anv/pipeline: Add shader lowering for multiview
authorJason Ekstrand <jason.ekstrand@intel.com>
Wed, 22 Mar 2017 22:37:17 +0000 (15:37 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Wed, 3 May 2017 18:25:46 +0000 (11:25 -0700)
v2 (Jason Ekstrand):
 - Take a view_mask rather than a whole subpass
 - Build the view mask into the VS shader key

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
src/intel/Makefile.sources
src/intel/vulkan/anv_nir.h
src/intel/vulkan/anv_nir_lower_multiview.c [new file with mode: 0644]
src/intel/vulkan/anv_pipeline.c

index e9a39a659acea014b5066498c3601eea393af916..83200c3d7d87566808151cf83d2081be685d5ddc 100644 (file)
@@ -197,6 +197,7 @@ VULKAN_FILES := \
        vulkan/anv_nir.h \
        vulkan/anv_nir_apply_pipeline_layout.c \
        vulkan/anv_nir_lower_input_attachments.c \
+       vulkan/anv_nir_lower_multiview.c \
        vulkan/anv_nir_lower_push_constants.c \
        vulkan/anv_pass.c \
        vulkan/anv_pipeline.c \
index 3f977017988c4c87308a288b3a777dcfbb703ab3..5b450b45cdf8bbc14021cd63d593ca19abf76d6e 100644 (file)
@@ -35,6 +35,8 @@ void anv_nir_lower_input_attachments(nir_shader *shader);
 
 void anv_nir_lower_push_constants(nir_shader *shader);
 
+bool anv_nir_lower_multiview(nir_shader *shader, uint32_t view_mask);
+
 void anv_nir_apply_pipeline_layout(struct anv_pipeline *pipeline,
                                    nir_shader *shader,
                                    struct brw_stage_prog_data *prog_data,
diff --git a/src/intel/vulkan/anv_nir_lower_multiview.c b/src/intel/vulkan/anv_nir_lower_multiview.c
new file mode 100644 (file)
index 0000000..f40e111
--- /dev/null
@@ -0,0 +1,234 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "anv_nir.h"
+#include "nir/nir_builder.h"
+
+/**
+ * This file implements the lowering required for VK_KHR_multiview.  We
+ * implement multiview using instanced rendering.  The number of instances in
+ * each draw call is multiplied by the number of views in the subpass.  Then,
+ * in the shader, we divide gl_InstanceId by the number of views and use
+ * gl_InstanceId % view_count to compute the actual ViewIndex.
+ */
+
+struct lower_multiview_state {
+   nir_builder builder;
+
+   uint32_t view_mask;
+
+   nir_ssa_def *instance_id;
+   nir_ssa_def *view_index;
+};
+
+static nir_ssa_def *
+build_instance_id(struct lower_multiview_state *state)
+{
+   assert(state->builder.shader->stage == MESA_SHADER_VERTEX);
+
+   if (state->instance_id == NULL) {
+      nir_builder *b = &state->builder;
+
+      b->cursor = nir_before_block(nir_start_block(b->impl));
+
+      /* We use instancing for implementing multiview.  The actual instance id
+       * is given by dividing instance_id by the number of views in this
+       * subpass.
+       */
+      state->instance_id =
+         nir_idiv(b, nir_load_instance_id(b),
+                     nir_imm_int(b, _mesa_bitcount(state->view_mask)));
+   }
+
+   return state->instance_id;
+}
+
+static nir_ssa_def *
+build_view_index(struct lower_multiview_state *state)
+{
+   if (state->view_index == NULL) {
+      nir_builder *b = &state->builder;
+
+      b->cursor = nir_before_block(nir_start_block(b->impl));
+
+      assert(state->view_mask != 0);
+      if (0 && _mesa_bitcount(state->view_mask) == 1) {
+         state->view_index = nir_imm_int(b, ffs(state->view_mask) - 1);
+      } else if (state->builder.shader->stage == MESA_SHADER_VERTEX) {
+         /* We only support 16 viewports */
+         assert((state->view_mask & 0xffff0000) == 0);
+
+         /* We use instancing for implementing multiview.  The compacted view
+          * id is given by instance_id % view_count.  We then have to convert
+          * that to an actual view id.
+          */
+         nir_ssa_def *compacted =
+            nir_umod(b, nir_load_instance_id(b),
+                        nir_imm_int(b, _mesa_bitcount(state->view_mask)));
+
+         if (0 && util_is_power_of_two(state->view_mask + 1)) {
+            /* If we have a full view mask, then compacted is what we want */
+            state->view_index = compacted;
+         } else {
+            /* Now we define a map from compacted view index to the actual
+             * view index that's based on the view_mask.  The map is given by
+             * 16 nibbles, each of which is a value from 0 to 15.
+             */
+            uint64_t remap = 0;
+            uint32_t bit, i = 0;
+            for_each_bit(bit, state->view_mask) {
+               assert(bit < 16);
+               remap |= (uint64_t)bit << (i++ * 4);
+            }
+
+            nir_ssa_def *shift = nir_imul(b, compacted, nir_imm_int(b, 4));
+
+            /* One of these days, when we have int64 everywhere, this will be
+             * easier.
+             */
+            nir_ssa_def *shifted;
+            if (remap <= UINT32_MAX) {
+               shifted = nir_ushr(b, nir_imm_int(b, remap), shift);
+            } else {
+               nir_ssa_def *shifted_low =
+                  nir_ushr(b, nir_imm_int(b, remap), shift);
+               nir_ssa_def *shifted_high =
+                  nir_ushr(b, nir_imm_int(b, remap >> 32),
+                              nir_isub(b, shift, nir_imm_int(b, 32)));
+               shifted = nir_bcsel(b, nir_ilt(b, shift, nir_imm_int(b, 32)),
+                                      shifted_low, shifted_high);
+            }
+            state->view_index = nir_iand(b, shifted, nir_imm_int(b, 0xf));
+         }
+      } else {
+         const struct glsl_type *type = glsl_int_type();
+         if (b->shader->stage == MESA_SHADER_TESS_CTRL ||
+             b->shader->stage == MESA_SHADER_GEOMETRY)
+            type = glsl_array_type(type, 1);
+
+         nir_variable *idx_var =
+            nir_variable_create(b->shader, nir_var_shader_in,
+                                type, "view index");
+         idx_var->data.location = VARYING_SLOT_VIEW_INDEX;
+         if (b->shader->stage == MESA_SHADER_FRAGMENT)
+            idx_var->data.interpolation = INTERP_MODE_FLAT;
+
+         if (glsl_type_is_array(type)) {
+            nir_deref_var *deref = nir_deref_var_create(b->shader, idx_var);
+            nir_deref_array *arr = nir_deref_array_create(b->shader);
+            arr->deref.type = glsl_int_type();
+            arr->deref_array_type = nir_deref_array_type_direct;
+            arr->base_offset = 0;
+            deref->deref.child = &arr->deref;
+
+            state->view_index = nir_load_deref_var(b, deref);
+         } else {
+            state->view_index = nir_load_var(b, idx_var);
+         }
+      }
+   }
+
+   return state->view_index;
+}
+
+bool
+anv_nir_lower_multiview(nir_shader *shader, uint32_t view_mask)
+{
+   assert(shader->stage != MESA_SHADER_COMPUTE);
+
+   /* If multiview isn't enabled, we have nothing to do. */
+   if (view_mask == 0)
+      return false;
+
+   struct lower_multiview_state state = {
+      .view_mask = view_mask,
+   };
+
+   /* This pass assumes a single entrypoint */
+   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
+
+   nir_builder_init(&state.builder, entrypoint);
+
+   bool progress = false;
+   nir_foreach_block(block, entrypoint) {
+      nir_foreach_instr_safe(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
+
+         if (load->intrinsic != nir_intrinsic_load_instance_id &&
+             load->intrinsic != nir_intrinsic_load_view_index)
+            continue;
+
+         assert(load->dest.is_ssa);
+
+         nir_ssa_def *value;
+         if (load->intrinsic == nir_intrinsic_load_instance_id) {
+            value = build_instance_id(&state);
+         } else {
+            assert(load->intrinsic == nir_intrinsic_load_view_index);
+            value = build_view_index(&state);
+         }
+
+         nir_ssa_def_rewrite_uses(&load->dest.ssa, nir_src_for_ssa(value));
+
+         nir_instr_remove(&load->instr);
+         progress = true;
+      }
+   }
+
+   /* The view index is available in all stages but the instance id is only
+    * available in the VS.  If it's not a fragment shader, we need to pass
+    * the view index on to the next stage.
+    */
+   if (shader->stage != MESA_SHADER_FRAGMENT) {
+      nir_ssa_def *view_index = build_view_index(&state);
+
+      nir_builder *b = &state.builder;
+
+      assert(view_index->parent_instr->block == nir_start_block(entrypoint));
+      b->cursor = nir_after_instr(view_index->parent_instr);
+
+      nir_variable *view_index_out =
+         nir_variable_create(shader, nir_var_shader_out,
+                             glsl_int_type(), "view index");
+      view_index_out->data.location = VARYING_SLOT_VIEW_INDEX;
+      nir_store_var(b, view_index_out, view_index, 0x1);
+
+      nir_variable *layer_id_out =
+         nir_variable_create(shader, nir_var_shader_out,
+                             glsl_int_type(), "layer ID");
+      layer_id_out->data.location = VARYING_SLOT_LAYER;
+      nir_store_var(b, layer_id_out, view_index, 0x1);
+
+      progress = true;
+   }
+
+   if (progress) {
+      nir_metadata_preserve(entrypoint, nir_metadata_block_index |
+                                        nir_metadata_dominance);
+   }
+
+   return progress;
+}
index 85b3290e7b4c7959cc98d5fc655720418ac58b2d..643db9cb9f2b5c57aea2f74d748d6b9bf275f3bf 100644 (file)
@@ -339,6 +339,10 @@ anv_pipeline_hash_shader(struct anv_pipeline *pipeline,
    struct mesa_sha1 ctx;
 
    _mesa_sha1_init(&ctx);
+   if (stage != MESA_SHADER_COMPUTE) {
+      _mesa_sha1_update(&ctx, &pipeline->subpass->view_mask,
+                        sizeof(pipeline->subpass->view_mask));
+   }
    if (pipeline->layout) {
       _mesa_sha1_update(&ctx, pipeline->layout->sha1,
                         sizeof(pipeline->layout->sha1));
@@ -372,6 +376,9 @@ anv_pipeline_compile(struct anv_pipeline *pipeline,
 
    NIR_PASS_V(nir, anv_nir_lower_push_constants);
 
+   if (stage != MESA_SHADER_COMPUTE)
+      NIR_PASS_V(nir, anv_nir_lower_multiview, pipeline->subpass->view_mask);
+
    nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 
    /* Figure out the number of parameters */