panvk: Implement texture/image queries

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16276>

src/panfrost/ci/deqp-panfrost-g52-vk.toml

@@ -24,7 +24,11 @@ include = [
     "dEQP-VK.glsl.derivate.*.constant.*",
     "dEQP-VK.glsl.derivate.*.linear.*",
     "dEQP-VK.glsl.derivate.*.uniform_*",
+    "dEQP-VK.glsl.texture_functions.query.texturesize*",
+    "dEQP-VK.glsl.texture_functions.query.texturesamples.*",
+    "dEQP-VK.glsl.texture_functions.query.texturequerylevels.*",
     "dEQP-VK.glsl.operator.*",
+    "dEQP-VK.image.image_size.*",
     "dEQP-VK.image.load_store.with_format.*",
     "dEQP-VK.pipeline.depth.format.d24_unorm_s8_uint.depth_test_disabled.depth_write_enabled",
     "dEQP-VK.pipeline.input_assembly.*",

src/panfrost/vulkan/panvk_vX_descriptor_set.c

@@ -42,6 +42,44 @@
 #include "pan_bo.h"
 #include "panvk_cs.h"
 
+#define PANVK_DESCRIPTOR_ALIGN 16
+
+struct panvk_bview_desc {
+   uint32_t elems;
+};
+
+static void
+panvk_fill_bview_desc(struct panvk_bview_desc *desc,
+                      struct panvk_buffer_view *view)
+{
+   desc->elems = view->elems;
+}
+
+struct panvk_image_desc {
+   uint32_t width;
+   uint32_t height;
+   uint32_t depth;
+   uint16_t levels;
+   uint16_t samples;
+};
+
+static void
+panvk_fill_image_desc(struct panvk_image_desc *desc,
+                      struct panvk_image_view *view)
+{
+   desc->width = view->vk.extent.width;
+   desc->height = view->vk.extent.height;
+   desc->depth = view->vk.extent.depth;
+   desc->levels = view->vk.level_count;
+   desc->samples = view->vk.image->samples;
+
+   /* Stick array layer count after the last valid size component */
+   if (view->vk.image->image_type == VK_IMAGE_TYPE_1D)
+      desc->height = view->vk.layer_count;
+   else if (view->vk.image->image_type == VK_IMAGE_TYPE_2D)
+      desc->depth = view->vk.layer_count;
+}
+
 VkResult
 panvk_per_arch(CreateDescriptorSetLayout)(VkDevice _device,
                                           const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
@@ -124,15 +162,18 @@ panvk_per_arch(CreateDescriptorSetLayout)(VkDevice _device,
          binding_layout->tex_idx = tex_idx;
          sampler_idx += binding_layout->array_size;
          tex_idx += binding_layout->array_size;
+         binding_layout->desc_ubo_stride = sizeof(struct panvk_image_desc);
          break;
       case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
          binding_layout->tex_idx = tex_idx;
          tex_idx += binding_layout->array_size;
+         binding_layout->desc_ubo_stride = sizeof(struct panvk_image_desc);
          break;
       case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
          binding_layout->tex_idx = tex_idx;
          tex_idx += binding_layout->array_size;
+         binding_layout->desc_ubo_stride = sizeof(struct panvk_bview_desc);
          break;
       case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
          binding_layout->dyn_ubo_idx = dyn_ubo_idx;
@@ -152,15 +193,18 @@ panvk_per_arch(CreateDescriptorSetLayout)(VkDevice _device,
       case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
          binding_layout->img_idx = img_idx;
          img_idx += binding_layout->array_size;
+         binding_layout->desc_ubo_stride = sizeof(struct panvk_image_desc);
          break;
       case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
          binding_layout->img_idx = img_idx;
          img_idx += binding_layout->array_size;
+         binding_layout->desc_ubo_stride = sizeof(struct panvk_bview_desc);
          break;
       default:
          unreachable("Invalid descriptor type");
       }
 
+      desc_ubo_size = ALIGN_POT(desc_ubo_size, PANVK_DESCRIPTOR_ALIGN);
       binding_layout->desc_ubo_offset = desc_ubo_size;
       desc_ubo_size += binding_layout->desc_ubo_stride *
                        binding_layout->array_size;
@@ -407,6 +451,12 @@ panvk_set_tex_desc(struct panvk_descriptor_set *set,
 #else
    ((mali_ptr *)set->textures)[tex_idx] = view->bo->ptr.gpu;
 #endif
+
+   void *desc = (char *)set->desc_bo->ptr.cpu +
+                binding_layout->desc_ubo_offset +
+                elem * binding_layout->desc_ubo_stride;
+
+   panvk_fill_image_desc(desc, view);
 }
 
 static void
@@ -425,6 +475,12 @@ panvk_set_tex_buf_desc(struct panvk_descriptor_set *set,
 #else
    ((mali_ptr *)set->textures)[tex_idx] = view->bo->ptr.gpu;
 #endif
+
+   void *desc = (char *)set->desc_bo->ptr.cpu +
+                binding_layout->desc_ubo_offset +
+                elem * binding_layout->desc_ubo_stride;
+
+   panvk_fill_bview_desc(desc, view);
 }
 
 static void
@@ -444,6 +500,12 @@ panvk_set_img_desc(struct panvk_device *dev,
 
    set->img_fmts[img_idx] = pdev->formats[view->pview.format].hw;
    memcpy(attrib_buf, view->descs.img_attrib_buf, pan_size(ATTRIBUTE_BUFFER) * 2);
+
+   void *desc = (char *)set->desc_bo->ptr.cpu +
+                binding_layout->desc_ubo_offset +
+                elem * binding_layout->desc_ubo_stride;
+
+   panvk_fill_image_desc(desc, view);
 }
 
 static void
@@ -463,6 +525,12 @@ panvk_set_img_buf_desc(struct panvk_device *dev,
 
    set->img_fmts[img_idx] = pdev->formats[view->fmt].hw;
    memcpy(attrib_buf, view->descs.img_attrib_buf, pan_size(ATTRIBUTE_BUFFER) * 2);
+
+   void *desc = (char *)set->desc_bo->ptr.cpu +
+                binding_layout->desc_ubo_offset +
+                elem * binding_layout->desc_ubo_stride;
+
+   panvk_fill_bview_desc(desc, view);
 }
 
 static void

src/panfrost/vulkan/panvk_vX_nir_lower_descriptors.c

@@ -337,16 +337,92 @@ get_resource_deref_binding(nir_deref_instr *deref,
    *binding = var->data.binding;
 }
 
+static nir_ssa_def *
+load_resource_deref_desc(nir_builder *b, nir_deref_instr *deref,
+                         unsigned desc_offset,
+                         unsigned num_components, unsigned bit_size,
+                         const struct apply_descriptors_ctx *ctx)
+{
+   uint32_t set, binding, index_imm;
+   nir_ssa_def *index_ssa;
+   get_resource_deref_binding(deref, &set, &binding,
+                              &index_imm, &index_ssa);
+
+   const struct panvk_descriptor_set_layout *set_layout =
+      ctx->layout->sets[set].layout;
+   const struct panvk_descriptor_set_binding_layout *bind_layout =
+      &set_layout->bindings[binding];
+
+   assert(index_ssa == NULL || index_imm == 0);
+   if (index_ssa == NULL)
+      index_ssa = nir_imm_int(b, index_imm);
+
+   const unsigned set_ubo_idx =
+      panvk_pipeline_layout_ubo_start(ctx->layout, set, false) +
+      set_layout->desc_ubo_index;
+
+   nir_ssa_def *desc_ubo_offset =
+      nir_iadd_imm(b, nir_imul_imm(b, index_ssa,
+                                      bind_layout->desc_ubo_stride),
+                      bind_layout->desc_ubo_offset + desc_offset);
+
+   assert(bind_layout->desc_ubo_stride > 0);
+   unsigned desc_align = (1 << (ffs(bind_layout->desc_ubo_stride) - 1));
+   desc_align = MIN2(desc_align, 16);
+
+   return nir_load_ubo(b, num_components, bit_size,
+                       nir_imm_int(b, set_ubo_idx),
+                       desc_ubo_offset,
+                       .align_mul=desc_align,
+                       .align_offset=(desc_offset % desc_align),
+                       .range=~0);
+}
 
 static bool
 lower_tex(nir_builder *b, nir_tex_instr *tex,
           const struct apply_descriptors_ctx *ctx)
 {
    bool progress = false;
-   int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
 
    b->cursor = nir_before_instr(&tex->instr);
 
+   if (tex->op == nir_texop_txs ||
+       tex->op == nir_texop_query_levels ||
+       tex->op == nir_texop_texture_samples) {
+      int tex_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
+      assert(tex_src_idx >= 0);
+      nir_deref_instr *deref = nir_src_as_deref(tex->src[tex_src_idx].src);
+
+      assert(tex->dest.is_ssa);
+      nir_ssa_def *desc = load_resource_deref_desc(b, deref, 0, 4, 32, ctx);
+
+      nir_ssa_def *res;
+      switch (tex->op) {
+      case nir_texop_txs:
+         assert(tex->dest.ssa.num_components <= 3);
+         res = nir_channels(b, desc,
+            nir_component_mask(tex->dest.ssa.num_components));
+         break;
+      case nir_texop_query_levels:
+         assert(tex->dest.ssa.num_components == 1);
+         res = nir_extract_u16(b, nir_channel(b, desc, 3),
+                                  nir_imm_int(b, 0));
+         break;
+      case nir_texop_texture_samples:
+         assert(tex->dest.ssa.num_components == 1);
+         res = nir_extract_u16(b, nir_channel(b, desc, 3),
+                                  nir_imm_int(b, 1));
+         break;
+      default:
+         unreachable("Unsupported texture query op");
+      }
+
+      nir_ssa_def_rewrite_uses(&tex->dest.ssa, res);
+      nir_instr_remove(&tex->instr);
+      return true;
+   }
+
+   int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
    if (sampler_src_idx >= 0) {
       nir_deref_instr *deref = nir_src_as_deref(tex->src[sampler_src_idx].src);
       nir_tex_instr_remove_src(tex, sampler_src_idx);
@@ -420,6 +496,42 @@ get_img_index(nir_builder *b, nir_deref_instr *deref,
    }
 }
 
+static bool
+lower_img_intrinsic(nir_builder *b, nir_intrinsic_instr *intr,
+                    struct apply_descriptors_ctx *ctx)
+{
+   b->cursor = nir_before_instr(&intr->instr);
+   nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
+
+   if (intr->intrinsic == nir_intrinsic_image_deref_size ||
+       intr->intrinsic == nir_intrinsic_image_deref_samples) {
+      assert(intr->dest.is_ssa);
+      nir_ssa_def *desc = load_resource_deref_desc(b, deref, 0, 4, 32, ctx);
+
+      nir_ssa_def *res;
+      switch (intr->intrinsic) {
+      case nir_intrinsic_image_deref_size:
+         res = nir_channels(b, desc,
+            nir_component_mask(intr->dest.ssa.num_components));
+         break;
+      case nir_intrinsic_image_deref_samples:
+         res = nir_extract_u16(b, nir_channel(b, desc, 3),
+                                  nir_imm_int(b, 1));
+         break;
+      default:
+         unreachable("Unsupported image query op");
+      }
+
+      nir_ssa_def_rewrite_uses(&intr->dest.ssa, res);
+      nir_instr_remove(&intr->instr);
+   } else {
+      nir_rewrite_image_intrinsic(intr, get_img_index(b, deref, ctx), false);
+      ctx->has_img_access = true;
+   }
+
+   return true;
+}
+
 static bool
 lower_intrinsic(nir_builder *b, nir_intrinsic_instr *intr,
                 struct apply_descriptors_ctx *ctx)
@@ -445,14 +557,8 @@ lower_intrinsic(nir_builder *b, nir_intrinsic_instr *intr,
    case nir_intrinsic_image_deref_atomic_comp_swap:
    case nir_intrinsic_image_deref_atomic_fadd:
    case nir_intrinsic_image_deref_size:
-   case nir_intrinsic_image_deref_samples: {
-      nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
-
-      b->cursor = nir_before_instr(&intr->instr);
-      nir_rewrite_image_intrinsic(intr, get_img_index(b, deref, ctx), false);
-      ctx->has_img_access = true;
-      return true;
-   }
+   case nir_intrinsic_image_deref_samples:
+      return lower_img_intrinsic(b, intr, ctx);
    default:
       return false;
    }

src/panfrost/vulkan/panvk_vX_shader.c

@@ -396,6 +396,11 @@ panvk_per_arch(shader_create)(struct panvk_device *dev,
       panvk_lower_blend(pdev, nir, &inputs, blend_state, static_blend_constants);
    }
 
+   nir_lower_tex_options lower_tex_options = {
+      .lower_txs_lod = true,
+   };
+   NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);
+
    NIR_PASS_V(nir, panvk_per_arch(nir_lower_descriptors),
               dev, layout, &shader->has_img_access);