radv,aco: add a workaround for binding 2D views of a 3D image on GFX9

The hardware can't bind a slice of a 3D image as a 2D image, GFX10+
introduced ARRAY_PITCH to allow this without a shader workaround.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16294>

src/amd/compiler/aco_instruction_selection.cpp

@@ -5890,6 +5890,22 @@ get_image_coords(isel_context* ctx, const nir_intrinsic_instr* instr)
          coords[i] = emit_extract_vector(ctx, src0, i, v1);
    }
 
+   if (ctx->options->key.image_2d_view_of_3d &&
+       dim == GLSL_SAMPLER_DIM_2D && !is_array) {
+      /* The hw can't bind a slice of a 3D image as a 2D image, because it
+       * ignores BASE_ARRAY if the target is 3D. The workaround is to read
+       * BASE_ARRAY and set it as the 3rd address operand for all 2D images.
+       */
+      assert(ctx->options->chip_class == GFX9);
+      Temp rsrc = bld.as_uniform(get_ssa_temp(ctx, instr->src[0].ssa));
+      Temp rsrc_word5 = emit_extract_vector(ctx, rsrc, 5, v1);
+      /* Extract the BASE_ARRAY field [0:12] from the descriptor. */
+      Temp first_layer =
+         bld.vop3(aco_opcode::v_bfe_u32, bld.def(v1), rsrc_word5,
+                  Operand::c32(0u), Operand::c32(13u));
+      coords.emplace_back(first_layer);
+   }
+
    if (instr->intrinsic == nir_intrinsic_bindless_image_load ||
        instr->intrinsic == nir_intrinsic_bindless_image_sparse_load ||
        instr->intrinsic == nir_intrinsic_bindless_image_store) {

src/amd/vulkan/radv_device.c

@@ -3124,6 +3124,7 @@ radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCr
    bool image_float32_atomics = false;
    bool vs_prologs = false;
    bool global_bo_list = false;
+   bool image_2d_view_of_3d = false;
 
    /* Check enabled features */
    if (pCreateInfo->pEnabledFeatures) {
@@ -3190,6 +3191,12 @@ radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCr
             global_bo_list = true;
          break;
       }
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_2D_VIEW_OF_3D_FEATURES_EXT: {
+         const VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *features = (const void *)ext;
+         if (features->image2DViewOf3D)
+            image_2d_view_of_3d = true;
+         break;
+      }
       default:
          break;
       }
@@ -3252,6 +3259,8 @@ radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCr
 
    device->image_float32_atomics = image_float32_atomics;
 
+   device->image_2d_view_of_3d = image_2d_view_of_3d;
+
    radv_init_shader_arenas(device);
 
    device->overallocation_disallowed = overallocation_disallowed;

src/amd/vulkan/radv_pipeline.c

@@ -3061,6 +3061,9 @@ radv_generate_pipeline_key(const struct radv_pipeline *pipeline, VkPipelineCreat
    key.disable_aniso_single_level = device->instance->disable_aniso_single_level &&
                                     device->physical_device->rad_info.chip_class < GFX8;
 
+   key.image_2d_view_of_3d = device->image_2d_view_of_3d &&
+                             device->physical_device->rad_info.chip_class == GFX9;
+
    return key;
 }
 

src/amd/vulkan/radv_private.h

@@ -839,6 +839,9 @@ struct radv_device {
    /* Whether shader image 32-bit float atomics are enabled. */
    bool image_float32_atomics;
 
+   /* Whether 2D views of 3D image is enabled. */
+   bool image_2d_view_of_3d;
+
    /* Whether anisotropy is forced with RADV_TEX_ANISO (-1 is disabled). */
    int force_aniso;
 

src/amd/vulkan/radv_shader.h

@@ -65,6 +65,7 @@ struct radv_pipeline_key {
    uint32_t adjust_frag_coord_z : 1;
    uint32_t disable_aniso_single_level : 1;
    uint32_t disable_sinking_load_input_fs : 1;
+   uint32_t image_2d_view_of_3d : 1;
 
    struct {
       uint32_t instance_rate_inputs;