tu: Add multiview lowering pass

For now this only handles an a630 quirk where PC_MULTIVIEW_MASK doesn't
exist. However in the future it will also handle multi-position output.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5720>

src/freedreno/vulkan/meson.build

@@ -51,6 +51,7 @@ libtu_files = files(
   'tu_fence.c',
   'tu_formats.c',
   'tu_image.c',
+  'tu_nir_lower_multiview.c',
   'tu_pass.c',
   'tu_pipeline.c',
   'tu_pipeline_cache.c',

src/freedreno/vulkan/tu_device.c

@@ -78,6 +78,7 @@ tu_physical_device_init(struct tu_physical_device *device,
       device->tile_align_w = 32;
       device->magic.PC_UNKNOWN_9805 = 0x0;
       device->magic.SP_UNKNOWN_A0F8 = 0x0;
+      device->supports_multiview_mask = false; /* TODO */
       break;
    case 630:
    case 640:
@@ -86,6 +87,7 @@ tu_physical_device_init(struct tu_physical_device *device,
       device->tile_align_w = 32;
       device->magic.PC_UNKNOWN_9805 = 0x1;
       device->magic.SP_UNKNOWN_A0F8 = 0x1;
+      device->supports_multiview_mask = device->gpu_id != 630;
       break;
    case 650:
       device->ccu_offset_gmem = 0x114000;
@@ -93,6 +95,7 @@ tu_physical_device_init(struct tu_physical_device *device,
       device->tile_align_w = 96;
       device->magic.PC_UNKNOWN_9805 = 0x2;
       device->magic.SP_UNKNOWN_A0F8 = 0x2;
+      device->supports_multiview_mask = true;
       break;
    default:
       result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,

src/freedreno/vulkan/tu_nir_lower_multiview.c

@@ -0,0 +1,76 @@
+/*
+ * Copyright © 2020 Valve Corporation
+ * SPDX-License-Identifier: MIT
+ */
+
+#include "tu_private.h"
+#include "nir_builder.h"
+
+/* Some a6xx variants cannot support a non-contiguous multiview mask. Instead,
+ * inside the shader something like this needs to be inserted:
+ *
+ * gl_Position = ((1ull << gl_ViewIndex) & view_mask) ? gl_Position : vec4(0.);
+ *
+ * Scan backwards until we find the gl_Position write (there should only be
+ * one).
+ */
+static bool
+lower_multiview_mask(nir_function_impl *impl, uint32_t mask)
+{
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   nir_foreach_block_reverse(block, impl) {
+      nir_foreach_instr_reverse(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+         if (intrin->intrinsic != nir_intrinsic_store_deref)
+            continue;
+
+         nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+         if (deref->mode != nir_var_shader_out)
+            continue;
+
+         nir_variable *var = nir_deref_instr_get_variable(deref);
+         if (var->data.location != VARYING_SLOT_POS)
+            continue;
+
+         assert(intrin->src[1].is_ssa);
+         nir_ssa_def *orig_src = intrin->src[1].ssa;
+         b.cursor = nir_before_instr(instr);
+
+         /* ((1ull << gl_ViewIndex) & mask) != 0 */
+         nir_ssa_def *cmp =
+            nir_i2b(&b, nir_iand(&b, nir_imm_int(&b, mask),
+                                  nir_ishl(&b, nir_imm_int(&b, 1),
+                                           nir_load_view_index(&b))));
+
+         nir_ssa_def *src = nir_bcsel(&b, cmp, orig_src, nir_imm_float(&b, 0.));
+         nir_instr_rewrite_src(instr, &intrin->src[1], nir_src_for_ssa(src));
+
+         nir_metadata_preserve(impl, nir_metadata_block_index |
+                                     nir_metadata_dominance);
+         return true;
+      }
+   }
+
+   nir_metadata_preserve(impl, nir_metadata_all);
+   return false;
+}
+
+bool
+tu_nir_lower_multiview(nir_shader *nir, uint32_t mask, struct tu_device *dev)
+{
+   nir_function_impl *entrypoint = nir_shader_get_entrypoint(nir);
+
+   if (!dev->physical_device->supports_multiview_mask &&
+       !util_is_power_of_two_or_zero(mask + 1)) {
+      return lower_multiview_mask(entrypoint, mask);
+   }
+
+   nir_metadata_preserve(entrypoint, nir_metadata_all);
+   return false;
+}
+

src/freedreno/vulkan/tu_private.h

@@ -200,6 +200,7 @@ struct tu_physical_device
    /* gmem store/load granularity */
 #define GMEM_ALIGN_W 16
 #define GMEM_ALIGN_H 4
+   bool supports_multiview_mask;
 
    struct {
       uint32_t PC_UNKNOWN_9805;
@@ -1078,6 +1079,9 @@ struct tu_shader
    uint8_t active_desc_sets;
 };
 
+bool
+tu_nir_lower_multiview(nir_shader *nir, uint32_t mask, struct tu_device *dev);
+
 struct tu_shader *
 tu_shader_create(struct tu_device *dev,
                  gl_shader_stage stage,