1/*
2 * Copyright © 2019 Google LLC
3 * SPDX-License-Identifier: MIT
4 */
5
6#include "tu_shader.h"
7
8#include "spirv/nir_spirv.h"
9#include "util/mesa-sha1.h"
10#include "nir/nir_xfb_info.h"
11#include "nir/nir_vulkan.h"
12#include "vk_util.h"
13
14#include "ir3/ir3_nir.h"
15
16#include "tu_device.h"
17#include "tu_descriptor_set.h"
18#include "tu_pipeline.h"
19
20nir_shader *
21tu_spirv_to_nir(struct tu_device *dev,
22                void *mem_ctx,
23                const VkPipelineShaderStageCreateInfo *stage_info,
24                gl_shader_stage stage)
25{
26   /* TODO these are made-up */
27   const struct spirv_to_nir_options spirv_options = {
28      .ubo_addr_format = nir_address_format_vec2_index_32bit_offset,
29      .ssbo_addr_format = nir_address_format_vec2_index_32bit_offset,
30
31      /* Accessed via stg/ldg */
32      .phys_ssbo_addr_format = nir_address_format_64bit_global,
33
34      /* Accessed via the const register file */
35      .push_const_addr_format = nir_address_format_logical,
36
37      /* Accessed via ldl/stl */
38      .shared_addr_format = nir_address_format_32bit_offset,
39
40      /* Accessed via stg/ldg (not used with Vulkan?) */
41      .global_addr_format = nir_address_format_64bit_global,
42
43      /* Use 16-bit math for RelaxedPrecision ALU ops */
44      .mediump_16bit_alu = true,
45
46      /* ViewID is a sysval in geometry stages and an input in the FS */
47      .view_index_is_input = stage == MESA_SHADER_FRAGMENT,
48      .caps = {
49         .transform_feedback = true,
50         .tessellation = true,
51         .draw_parameters = true,
52         .image_read_without_format = true,
53         .image_write_without_format = true,
54         .variable_pointers = true,
55         .stencil_export = true,
56         .multiview = true,
57         .shader_viewport_index_layer = true,
58         .geometry_streams = true,
59         .device_group = true,
60         .descriptor_indexing = true,
61         .descriptor_array_dynamic_indexing = true,
62         .descriptor_array_non_uniform_indexing = true,
63         .runtime_descriptor_array = true,
64         .float_controls = true,
65         .float16 = true,
66         .int16 = true,
67         .storage_16bit = dev->physical_device->info->a6xx.storage_16bit,
68         .demote_to_helper_invocation = true,
69         .vk_memory_model = true,
70         .vk_memory_model_device_scope = true,
71         .subgroup_basic = true,
72         .subgroup_ballot = true,
73         .subgroup_vote = true,
74         .subgroup_quad = true,
75         .subgroup_shuffle = true,
76         .subgroup_arithmetic = true,
77         .physical_storage_buffer_address = true,
78      },
79   };
80
81   const nir_shader_compiler_options *nir_options =
82      ir3_get_compiler_options(dev->compiler);
83
84   struct vk_shader_module *module =
85      vk_shader_module_from_handle(stage_info->module);
86
87   nir_shader *nir;
88   VkResult result = vk_shader_module_to_nir(&dev->vk, module,
89                                             stage, stage_info->pName,
90                                             stage_info->pSpecializationInfo,
91                                             &spirv_options, nir_options,
92                                             mem_ctx, &nir);
93   if (result != VK_SUCCESS)
94      return NULL;
95
96   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_NIR)) {
97      fprintf(stderr, "translated nir:\n");
98      nir_print_shader(nir, stderr);
99   }
100
101   const struct nir_lower_sysvals_to_varyings_options sysvals_to_varyings = {
102      .point_coord = true,
103   };
104   NIR_PASS_V(nir, nir_lower_sysvals_to_varyings, &sysvals_to_varyings);
105
106   NIR_PASS_V(nir, nir_lower_global_vars_to_local);
107   NIR_PASS_V(nir, nir_split_var_copies);
108   NIR_PASS_V(nir, nir_lower_var_copies);
109
110   NIR_PASS_V(nir, nir_opt_copy_prop_vars);
111   NIR_PASS_V(nir, nir_opt_combine_stores, nir_var_all);
112
113   NIR_PASS_V(nir, nir_lower_is_helper_invocation);
114
115   NIR_PASS_V(nir, nir_lower_system_values);
116
117   NIR_PASS_V(nir, nir_lower_frexp);
118
119   ir3_optimize_loop(dev->compiler, nir);
120
121   NIR_PASS_V(nir, nir_opt_conditional_discard);
122
123   return nir;
124}
125
126static void
127lower_load_push_constant(struct tu_device *dev,
128                         nir_builder *b,
129                         nir_intrinsic_instr *instr,
130                         struct tu_shader *shader,
131                         const struct tu_pipeline_layout *layout)
132{
133   uint32_t base = nir_intrinsic_base(instr);
134   assert(base % 4 == 0);
135
136   if (tu6_shared_constants_enable(layout, dev->compiler)) {
137      /* All stages share the same range.  We could potentially add
138       * push_constant_offset to layout and apply it, but this is good for
139       * now.
140       */
141      base += dev->compiler->shared_consts_base_offset * 4;
142   } else {
143      assert(base >= shader->push_consts.lo * 4);
144      base -= shader->push_consts.lo * 4;
145   }
146
147   nir_ssa_def *load =
148      nir_load_uniform(b, instr->num_components,
149            instr->dest.ssa.bit_size,
150            nir_ushr(b, instr->src[0].ssa, nir_imm_int(b, 2)),
151            .base = base);
152
153   nir_ssa_def_rewrite_uses(&instr->dest.ssa, load);
154
155   nir_instr_remove(&instr->instr);
156}
157
158static void
159lower_vulkan_resource_index(nir_builder *b, nir_intrinsic_instr *instr,
160                            struct tu_shader *shader,
161                            const struct tu_pipeline_layout *layout)
162{
163   nir_ssa_def *vulkan_idx = instr->src[0].ssa;
164
165   unsigned set = nir_intrinsic_desc_set(instr);
166   unsigned binding = nir_intrinsic_binding(instr);
167   struct tu_descriptor_set_layout *set_layout = layout->set[set].layout;
168   struct tu_descriptor_set_binding_layout *binding_layout =
169      &set_layout->binding[binding];
170   uint32_t base;
171
172   shader->active_desc_sets |= 1u << set;
173
174   switch (binding_layout->type) {
175   case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
176   case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
177      base = (layout->set[set].dynamic_offset_start +
178         binding_layout->dynamic_offset_offset) / (4 * A6XX_TEX_CONST_DWORDS);
179      set = MAX_SETS;
180      break;
181   default:
182      base = binding_layout->offset / (4 * A6XX_TEX_CONST_DWORDS);
183      break;
184   }
185
186   unsigned stride = binding_layout->size / (4 * A6XX_TEX_CONST_DWORDS);
187   assert(util_is_power_of_two_nonzero(stride));
188   nir_ssa_def *shift = nir_imm_int(b, util_logbase2(stride));
189   nir_ssa_def *def = nir_vec3(b, nir_imm_int(b, set),
190                               nir_iadd(b, nir_imm_int(b, base),
191                                        nir_ishl(b, vulkan_idx, shift)),
192                               shift);
193
194   nir_ssa_def_rewrite_uses(&instr->dest.ssa, def);
195   nir_instr_remove(&instr->instr);
196}
197
198static void
199lower_vulkan_resource_reindex(nir_builder *b, nir_intrinsic_instr *instr)
200{
201   nir_ssa_def *old_index = instr->src[0].ssa;
202   nir_ssa_def *delta = instr->src[1].ssa;
203   nir_ssa_def *shift = nir_channel(b, old_index, 2);
204
205   nir_ssa_def *new_index =
206      nir_vec3(b, nir_channel(b, old_index, 0),
207               nir_iadd(b, nir_channel(b, old_index, 1),
208                        nir_ishl(b, delta, shift)),
209               shift);
210
211   nir_ssa_def_rewrite_uses(&instr->dest.ssa, new_index);
212   nir_instr_remove(&instr->instr);
213}
214
215static void
216lower_load_vulkan_descriptor(nir_builder *b, nir_intrinsic_instr *intrin)
217{
218   nir_ssa_def *old_index = intrin->src[0].ssa;
219   /* Loading the descriptor happens as part of the load/store instruction so
220    * this is a no-op. We just need to turn the shift into an offset of 0.
221    */
222   nir_ssa_def *new_index =
223      nir_vec3(b, nir_channel(b, old_index, 0),
224               nir_channel(b, old_index, 1),
225               nir_imm_int(b, 0));
226   nir_ssa_def_rewrite_uses(&intrin->dest.ssa, new_index);
227   nir_instr_remove(&intrin->instr);
228}
229
230static void
231lower_ssbo_ubo_intrinsic(struct tu_device *dev,
232                         nir_builder *b, nir_intrinsic_instr *intrin)
233{
234   const nir_intrinsic_info *info = &nir_intrinsic_infos[intrin->intrinsic];
235
236   /* The bindless base is part of the instruction, which means that part of
237    * the "pointer" has to be constant. We solve this in the same way the blob
238    * does, by generating a bunch of if-statements. In the usual case where
239    * the descriptor set is constant we can skip that, though).
240    */
241
242   unsigned buffer_src;
243   if (intrin->intrinsic == nir_intrinsic_store_ssbo) {
244      /* This has the value first */
245      buffer_src = 1;
246   } else {
247      buffer_src = 0;
248   }
249
250   nir_ssa_scalar scalar_idx = nir_ssa_scalar_resolved(intrin->src[buffer_src].ssa, 0);
251   nir_ssa_def *descriptor_idx = nir_channel(b, intrin->src[buffer_src].ssa, 1);
252
253   /* For isam, we need to use the appropriate descriptor if 16-bit storage is
254    * enabled. Descriptor 0 is the 16-bit one, descriptor 1 is the 32-bit one.
255    */
256   if (dev->physical_device->info->a6xx.storage_16bit &&
257       intrin->intrinsic == nir_intrinsic_load_ssbo &&
258       (nir_intrinsic_access(intrin) & ACCESS_CAN_REORDER) &&
259       intrin->dest.ssa.bit_size > 16) {
260      descriptor_idx = nir_iadd(b, descriptor_idx, nir_imm_int(b, 1));
261   }
262
263   nir_ssa_def *results[MAX_SETS + 1] = { NULL };
264
265   if (nir_ssa_scalar_is_const(scalar_idx)) {
266      nir_ssa_def *bindless =
267         nir_bindless_resource_ir3(b, 32, descriptor_idx, .desc_set = nir_ssa_scalar_as_uint(scalar_idx));
268      nir_instr_rewrite_src_ssa(&intrin->instr, &intrin->src[buffer_src], bindless);
269      return;
270   }
271
272   nir_ssa_def *base_idx = nir_channel(b, scalar_idx.def, scalar_idx.comp);
273   for (unsigned i = 0; i < MAX_SETS + 1; i++) {
274      /* if (base_idx == i) { ... */
275      nir_if *nif = nir_push_if(b, nir_ieq_imm(b, base_idx, i));
276
277      nir_ssa_def *bindless =
278         nir_bindless_resource_ir3(b, 32, descriptor_idx, .desc_set = i);
279
280      nir_intrinsic_instr *copy =
281         nir_intrinsic_instr_create(b->shader, intrin->intrinsic);
282
283      copy->num_components = intrin->num_components;
284
285      for (unsigned src = 0; src < info->num_srcs; src++) {
286         if (src == buffer_src)
287            copy->src[src] = nir_src_for_ssa(bindless);
288         else
289            copy->src[src] = nir_src_for_ssa(intrin->src[src].ssa);
290      }
291
292      for (unsigned idx = 0; idx < info->num_indices; idx++) {
293         copy->const_index[idx] = intrin->const_index[idx];
294      }
295
296      if (info->has_dest) {
297         nir_ssa_dest_init(&copy->instr, &copy->dest,
298                           intrin->dest.ssa.num_components,
299                           intrin->dest.ssa.bit_size,
300                           NULL);
301         results[i] = &copy->dest.ssa;
302      }
303
304      nir_builder_instr_insert(b, &copy->instr);
305
306      /* } else { ... */
307      nir_push_else(b, nif);
308   }
309
310   nir_ssa_def *result =
311      nir_ssa_undef(b, intrin->dest.ssa.num_components, intrin->dest.ssa.bit_size);
312   for (int i = MAX_SETS; i >= 0; i--) {
313      nir_pop_if(b, NULL);
314      if (info->has_dest)
315         result = nir_if_phi(b, results[i], result);
316   }
317
318   if (info->has_dest)
319      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, result);
320   nir_instr_remove(&intrin->instr);
321}
322
323static nir_ssa_def *
324build_bindless(struct tu_device *dev, nir_builder *b,
325               nir_deref_instr *deref, bool is_sampler,
326               struct tu_shader *shader,
327               const struct tu_pipeline_layout *layout)
328{
329   nir_variable *var = nir_deref_instr_get_variable(deref);
330
331   unsigned set = var->data.descriptor_set;
332   unsigned binding = var->data.binding;
333   const struct tu_descriptor_set_binding_layout *bind_layout =
334      &layout->set[set].layout->binding[binding];
335
336   /* input attachments use non bindless workaround */
337   if (bind_layout->type == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT &&
338       likely(!(dev->instance->debug_flags & TU_DEBUG_DYNAMIC))) {
339      const struct glsl_type *glsl_type = glsl_without_array(var->type);
340      uint32_t idx = var->data.index * 2;
341
342      BITSET_SET_RANGE_INSIDE_WORD(b->shader->info.textures_used, idx * 2, ((idx * 2) + (bind_layout->array_size * 2)) - 1);
343
344      /* D24S8 workaround: stencil of D24S8 will be sampled as uint */
345      if (glsl_get_sampler_result_type(glsl_type) == GLSL_TYPE_UINT)
346         idx += 1;
347
348      if (deref->deref_type == nir_deref_type_var)
349         return nir_imm_int(b, idx);
350
351      nir_ssa_def *arr_index = nir_ssa_for_src(b, deref->arr.index, 1);
352      return nir_iadd(b, nir_imm_int(b, idx),
353                      nir_imul_imm(b, arr_index, 2));
354   }
355
356   shader->active_desc_sets |= 1u << set;
357
358   nir_ssa_def *desc_offset;
359   unsigned descriptor_stride;
360   unsigned offset = 0;
361   /* Samplers come second in combined image/sampler descriptors, see
362      * write_combined_image_sampler_descriptor().
363      */
364   if (is_sampler && bind_layout->type ==
365         VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
366      offset = 1;
367   }
368   desc_offset =
369      nir_imm_int(b, (bind_layout->offset / (4 * A6XX_TEX_CONST_DWORDS)) +
370                  offset);
371   descriptor_stride = bind_layout->size / (4 * A6XX_TEX_CONST_DWORDS);
372
373   if (deref->deref_type != nir_deref_type_var) {
374      assert(deref->deref_type == nir_deref_type_array);
375
376      nir_ssa_def *arr_index = nir_ssa_for_src(b, deref->arr.index, 1);
377      desc_offset = nir_iadd(b, desc_offset,
378                             nir_imul_imm(b, arr_index, descriptor_stride));
379   }
380
381   return nir_bindless_resource_ir3(b, 32, desc_offset, .desc_set = set);
382}
383
384static void
385lower_image_deref(struct tu_device *dev, nir_builder *b,
386                  nir_intrinsic_instr *instr, struct tu_shader *shader,
387                  const struct tu_pipeline_layout *layout)
388{
389   nir_deref_instr *deref = nir_src_as_deref(instr->src[0]);
390   nir_ssa_def *bindless = build_bindless(dev, b, deref, false, shader, layout);
391   nir_rewrite_image_intrinsic(instr, bindless, true);
392}
393
394static bool
395lower_intrinsic(nir_builder *b, nir_intrinsic_instr *instr,
396                struct tu_device *dev,
397                struct tu_shader *shader,
398                const struct tu_pipeline_layout *layout)
399{
400   switch (instr->intrinsic) {
401   case nir_intrinsic_load_push_constant:
402      lower_load_push_constant(dev, b, instr, shader, layout);
403      return true;
404
405   case nir_intrinsic_load_vulkan_descriptor:
406      lower_load_vulkan_descriptor(b, instr);
407      return true;
408
409   case nir_intrinsic_vulkan_resource_index:
410      lower_vulkan_resource_index(b, instr, shader, layout);
411      return true;
412   case nir_intrinsic_vulkan_resource_reindex:
413      lower_vulkan_resource_reindex(b, instr);
414      return true;
415
416   case nir_intrinsic_load_ubo:
417   case nir_intrinsic_load_ssbo:
418   case nir_intrinsic_store_ssbo:
419   case nir_intrinsic_ssbo_atomic_add:
420   case nir_intrinsic_ssbo_atomic_imin:
421   case nir_intrinsic_ssbo_atomic_umin:
422   case nir_intrinsic_ssbo_atomic_imax:
423   case nir_intrinsic_ssbo_atomic_umax:
424   case nir_intrinsic_ssbo_atomic_and:
425   case nir_intrinsic_ssbo_atomic_or:
426   case nir_intrinsic_ssbo_atomic_xor:
427   case nir_intrinsic_ssbo_atomic_exchange:
428   case nir_intrinsic_ssbo_atomic_comp_swap:
429   case nir_intrinsic_ssbo_atomic_fadd:
430   case nir_intrinsic_ssbo_atomic_fmin:
431   case nir_intrinsic_ssbo_atomic_fmax:
432   case nir_intrinsic_ssbo_atomic_fcomp_swap:
433   case nir_intrinsic_get_ssbo_size:
434      lower_ssbo_ubo_intrinsic(dev, b, instr);
435      return true;
436
437   case nir_intrinsic_image_deref_load:
438   case nir_intrinsic_image_deref_store:
439   case nir_intrinsic_image_deref_atomic_add:
440   case nir_intrinsic_image_deref_atomic_imin:
441   case nir_intrinsic_image_deref_atomic_umin:
442   case nir_intrinsic_image_deref_atomic_imax:
443   case nir_intrinsic_image_deref_atomic_umax:
444   case nir_intrinsic_image_deref_atomic_and:
445   case nir_intrinsic_image_deref_atomic_or:
446   case nir_intrinsic_image_deref_atomic_xor:
447   case nir_intrinsic_image_deref_atomic_exchange:
448   case nir_intrinsic_image_deref_atomic_comp_swap:
449   case nir_intrinsic_image_deref_size:
450   case nir_intrinsic_image_deref_samples:
451      lower_image_deref(dev, b, instr, shader, layout);
452      return true;
453
454   default:
455      return false;
456   }
457}
458
459static void
460lower_tex_ycbcr(const struct tu_pipeline_layout *layout,
461                nir_builder *builder,
462                nir_tex_instr *tex)
463{
464   int deref_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
465   assert(deref_src_idx >= 0);
466   nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
467
468   nir_variable *var = nir_deref_instr_get_variable(deref);
469   const struct tu_descriptor_set_layout *set_layout =
470      layout->set[var->data.descriptor_set].layout;
471   const struct tu_descriptor_set_binding_layout *binding =
472      &set_layout->binding[var->data.binding];
473   const struct tu_sampler_ycbcr_conversion *ycbcr_samplers =
474      tu_immutable_ycbcr_samplers(set_layout, binding);
475
476   if (!ycbcr_samplers)
477      return;
478
479   /* For the following instructions, we don't apply any change */
480   if (tex->op == nir_texop_txs ||
481       tex->op == nir_texop_query_levels ||
482       tex->op == nir_texop_lod)
483      return;
484
485   assert(tex->texture_index == 0);
486   unsigned array_index = 0;
487   if (deref->deref_type != nir_deref_type_var) {
488      assert(deref->deref_type == nir_deref_type_array);
489      if (!nir_src_is_const(deref->arr.index))
490         return;
491      array_index = nir_src_as_uint(deref->arr.index);
492      array_index = MIN2(array_index, binding->array_size - 1);
493   }
494   const struct tu_sampler_ycbcr_conversion *ycbcr_sampler = ycbcr_samplers + array_index;
495
496   if (ycbcr_sampler->ycbcr_model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY)
497      return;
498
499   builder->cursor = nir_after_instr(&tex->instr);
500
501   uint8_t bits = vk_format_get_component_bits(ycbcr_sampler->format,
502                                               UTIL_FORMAT_COLORSPACE_RGB,
503                                               PIPE_SWIZZLE_X);
504
505   switch (ycbcr_sampler->format) {
506   case VK_FORMAT_G8B8G8R8_422_UNORM:
507   case VK_FORMAT_B8G8R8G8_422_UNORM:
508   case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
509   case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
510      /* util_format_get_component_bits doesn't return what we want */
511      bits = 8;
512      break;
513   default:
514      break;
515   }
516
517   uint32_t bpcs[3] = {bits, bits, bits}; /* TODO: use right bpc for each channel ? */
518   nir_ssa_def *result = nir_convert_ycbcr_to_rgb(builder,
519                                                  ycbcr_sampler->ycbcr_model,
520                                                  ycbcr_sampler->ycbcr_range,
521                                                  &tex->dest.ssa,
522                                                  bpcs);
523   nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, result,
524                                  result->parent_instr);
525
526   builder->cursor = nir_before_instr(&tex->instr);
527}
528
529static bool
530lower_tex(nir_builder *b, nir_tex_instr *tex, struct tu_device *dev,
531          struct tu_shader *shader, const struct tu_pipeline_layout *layout)
532{
533   lower_tex_ycbcr(layout, b, tex);
534
535   int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
536   if (sampler_src_idx >= 0) {
537      nir_deref_instr *deref = nir_src_as_deref(tex->src[sampler_src_idx].src);
538      nir_ssa_def *bindless = build_bindless(dev, b, deref, true, shader, layout);
539      nir_instr_rewrite_src(&tex->instr, &tex->src[sampler_src_idx].src,
540                            nir_src_for_ssa(bindless));
541      tex->src[sampler_src_idx].src_type = nir_tex_src_sampler_handle;
542   }
543
544   int tex_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
545   if (tex_src_idx >= 0) {
546      nir_deref_instr *deref = nir_src_as_deref(tex->src[tex_src_idx].src);
547      nir_ssa_def *bindless = build_bindless(dev, b, deref, false, shader, layout);
548      nir_instr_rewrite_src(&tex->instr, &tex->src[tex_src_idx].src,
549                            nir_src_for_ssa(bindless));
550      tex->src[tex_src_idx].src_type = nir_tex_src_texture_handle;
551
552      /* for the input attachment case: */
553      if (bindless->parent_instr->type != nir_instr_type_intrinsic)
554         tex->src[tex_src_idx].src_type = nir_tex_src_texture_offset;
555   }
556
557   return true;
558}
559
560struct lower_instr_params {
561   struct tu_device *dev;
562   struct tu_shader *shader;
563   const struct tu_pipeline_layout *layout;
564};
565
566static bool
567lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
568{
569   struct lower_instr_params *params = cb_data;
570   b->cursor = nir_before_instr(instr);
571   switch (instr->type) {
572   case nir_instr_type_tex:
573      return lower_tex(b, nir_instr_as_tex(instr), params->dev, params->shader, params->layout);
574   case nir_instr_type_intrinsic:
575      return lower_intrinsic(b, nir_instr_as_intrinsic(instr), params->dev, params->shader, params->layout);
576   default:
577      return false;
578   }
579}
580
581/* Figure out the range of push constants that we're actually going to push to
582 * the shader, and tell the backend to reserve this range when pushing UBO
583 * constants.
584 */
585
586static void
587gather_push_constants(nir_shader *shader, struct tu_shader *tu_shader)
588{
589   uint32_t min = UINT32_MAX, max = 0;
590   nir_foreach_function(function, shader) {
591      if (!function->impl)
592         continue;
593
594      nir_foreach_block(block, function->impl) {
595         nir_foreach_instr_safe(instr, block) {
596            if (instr->type != nir_instr_type_intrinsic)
597               continue;
598
599            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
600            if (intrin->intrinsic != nir_intrinsic_load_push_constant)
601               continue;
602
603            uint32_t base = nir_intrinsic_base(intrin);
604            uint32_t range = nir_intrinsic_range(intrin);
605            min = MIN2(min, base);
606            max = MAX2(max, base + range);
607            break;
608         }
609      }
610   }
611
612   if (min >= max) {
613      tu_shader->push_consts.lo = 0;
614      tu_shader->push_consts.dwords = 0;
615      return;
616   }
617
618   /* CP_LOAD_STATE OFFSET and NUM_UNIT for SHARED_CONSTS are in units of
619    * dwords while loading regular consts is in units of vec4's.
620    * So we unify the unit here as dwords for tu_push_constant_range, then
621    * we should consider correct unit when emitting.
622    *
623    * Note there's an alignment requirement of 16 dwords on OFFSET. Expand
624    * the range and change units accordingly.
625    */
626   tu_shader->push_consts.lo = (min / 4) / 4 * 4;
627   tu_shader->push_consts.dwords =
628      align(max, 16) / 4 - tu_shader->push_consts.lo;
629}
630
631static bool
632tu_lower_io(nir_shader *shader, struct tu_device *dev,
633            struct tu_shader *tu_shader,
634            const struct tu_pipeline_layout *layout)
635{
636   if (!tu6_shared_constants_enable(layout, dev->compiler))
637      gather_push_constants(shader, tu_shader);
638
639   struct lower_instr_params params = {
640      .dev = dev,
641      .shader = tu_shader,
642      .layout = layout,
643   };
644
645   bool progress = nir_shader_instructions_pass(shader,
646                                                lower_instr,
647                                                nir_metadata_none,
648                                                &params);
649
650   /* Remove now-unused variables so that when we gather the shader info later
651    * they won't be counted.
652    */
653
654   if (progress)
655      nir_opt_dce(shader);
656
657   progress |=
658      nir_remove_dead_variables(shader,
659                                nir_var_uniform | nir_var_mem_ubo | nir_var_mem_ssbo,
660                                NULL);
661
662   return progress;
663}
664
665static void
666shared_type_info(const struct glsl_type *type, unsigned *size, unsigned *align)
667{
668   assert(glsl_type_is_vector_or_scalar(type));
669
670   unsigned comp_size =
671      glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
672   unsigned length = glsl_get_vector_elements(type);
673   *size = comp_size * length;
674   *align = comp_size;
675}
676
677static void
678tu_gather_xfb_info(nir_shader *nir, struct ir3_stream_output_info *info)
679{
680   nir_shader_gather_xfb_info(nir);
681
682   if (!nir->xfb_info)
683      return;
684
685   nir_xfb_info *xfb = nir->xfb_info;
686
687   uint8_t output_map[VARYING_SLOT_TESS_MAX];
688   memset(output_map, 0, sizeof(output_map));
689
690   nir_foreach_shader_out_variable(var, nir) {
691      unsigned slots =
692         var->data.compact ? DIV_ROUND_UP(glsl_get_length(var->type), 4)
693                           : glsl_count_attribute_slots(var->type, false);
694      for (unsigned i = 0; i < slots; i++)
695         output_map[var->data.location + i] = var->data.driver_location + i;
696   }
697
698   assert(xfb->output_count <= IR3_MAX_SO_OUTPUTS);
699   info->num_outputs = xfb->output_count;
700
701   for (int i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
702      info->stride[i] = xfb->buffers[i].stride / 4;
703      info->buffer_to_stream[i] = xfb->buffer_to_stream[i];
704   }
705
706   info->streams_written = xfb->streams_written;
707
708   for (int i = 0; i < xfb->output_count; i++) {
709      info->output[i].register_index = output_map[xfb->outputs[i].location];
710      info->output[i].start_component = xfb->outputs[i].component_offset;
711      info->output[i].num_components =
712                           util_bitcount(xfb->outputs[i].component_mask);
713      info->output[i].output_buffer  = xfb->outputs[i].buffer;
714      info->output[i].dst_offset = xfb->outputs[i].offset / 4;
715      info->output[i].stream = xfb->buffer_to_stream[xfb->outputs[i].buffer];
716   }
717}
718
719struct tu_shader *
720tu_shader_create(struct tu_device *dev,
721                 nir_shader *nir,
722                 const struct tu_shader_key *key,
723                 struct tu_pipeline_layout *layout,
724                 const VkAllocationCallbacks *alloc)
725{
726   struct tu_shader *shader;
727
728   shader = vk_zalloc2(
729      &dev->vk.alloc, alloc,
730      sizeof(*shader),
731      8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
732   if (!shader)
733      return NULL;
734
735   NIR_PASS_V(nir, nir_opt_access, &(nir_opt_access_options) {
736               .is_vulkan = true,
737               .infer_non_readable = true,
738             });
739
740   if (nir->info.stage == MESA_SHADER_FRAGMENT) {
741      NIR_PASS_V(nir, nir_lower_input_attachments,
742                 &(nir_input_attachment_options) {
743                     .use_fragcoord_sysval = true,
744                     .use_layer_id_sysval = false,
745                     /* When using multiview rendering, we must use
746                      * gl_ViewIndex as the layer id to pass to the texture
747                      * sampling function. gl_Layer doesn't work when
748                      * multiview is enabled.
749                      */
750                     .use_view_id_for_layer = key->multiview_mask != 0,
751                 });
752   }
753
754   /* This needs to happen before multiview lowering which rewrites store
755    * instructions of the position variable, so that we can just rewrite one
756    * store at the end instead of having to rewrite every store specified by
757    * the user.
758    */
759   ir3_nir_lower_io_to_temporaries(nir);
760
761   if (nir->info.stage == MESA_SHADER_VERTEX && key->multiview_mask) {
762      tu_nir_lower_multiview(nir, key->multiview_mask,
763                             &shader->multi_pos_output, dev);
764   }
765
766   if (nir->info.stage == MESA_SHADER_FRAGMENT && key->force_sample_interp) {
767      nir_foreach_shader_in_variable(var, nir) {
768         if (!var->data.centroid)
769            var->data.sample = true;
770      }
771   }
772
773   NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_push_const,
774              nir_address_format_32bit_offset);
775
776   NIR_PASS_V(nir, nir_lower_explicit_io,
777              nir_var_mem_ubo | nir_var_mem_ssbo,
778              nir_address_format_vec2_index_32bit_offset);
779
780   NIR_PASS_V(nir, nir_lower_explicit_io,
781              nir_var_mem_global,
782              nir_address_format_64bit_global);
783
784   if (nir->info.stage == MESA_SHADER_COMPUTE) {
785      NIR_PASS_V(nir, nir_lower_vars_to_explicit_types,
786                 nir_var_mem_shared, shared_type_info);
787      NIR_PASS_V(nir, nir_lower_explicit_io,
788                 nir_var_mem_shared,
789                 nir_address_format_32bit_offset);
790
791      if (nir->info.zero_initialize_shared_memory && nir->info.shared_size > 0) {
792         const unsigned chunk_size = 16; /* max single store size */
793         /* Shared memory is allocated in 1024b chunks in HW, but the zero-init
794          * extension only requires us to initialize the memory that the shader
795          * is allocated at the API level, and it's up to the user to ensure
796          * that accesses are limited to those bounds.
797          */
798         const unsigned shared_size = ALIGN(nir->info.shared_size, chunk_size);
799         NIR_PASS_V(nir, nir_zero_initialize_shared_memory, shared_size, chunk_size);
800      }
801
802      const struct nir_lower_compute_system_values_options compute_sysval_options = {
803         .has_base_workgroup_id = true,
804      };
805      NIR_PASS_V(nir, nir_lower_compute_system_values, &compute_sysval_options);
806   }
807
808   nir_assign_io_var_locations(nir, nir_var_shader_in, &nir->num_inputs, nir->info.stage);
809   nir_assign_io_var_locations(nir, nir_var_shader_out, &nir->num_outputs, nir->info.stage);
810
811  /* Gather information for transform feedback. This should be called after:
812    * - nir_split_per_member_structs.
813    * - nir_remove_dead_variables with varyings, so that we could align
814    *   stream outputs correctly.
815    * - nir_assign_io_var_locations - to have valid driver_location
816    */
817   struct ir3_stream_output_info so_info = {};
818   if (nir->info.stage == MESA_SHADER_VERTEX ||
819         nir->info.stage == MESA_SHADER_TESS_EVAL ||
820         nir->info.stage == MESA_SHADER_GEOMETRY)
821      tu_gather_xfb_info(nir, &so_info);
822
823   NIR_PASS_V(nir, tu_lower_io, dev, shader, layout);
824
825   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
826
827   ir3_finalize_nir(dev->compiler, nir);
828
829   uint32_t reserved_consts_vec4 = align(shader->push_consts.dwords, 16) / 4;
830   bool shared_consts_enable = tu6_shared_constants_enable(layout, dev->compiler);
831   if (shared_consts_enable)
832      assert(!shader->push_consts.dwords);
833
834   shader->ir3_shader =
835      ir3_shader_from_nir(dev->compiler, nir, &(struct ir3_shader_options) {
836                           .reserved_user_consts = reserved_consts_vec4,
837                           .shared_consts_enable = shared_consts_enable,
838                           .api_wavesize = key->api_wavesize,
839                           .real_wavesize = key->real_wavesize,
840                          }, &so_info);
841
842   return shader;
843}
844
845void
846tu_shader_destroy(struct tu_device *dev,
847                  struct tu_shader *shader,
848                  const VkAllocationCallbacks *alloc)
849{
850   ir3_shader_destroy(shader->ir3_shader);
851
852   vk_free2(&dev->vk.alloc, alloc, shader);
853}
854