703   enum tgsi_texture_type target = is_array ? TGSI_TEXTURE_2D_ARRAY_MSAA : TGSI_TEXTURE_2D_MSAA;
704   struct ureg_program *ureg = ureg_create(PIPE_SHADER_COMPUTE);
705   if (!ureg)
706      return NULL;
707
708   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH, 8);
709   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT, 8);
710   ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH, 1);
711
712   /* Compute the image coordinates. */
713   struct ureg_src image = ureg_DECL_image(ureg, 0, target, 0, true, false);
714   struct ureg_src tid = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_THREAD_ID, 0);
715   struct ureg_src blk = ureg_DECL_system_value(ureg, TGSI_SEMANTIC_BLOCK_ID, 0);
716   struct ureg_dst coord = ureg_writemask(ureg_DECL_temporary(ureg), TGSI_WRITEMASK_XYZW);
717   ureg_UMAD(ureg, ureg_writemask(coord, TGSI_WRITEMASK_XY), ureg_swizzle(blk, 0, 1, 1, 1),
718             ureg_imm2u(ureg, 8, 8), ureg_swizzle(tid, 0, 1, 1, 1));
719   if (is_array) {
720      ureg_MOV(ureg, ureg_writemask(coord, TGSI_WRITEMASK_Z), ureg_scalar(blk, TGSI_SWIZZLE_Z));
721   }
722
723   /* Load samples, resolving FMASK. */
724   struct ureg_dst sample[8];
725   assert(num_samples <= ARRAY_SIZE(sample));
726
727   for (unsigned i = 0; i < num_samples; i++) {
728      sample[i] = ureg_DECL_temporary(ureg);
729
730      ureg_MOV(ureg, ureg_writemask(coord, TGSI_WRITEMASK_W), ureg_imm1u(ureg, i));
731
732      struct ureg_src srcs[] = {image, ureg_src(coord)};
733      ureg_memory_insn(ureg, TGSI_OPCODE_LOAD, &sample[i], 1, srcs, 2, TGSI_MEMORY_RESTRICT, target,
734                       0);
735   }
736
737   /* Store samples, ignoring FMASK. */
738   for (unsigned i = 0; i < num_samples; i++) {
739      ureg_MOV(ureg, ureg_writemask(coord, TGSI_WRITEMASK_W), ureg_imm1u(ureg, i));
740
741      struct ureg_dst dst_image = ureg_dst(image);
742      struct ureg_src srcs[] = {ureg_src(coord), ureg_src(sample[i])};
743      ureg_memory_insn(ureg, TGSI_OPCODE_STORE, &dst_image, 1, srcs, 2, TGSI_MEMORY_RESTRICT,
744                       target, 0);
745   }
746   ureg_END(ureg);
747
748   struct pipe_compute_state state = {};
749   state.ir_type = PIPE_SHADER_IR_TGSI;
750   state.prog = ureg_get_tokens(ureg, NULL);
751
752   void *cs = ctx->create_compute_state(ctx, &state);
753   ureg_destroy(ureg);
754   return cs;
755}
756
757/* Create the compute shader that is used to collect the results of gfx10+
758 * shader queries.
759 *
760 * One compute grid with a single thread is launched for every query result
761 * buffer. The thread (optionally) reads a previous summary buffer, then
762 * accumulates data from the query result buffer, and writes the result either
763 * to a summary buffer to be consumed by the next grid invocation or to the
764 * user-supplied buffer.
765 *
766 * Data layout:
767 *
768 * BUFFER[0] = query result buffer (layout is defined by gfx10_sh_query_buffer_mem)
769 * BUFFER[1] = previous summary buffer
770 * BUFFER[2] = next summary buffer or user-supplied buffer
771 *
772 * CONST
773 *  0.x = config; the low 3 bits indicate the mode:
774 *          0: sum up counts
775 *          1: determine result availability and write it as a boolean
776 *          2: SO_OVERFLOW
777 *          3: SO_ANY_OVERFLOW
778 *        the remaining bits form a bitfield:
779 *          8: write result as a 64-bit value
780 *  0.y = offset in bytes to counts or stream for SO_OVERFLOW mode
781 *  0.z = chain bit field:
782 *          1: have previous summary buffer
783 *          2: write next summary buffer
784 *  0.w = result_count
785 */
786void *gfx10_create_sh_query_result_cs(struct si_context *sctx)
787{
788   /* TEMP[0].x = accumulated result so far
789    * TEMP[0].y = result missing
790    * TEMP[0].z = whether we're in overflow mode
791    */
792   static const char text_tmpl[] = "COMP\n"
793                                   "PROPERTY CS_FIXED_BLOCK_WIDTH 1\n"
794                                   "PROPERTY CS_FIXED_BLOCK_HEIGHT 1\n"
795                                   "PROPERTY CS_FIXED_BLOCK_DEPTH 1\n"
796                                   "DCL BUFFER[0]\n"
797                                   "DCL BUFFER[1]\n"
798                                   "DCL BUFFER[2]\n"
799                                   "DCL CONST[0][0..0]\n"
800                                   "DCL TEMP[0..5]\n"
801                                   "IMM[0] UINT32 {0, 7, 256, 4294967295}\n"
802                                   "IMM[1] UINT32 {1, 2, 4, 8}\n"
803                                   "IMM[2] UINT32 {16, 32, 64, 128}\n"
804
805                                   /*
806                                   acc_result = 0;
807                                   acc_missing = 0;
808                                   if (chain & 1) {
809                                           acc_result = buffer[1][0];
810                                           acc_missing = buffer[1][1];
811                                   }
812                                   */
813                                   "MOV TEMP[0].xy, IMM[0].xxxx\n"
814                                   "AND TEMP[5], CONST[0][0].zzzz, IMM[1].xxxx\n"
815                                   "UIF TEMP[5]\n"
816                                   "LOAD TEMP[0].xy, BUFFER[1], IMM[0].xxxx\n"
817                                   "ENDIF\n"
818
819                                   /*
820                                   is_overflow (TEMP[0].z) = (config & 7) >= 2;
821                                   result_remaining (TEMP[1].x) = (is_overflow && acc_result) ? 0 :
822                                   result_count; base_offset (TEMP[1].y) = 0; for (;;) { if
823                                   (!result_remaining) break; result_remaining--;
824                                   */
825                                   "AND TEMP[5].x, CONST[0][0].xxxx, IMM[0].yyyy\n"
826                                   "USGE TEMP[0].z, TEMP[5].xxxx, IMM[1].yyyy\n"
827
828                                   "AND TEMP[5].x, TEMP[0].zzzz, TEMP[0].xxxx\n"
829                                   "UCMP TEMP[1].x, TEMP[5].xxxx, IMM[0].xxxx, CONST[0][0].wwww\n"
830                                   "MOV TEMP[1].y, IMM[0].xxxx\n"
831
832                                   "BGNLOOP\n"
833                                   "USEQ TEMP[5], TEMP[1].xxxx, IMM[0].xxxx\n"
834                                   "UIF TEMP[5]\n"
835                                   "BRK\n"
836                                   "ENDIF\n"
837                                   "UADD TEMP[1].x, TEMP[1].xxxx, IMM[0].wwww\n"
838
839                                   /*
840                                   fence = buffer[0]@(base_offset + sizeof(gfx10_sh_query_buffer_mem.stream));
841                                   if (!fence) {
842                                           acc_missing = ~0u;
843                                           break;
844                                   }
845                                   */
846                                   "UADD TEMP[5].x, TEMP[1].yyyy, IMM[2].wwww\n"
847                                   "LOAD TEMP[5].x, BUFFER[0], TEMP[5].xxxx\n"
848                                   "USEQ TEMP[5], TEMP[5].xxxx, IMM[0].xxxx\n"
849                                   "UIF TEMP[5]\n"
850                                   "MOV TEMP[0].y, TEMP[5].xxxx\n"
851                                   "BRK\n"
852                                   "ENDIF\n"
853
854                                   /*
855                                   stream_offset (TEMP[2].x) = base_offset + offset;
856
857                                   if (!(config & 7)) {
858                                           acc_result += buffer[0]@stream_offset;
859                                   }
860                                   */
861                                   "UADD TEMP[2].x, TEMP[1].yyyy, CONST[0][0].yyyy\n"
862
863                                   "AND TEMP[5].x, CONST[0][0].xxxx, IMM[0].yyyy\n"
864                                   "USEQ TEMP[5], TEMP[5].xxxx, IMM[0].xxxx\n"
865                                   "UIF TEMP[5]\n"
866                                   "LOAD TEMP[5].x, BUFFER[0], TEMP[2].xxxx\n"
867                                   "UADD TEMP[0].x, TEMP[0].xxxx, TEMP[5].xxxx\n"
868                                   "ENDIF\n"
869
870                                   /*
871                                   if ((config & 7) >= 2) {
872                                           count (TEMP[2].y) = (config & 1) ? 4 : 1;
873                                   */
874                                   "AND TEMP[5].x, CONST[0][0].xxxx, IMM[0].yyyy\n"
875                                   "USGE TEMP[5], TEMP[5].xxxx, IMM[1].yyyy\n"
876                                   "UIF TEMP[5]\n"
877                                   "AND TEMP[5].x, CONST[0][0].xxxx, IMM[1].xxxx\n"
878                                   "UCMP TEMP[2].y, TEMP[5].xxxx, IMM[1].zzzz, IMM[1].xxxx\n"
879
880                                   /*
881                                   do {
882                                           generated = buffer[0]@(stream_offset + 2 * sizeof(uint64_t));
883                                           emitted = buffer[0]@(stream_offset + 3 * sizeof(uint64_t));
884                                           if (generated != emitted) {
885                                                   acc_result = 1;
886                                                   result_remaining = 0;
887                                                   break;
888                                           }
889
890                                           stream_offset += sizeof(gfx10_sh_query_buffer_mem.stream[0]);
891                                   } while (--count);
892                                   */
893                                   "BGNLOOP\n"
894                                   "UADD TEMP[5].x, TEMP[2].xxxx, IMM[2].xxxx\n"
895                                   "LOAD TEMP[4].xyzw, BUFFER[0], TEMP[5].xxxx\n"
896                                   "USNE TEMP[5], TEMP[4].xyxy, TEMP[4].zwzw\n"
897                                   "UIF TEMP[5]\n"
898                                   "MOV TEMP[0].x, IMM[1].xxxx\n"
899                                   "MOV TEMP[1].y, IMM[0].xxxx\n"
900                                   "BRK\n"
901                                   "ENDIF\n"
902
903                                   "UADD TEMP[2].y, TEMP[2].yyyy, IMM[0].wwww\n"
904                                   "USEQ TEMP[5], TEMP[2].yyyy, IMM[0].xxxx\n"
905                                   "UIF TEMP[5]\n"
906                                   "BRK\n"
907                                   "ENDIF\n"
908                                   "UADD TEMP[2].x, TEMP[2].xxxx, IMM[2].yyyy\n"
909                                   "ENDLOOP\n"
910                                   "ENDIF\n"
911
912                                   /*
913                                           base_offset += sizeof(gfx10_sh_query_buffer_mem);
914                                   } // end outer loop
915                                   */
916                                   "UADD TEMP[1].y, TEMP[1].yyyy, IMM[0].zzzz\n"
917                                   "ENDLOOP\n"
918
919                                   /*
920                                   if (chain & 2) {
921                                           buffer[2][0] = acc_result;
922                                           buffer[2][1] = acc_missing;
923                                   } else {
924                                   */
925                                   "AND TEMP[5], CONST[0][0].zzzz, IMM[1].yyyy\n"
926                                   "UIF TEMP[5]\n"
927                                   "STORE BUFFER[2].xy, IMM[0].xxxx, TEMP[0]\n"
928                                   "ELSE\n"
929
930                                   /*
931                                   if ((config & 7) == 1) {
932                                           acc_result = acc_missing ? 0 : 1;
933                                           acc_missing = 0;
934                                   }
935                                   */
936                                   "AND TEMP[5], CONST[0][0].xxxx, IMM[0].yyyy\n"
937                                   "USEQ TEMP[5], TEMP[5].xxxx, IMM[1].xxxx\n"
938                                   "UIF TEMP[5]\n"
939                                   "UCMP TEMP[0].x, TEMP[0].yyyy, IMM[0].xxxx, IMM[1].xxxx\n"
940                                   "MOV TEMP[0].y, IMM[0].xxxx\n"
941                                   "ENDIF\n"
942
943                                   /*
944                                   if (!acc_missing) {
945                                           buffer[2][0] = acc_result;
946                                           if (config & 8)
947                                                   buffer[2][1] = 0;
948                                   }
949                                   */
950                                   "USEQ TEMP[5], TEMP[0].yyyy, IMM[0].xxxx\n"
951                                   "UIF TEMP[5]\n"
952                                   "STORE BUFFER[2].x, IMM[0].xxxx, TEMP[0].xxxx\n"
953
954                                   "AND TEMP[5], CONST[0][0].xxxx, IMM[1].wwww\n"
955                                   "UIF TEMP[5]\n"
956                                   "STORE BUFFER[2].x, IMM[1].zzzz, TEMP[0].yyyy\n"
957                                   "ENDIF\n"
958                                   "ENDIF\n"
959                                   "ENDIF\n"
960
961                                   "END\n";
962
963   struct tgsi_token tokens[1024];
964   struct pipe_compute_state state = {};
965
966   if (!tgsi_text_translate(text_tmpl, tokens, ARRAY_SIZE(tokens))) {
967      assert(false);
968      return NULL;
969   }
970
971   state.ir_type = PIPE_SHADER_IR_TGSI;
972   state.prog = tokens;
973
974   return sctx->b.create_compute_state(&sctx->b, &state);
975}
976