anv/nir: Add a pass for applying a applying a pipeline layout to a shader
authorJason Ekstrand <jason.ekstrand@intel.com>
Wed, 14 Oct 2015 23:39:58 +0000 (16:39 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Thu, 15 Oct 2015 01:38:33 +0000 (18:38 -0700)
This new pass lowers the _vk intrinsics which take a (set, binding, index)
tripple to the single-index non-vk intrinsics based on the pipeline layout.

src/glsl/nir/nir.h
src/glsl/nir/nir_lower_samplers.c
src/mesa/drivers/dri/i965/brw_nir.c
src/vulkan/Makefile.am
src/vulkan/anv_compiler.cpp
src/vulkan/anv_nir.h
src/vulkan/anv_nir_apply_pipeline_layout.c [new file with mode: 0644]

index f2794cb7afb4484b27c1c59445b5afa2c086e6d3..f7b9483d74ae15c790a4953c0a8d28fb95949c43 100644 (file)
@@ -1916,7 +1916,6 @@ void nir_lower_phis_to_scalar(nir_shader *shader);
 
 void nir_lower_samplers(nir_shader *shader,
                         const struct gl_shader_program *shader_program);
-void nir_lower_samplers_for_vk(nir_shader *shader);
 
 bool nir_lower_system_values(nir_shader *shader);
 
index aaeff2c223ef5ab33000eea614929b7821f9462e..5df79a69a065bbaf9c5c53ab9af40fc657656ea6 100644 (file)
 #include "program/prog_parameter.h"
 #include "program/program.h"
 
-static void
-add_indirect_to_tex(nir_tex_instr *instr, nir_src indirect)
-{
-   /* First, we have to resize the array of texture sources */
-   nir_tex_src *new_srcs = rzalloc_array(instr, nir_tex_src,
-                                         instr->num_srcs + 1);
-
-   for (unsigned i = 0; i < instr->num_srcs; i++) {
-      new_srcs[i].src_type = instr->src[i].src_type;
-      nir_instr_move_src(&instr->instr, &new_srcs[i].src, &instr->src[i].src);
-   }
-
-   ralloc_free(instr->src);
-   instr->src = new_srcs;
-
-   /* Now we can go ahead and move the source over to being a
-    * first-class texture source.
-    */
-   instr->src[instr->num_srcs].src_type = nir_tex_src_sampler_offset;
-   instr->num_srcs++;
-   nir_instr_rewrite_src(&instr->instr, &instr->src[instr->num_srcs - 1].src,
-                         indirect);
-}
-
 /* Calculate the sampler index based on array indicies and also
  * calculate the base uniform location for struct members.
  */
@@ -210,49 +186,3 @@ nir_lower_samplers(nir_shader *shader,
          lower_impl(overload->impl, shader_program, shader->stage);
    }
 }
-
-static bool
-lower_samplers_for_vk_block(nir_block *block, void *data)
-{
-   nir_foreach_instr(block, instr) {
-      if (instr->type != nir_instr_type_tex)
-         continue;
-
-      nir_tex_instr *tex = nir_instr_as_tex(instr);
-
-      assert(tex->sampler);
-
-      tex->sampler_set = tex->sampler->var->data.descriptor_set;
-      tex->sampler_index = tex->sampler->var->data.binding;
-
-      if (tex->sampler->deref.child) {
-         assert(tex->sampler->deref.child->deref_type == nir_deref_type_array);
-         nir_deref_array *arr = nir_deref_as_array(tex->sampler->deref.child);
-
-         /* Only one-level arrays are allowed in vulkan */
-         assert(arr->deref.child == NULL);
-
-         tex->sampler_index += arr->base_offset;
-         if (arr->deref_array_type == nir_deref_array_type_indirect) {
-            add_indirect_to_tex(tex, arr->indirect);
-            nir_instr_rewrite_src(instr, &arr->indirect, NIR_SRC_INIT);
-
-            tex->sampler_array_size = glsl_get_length(tex->sampler->deref.type);
-         }
-      }
-
-      tex->sampler = NULL;
-   }
-
-   return true;
-}
-
-void
-nir_lower_samplers_for_vk(nir_shader *shader)
-{
-   nir_foreach_overload(shader, overload) {
-      if (overload->impl) {
-         nir_foreach_block(overload->impl, lower_samplers_for_vk_block, NULL);
-      }
-   }
-}
index 19206600e6476fd4b8b4c8fbc72edef311cc8dec..0a9c09f10756bc6d2a086d02bb4b1136c92e4f89 100644 (file)
@@ -181,10 +181,8 @@ brw_process_nir(nir_shader *nir,
 
    if (shader_prog) {
       nir_lower_samplers(nir, shader_prog);
-   } else {
-      nir_lower_samplers_for_vk(nir);
+      nir_validate_shader(nir);
    }
-   nir_validate_shader(nir);
 
    nir_lower_system_values(nir);
    nir_validate_shader(nir);
index a691329065cd9d626662605409b1a0f56b6c4706..aeed78ae840a616a311fed0e5870cdfc77e8c0e3 100644 (file)
@@ -70,6 +70,7 @@ VULKAN_SOURCES =                                        \
        anv_intel.c                                     \
        anv_meta.c                                      \
        anv_nir_apply_dynamic_offsets.c                 \
+       anv_nir_apply_pipeline_layout.c                 \
        anv_pipeline.c                                  \
        anv_private.h                                   \
        anv_query.c                                     \
index bf931fe886f53be895bba449c45dc16a1cd91401..34023a7369f45fad4e5685206892b086fdc759a5 100644 (file)
@@ -285,6 +285,7 @@ really_do_vs_prog(struct brw_context *brw,
    create_params_array(pipeline, vs, &prog_data->base.base);
    anv_nir_apply_dynamic_offsets(pipeline, vs->Program->nir,
                                  &prog_data->base.base);
+   anv_nir_apply_pipeline_layout(vs->Program->nir, pipeline->layout);
 
    GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
    prog_data->inputs_read = vp->program.Base.InputsRead;
@@ -571,6 +572,7 @@ really_do_wm_prog(struct brw_context *brw,
 
    create_params_array(pipeline, fs, &prog_data->base);
    anv_nir_apply_dynamic_offsets(pipeline, fs->Program->nir, &prog_data->base);
+   anv_nir_apply_pipeline_layout(fs->Program->nir, pipeline->layout);
 
    prog_data->barycentric_interp_modes =
       brw_compute_barycentric_interp_modes(brw->intelScreen->devinfo,
@@ -888,6 +890,7 @@ brw_codegen_cs_prog(struct brw_context *brw,
 
    create_params_array(pipeline, cs, &prog_data->base);
    anv_nir_apply_dynamic_offsets(pipeline, cs->Program->nir, &prog_data->base);
+   anv_nir_apply_pipeline_layout(cs->Program->nir, pipeline->layout);
 
    program = brw_cs_emit(brw, mem_ctx, key, prog_data,
                          &cp->program, prog, -1, &program_size);
index 1fd3484e0595776752cc47b2f0562bb6010d5a2d..af95e3a8849434a94e536157b51adafbea268b71 100644 (file)
@@ -40,6 +40,8 @@ anv_vk_shader_stage_for_mesa_stage(gl_shader_stage stage)
 void anv_nir_apply_dynamic_offsets(struct anv_pipeline *pipeline,
                                    nir_shader *shader,
                                    struct brw_stage_prog_data *prog_data);
+bool anv_nir_apply_pipeline_layout(nir_shader *shader,
+                                   const struct anv_pipeline_layout *layout);
 
 #ifdef __cplusplus
 }
diff --git a/src/vulkan/anv_nir_apply_pipeline_layout.c b/src/vulkan/anv_nir_apply_pipeline_layout.c
new file mode 100644 (file)
index 0000000..dea2dee
--- /dev/null
@@ -0,0 +1,217 @@
+/*
+ * Copyright © 2015 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 "glsl/nir/nir_builder.h"
+
+struct apply_pipeline_layout_state {
+   nir_shader *shader;
+   nir_builder builder;
+
+   VkShaderStage stage;
+   const struct anv_pipeline_layout *layout;
+
+   bool progress;
+};
+
+static nir_intrinsic_op
+lowered_op(nir_intrinsic_op op)
+{
+   switch (op) {
+   case nir_intrinsic_load_ubo_vk:
+      return nir_intrinsic_load_ubo;
+   case nir_intrinsic_load_ubo_vk_indirect:
+      return nir_intrinsic_load_ubo_indirect;
+   case nir_intrinsic_load_ssbo_vk:
+      return nir_intrinsic_load_ssbo;
+   case nir_intrinsic_load_ssbo_vk_indirect:
+      return nir_intrinsic_load_ssbo_indirect;
+   case nir_intrinsic_store_ssbo_vk:
+      return nir_intrinsic_store_ssbo;
+   case nir_intrinsic_store_ssbo_vk_indirect:
+      return nir_intrinsic_store_ssbo_indirect;
+   default:
+      unreachable("Invalid intrinsic for lowering");
+   }
+}
+
+static uint32_t
+get_surface_index(unsigned set, unsigned binding,
+                  struct apply_pipeline_layout_state *state)
+{
+   assert(set < state->layout->num_sets);
+   struct anv_descriptor_set_layout *set_layout =
+      state->layout->set[set].layout;
+
+   assert(binding < set_layout->binding_count);
+
+   assert(set_layout->binding[binding].stage[state->stage].surface_index >= 0);
+
+   uint32_t surface_index =
+      state->layout->set[set].stage[state->stage].surface_start +
+      set_layout->binding[binding].stage[state->stage].surface_index;
+
+   assert(surface_index < state->layout->stage[state->stage].surface_count);
+
+   return surface_index;
+}
+
+static bool
+try_lower_intrinsic(nir_intrinsic_instr *intrin,
+                    struct apply_pipeline_layout_state *state)
+{
+   nir_builder *b = &state->builder;
+
+   int block_idx_src;
+   switch (intrin->intrinsic) {
+   case nir_intrinsic_load_ubo_vk:
+   case nir_intrinsic_load_ubo_vk_indirect:
+   case nir_intrinsic_load_ssbo_vk:
+   case nir_intrinsic_load_ssbo_vk_indirect:
+      block_idx_src = 0;
+      break;
+   case nir_intrinsic_store_ssbo_vk:
+   case nir_intrinsic_store_ssbo_vk_indirect:
+      block_idx_src = 1;
+      break;
+   default:
+      return false;
+   }
+
+   b->cursor = nir_before_instr(&intrin->instr);
+
+   uint32_t set = intrin->const_index[0];
+   uint32_t binding = intrin->const_index[1];
+
+   uint32_t surface_index = get_surface_index(set, binding, state);
+
+   nir_const_value *const_block_idx =
+      nir_src_as_const_value(intrin->src[block_idx_src]);
+
+   nir_ssa_def *block_index;
+   if (const_block_idx) {
+      block_index = nir_imm_int(b, surface_index + const_block_idx->u[0]);
+   } else {
+      block_index = nir_iadd(b, nir_imm_int(b, surface_index),
+                             nir_ssa_for_src(b, intrin->src[block_idx_src], 1));
+   }
+
+   nir_instr_rewrite_src(&intrin->instr, &intrin->src[block_idx_src],
+                         nir_src_for_ssa(block_index));
+
+   intrin->intrinsic = lowered_op(intrin->intrinsic);
+   /* Shift the offset indices down */
+   intrin->const_index[0] = intrin->const_index[2];
+   intrin->const_index[1] = intrin->const_index[3];
+
+   return true;
+}
+
+static void
+lower_tex(nir_tex_instr *tex, struct apply_pipeline_layout_state *state)
+{
+   /* No one should have come by and lowered it already */
+   assert(tex->sampler);
+
+   unsigned set = tex->sampler->var->data.descriptor_set;
+   unsigned binding = tex->sampler->var->data.binding;
+
+   tex->sampler_index = get_surface_index(set, binding, state);
+
+   if (tex->sampler->deref.child) {
+      assert(tex->sampler->deref.child->deref_type == nir_deref_type_array);
+      nir_deref_array *deref_array =
+         nir_deref_as_array(tex->sampler->deref.child);
+
+      tex->sampler_index += deref_array->base_offset;
+
+      if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
+         nir_tex_src *new_srcs = rzalloc_array(tex, nir_tex_src,
+                                               tex->num_srcs + 1);
+
+         for (unsigned i = 0; i < tex->num_srcs; i++) {
+            new_srcs[i].src_type = tex->src[i].src_type;
+            nir_instr_move_src(&tex->instr, &new_srcs[i].src, &tex->src[i].src);
+         }
+
+         ralloc_free(tex->src);
+         tex->src = new_srcs;
+
+         /* Now we can go ahead and move the source over to being a
+          * first-class texture source.
+          */
+         tex->src[tex->num_srcs].src_type = nir_tex_src_sampler_offset;
+         tex->num_srcs++;
+         nir_instr_move_src(&tex->instr, &tex->src[tex->num_srcs - 1].src,
+                            &deref_array->indirect);
+      }
+   }
+
+   tex->sampler = NULL;
+}
+
+static bool
+apply_pipeline_layout_block(nir_block *block, void *void_state)
+{
+   struct apply_pipeline_layout_state *state = void_state;
+
+   nir_foreach_instr_safe(block, instr) {
+      switch (instr->type) {
+      case nir_instr_type_intrinsic:
+         if (try_lower_intrinsic(nir_instr_as_intrinsic(instr), state))
+            state->progress = true;
+         break;
+      case nir_instr_type_tex:
+         lower_tex(nir_instr_as_tex(instr), state);
+         /* All texture instructions need lowering */
+         state->progress = true;
+         break;
+      default:
+         continue;
+      }
+   }
+
+   return true;
+}
+
+bool
+anv_nir_apply_pipeline_layout(nir_shader *shader,
+                              const struct anv_pipeline_layout *layout)
+{
+   struct apply_pipeline_layout_state state = {
+      .shader = shader,
+      .stage = anv_vk_shader_stage_for_mesa_stage(shader->stage),
+      .layout = layout,
+   };
+
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl) {
+         nir_builder_init(&state.builder, overload->impl);
+         nir_foreach_block(overload->impl, apply_pipeline_layout_block, &state);
+         nir_metadata_preserve(overload->impl, nir_metadata_block_index |
+                                               nir_metadata_dominance);
+      }
+   }
+
+   return state.progress;
+}