nir/lower_io: Add "explicit" IO lowering
authorJason Ekstrand <jason.ekstrand@intel.com>
Wed, 28 Nov 2018 03:31:42 +0000 (21:31 -0600)
committerJason Ekstrand <jason@jlekstrand.net>
Tue, 8 Jan 2019 00:38:30 +0000 (00:38 +0000)
This new pass is for lowering explicitly laid out memory coming in from
SPIR-V or a similar source.  It's quite a bit more complicated than the
normal lower_io because we have to be able to handle matrices.  The
way the stride information is stored for matrices is awkward and dealing
with row-major matrices is especially painful.

Reviewed-by: Alejandro PiƱeiro <apinheiro@igalia.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
src/compiler/nir/nir.h
src/compiler/nir/nir_lower_io.c

index 29b4da22330de8d10e8d4dd065ebeaa57f9ef611..553410b92d16e22f419ddb9149aa4b23da0713e5 100644 (file)
@@ -2896,6 +2896,18 @@ bool nir_lower_io(nir_shader *shader,
                   nir_variable_mode modes,
                   int (*type_size)(const struct glsl_type *),
                   nir_lower_io_options);
+
+typedef enum {
+   /**
+    * An address format which is comprised of a vec2 where the first
+    * component is a vulkan descriptor index and the second is an offset.
+    */
+   nir_address_format_vk_index_offset,
+} nir_address_format;
+bool nir_lower_explicit_io(nir_shader *shader,
+                           nir_variable_mode modes,
+                           nir_address_format);
+
 nir_src *nir_get_io_offset_src(nir_intrinsic_instr *instr);
 nir_src *nir_get_io_vertex_index_src(nir_intrinsic_instr *instr);
 
index 2ccba8c032be0a23c29eeda43b339d43ac7a59a1..bcbfebdfa3b75eedc89e32355c1d5630f4289398 100644 (file)
@@ -528,6 +528,411 @@ nir_lower_io(nir_shader *shader, nir_variable_mode modes,
    return progress;
 }
 
+static unsigned
+type_scalar_size_bytes(const struct glsl_type *type)
+{
+   assert(glsl_type_is_vector_or_scalar(type) ||
+          glsl_type_is_matrix(type));
+   return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
+}
+
+static nir_ssa_def *
+build_addr_iadd(nir_builder *b, nir_ssa_def *addr,
+                nir_address_format addr_format, nir_ssa_def *offset)
+{
+   assert(offset->num_components == 1);
+   assert(addr->bit_size == offset->bit_size);
+
+   switch (addr_format) {
+   case nir_address_format_vk_index_offset:
+      assert(addr->num_components == 2);
+      return nir_vec2(b, nir_channel(b, addr, 0),
+                         nir_iadd(b, nir_channel(b, addr, 1), offset));
+   }
+   unreachable("Invalid address format");
+}
+
+static nir_ssa_def *
+build_addr_iadd_imm(nir_builder *b, nir_ssa_def *addr,
+                    nir_address_format addr_format, int64_t offset)
+{
+   return build_addr_iadd(b, addr, addr_format,
+                             nir_imm_intN_t(b, offset, addr->bit_size));
+}
+
+static nir_ssa_def *
+addr_to_index(nir_builder *b, nir_ssa_def *addr,
+              nir_address_format addr_format)
+{
+   assert(addr_format == nir_address_format_vk_index_offset);
+   assert(addr->num_components == 2);
+   return nir_channel(b, addr, 0);
+}
+
+static nir_ssa_def *
+addr_to_offset(nir_builder *b, nir_ssa_def *addr,
+               nir_address_format addr_format)
+{
+   assert(addr_format == nir_address_format_vk_index_offset);
+   assert(addr->num_components == 2);
+   return nir_channel(b, addr, 1);
+}
+
+static nir_ssa_def *
+build_explicit_io_load(nir_builder *b, nir_intrinsic_instr *intrin,
+                       nir_ssa_def *addr, nir_address_format addr_format,
+                       unsigned num_components)
+{
+   nir_variable_mode mode = nir_src_as_deref(intrin->src[0])->mode;
+
+   nir_intrinsic_op op;
+   switch (mode) {
+   case nir_var_ubo:
+      op = nir_intrinsic_load_ubo;
+      break;
+   case nir_var_ssbo:
+      op = nir_intrinsic_load_ssbo;
+      break;
+   default:
+      unreachable("Unsupported explicit IO variable mode");
+   }
+
+   nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
+
+   load->src[0] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
+   load->src[1] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
+
+   if (mode != nir_var_ubo)
+      nir_intrinsic_set_access(load, nir_intrinsic_access(intrin));
+
+   /* TODO: We should try and provide a better alignment.  For OpenCL, we need
+    * to plumb the alignment through from SPIR-V when we have one.
+    */
+   nir_intrinsic_set_align(load, intrin->dest.ssa.bit_size / 8, 0);
+
+   assert(intrin->dest.is_ssa);
+   load->num_components = num_components;
+   nir_ssa_dest_init(&load->instr, &load->dest, num_components,
+                     intrin->dest.ssa.bit_size, intrin->dest.ssa.name);
+   nir_builder_instr_insert(b, &load->instr);
+
+   return &load->dest.ssa;
+}
+
+static void
+build_explicit_io_store(nir_builder *b, nir_intrinsic_instr *intrin,
+                        nir_ssa_def *addr, nir_address_format addr_format,
+                        nir_ssa_def *value, nir_component_mask_t write_mask)
+{
+   nir_variable_mode mode = nir_src_as_deref(intrin->src[0])->mode;
+
+   nir_intrinsic_op op;
+   switch (mode) {
+   case nir_var_ssbo:
+      op = nir_intrinsic_store_ssbo;
+      break;
+   default:
+      unreachable("Unsupported explicit IO variable mode");
+   }
+
+   nir_intrinsic_instr *store = nir_intrinsic_instr_create(b->shader, op);
+
+   store->src[0] = nir_src_for_ssa(value);
+   store->src[1] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
+   store->src[2] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
+
+   nir_intrinsic_set_write_mask(store, write_mask);
+
+   nir_intrinsic_set_access(store, nir_intrinsic_access(intrin));
+
+   /* TODO: We should try and provide a better alignment.  For OpenCL, we need
+    * to plumb the alignment through from SPIR-V when we have one.
+    */
+   nir_intrinsic_set_align(store, value->bit_size / 8, 0);
+
+   assert(value->num_components == 1 ||
+          value->num_components == intrin->num_components);
+   store->num_components = value->num_components;
+   nir_builder_instr_insert(b, &store->instr);
+}
+
+static nir_ssa_def *
+build_explicit_io_atomic(nir_builder *b, nir_intrinsic_instr *intrin,
+                         nir_ssa_def *addr, nir_address_format addr_format)
+{
+   nir_variable_mode mode = nir_src_as_deref(intrin->src[0])->mode;
+   const unsigned num_data_srcs =
+      nir_intrinsic_infos[intrin->intrinsic].num_srcs - 1;
+
+   nir_intrinsic_op op;
+   switch (mode) {
+   case nir_var_ssbo:
+      switch (intrin->intrinsic) {
+#define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_ssbo_##O; break;
+      OP(atomic_exchange)
+      OP(atomic_comp_swap)
+      OP(atomic_add)
+      OP(atomic_imin)
+      OP(atomic_umin)
+      OP(atomic_imax)
+      OP(atomic_umax)
+      OP(atomic_and)
+      OP(atomic_or)
+      OP(atomic_xor)
+      OP(atomic_fadd)
+      OP(atomic_fmin)
+      OP(atomic_fmax)
+      OP(atomic_fcomp_swap)
+#undef OP
+      default:
+         unreachable("Invalid SSBO atomic");
+      }
+      break;
+   default:
+      unreachable("Unsupported explicit IO variable mode");
+   }
+
+   nir_intrinsic_instr *atomic = nir_intrinsic_instr_create(b->shader, op);
+
+   atomic->src[0] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
+   atomic->src[1] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
+   for (unsigned i = 0; i < num_data_srcs; i++) {
+      assert(intrin->src[1 + i].is_ssa);
+      atomic->src[2 + i] = nir_src_for_ssa(intrin->src[1 + i].ssa);
+   }
+
+   assert(intrin->dest.ssa.num_components == 1);
+   nir_ssa_dest_init(&atomic->instr, &atomic->dest,
+                     1, intrin->dest.ssa.bit_size, intrin->dest.ssa.name);
+   nir_builder_instr_insert(b, &atomic->instr);
+
+   return &atomic->dest.ssa;
+}
+
+static void
+lower_explicit_io_deref(nir_builder *b, nir_deref_instr *deref,
+                        nir_address_format addr_format)
+{
+   /* Just delete the deref if it's not used.  We can't use
+    * nir_deref_instr_remove_if_unused here because it may remove more than
+    * one deref which could break our list walking since we walk the list
+    * backwards.
+    */
+   assert(list_empty(&deref->dest.ssa.if_uses));
+   if (list_empty(&deref->dest.ssa.uses)) {
+      nir_instr_remove(&deref->instr);
+      return;
+   }
+
+   b->cursor = nir_after_instr(&deref->instr);
+
+   /* Var derefs must be lowered away by the driver */
+   assert(deref->deref_type != nir_deref_type_var);
+
+   assert(deref->parent.is_ssa);
+   nir_ssa_def *parent_addr = deref->parent.ssa;
+
+   nir_ssa_def *addr;
+   assert(deref->dest.is_ssa);
+   switch (deref->deref_type) {
+   case nir_deref_type_var:
+      unreachable("Must be lowered by the driver");
+      break;
+
+   case nir_deref_type_array: {
+      nir_deref_instr *parent = nir_deref_instr_parent(deref);
+
+      unsigned stride = glsl_get_explicit_stride(parent->type);
+      if ((glsl_type_is_matrix(parent->type) &&
+           glsl_matrix_type_is_row_major(parent->type)) ||
+          (glsl_type_is_vector(parent->type) && stride == 0))
+         stride = type_scalar_size_bytes(parent->type);
+
+      assert(stride > 0);
+
+      nir_ssa_def *index = nir_ssa_for_src(b, deref->arr.index, 1);
+      index = nir_i2i(b, index, parent_addr->bit_size);
+      addr = build_addr_iadd(b, parent_addr, addr_format,
+                                nir_imul_imm(b, index, stride));
+      break;
+   }
+
+   case nir_deref_type_ptr_as_array: {
+      nir_ssa_def *index = nir_ssa_for_src(b, deref->arr.index, 1);
+      index = nir_i2i(b, index, parent_addr->bit_size);
+      unsigned stride = nir_deref_instr_ptr_as_array_stride(deref);
+      addr = build_addr_iadd(b, parent_addr, addr_format,
+                                nir_imul_imm(b, index, stride));
+      break;
+   }
+
+   case nir_deref_type_array_wildcard:
+      unreachable("Wildcards should be lowered by now");
+      break;
+
+   case nir_deref_type_struct: {
+      nir_deref_instr *parent = nir_deref_instr_parent(deref);
+      int offset = glsl_get_struct_field_offset(parent->type,
+                                                deref->strct.index);
+      assert(offset >= 0);
+      addr = build_addr_iadd_imm(b, parent_addr, addr_format, offset);
+      break;
+   }
+
+   case nir_deref_type_cast:
+      /* Nothing to do here */
+      addr = parent_addr;
+      break;
+   }
+
+   nir_instr_remove(&deref->instr);
+   nir_ssa_def_rewrite_uses(&deref->dest.ssa, nir_src_for_ssa(addr));
+}
+
+static void
+lower_explicit_io_access(nir_builder *b, nir_intrinsic_instr *intrin,
+                         nir_address_format addr_format)
+{
+   b->cursor = nir_after_instr(&intrin->instr);
+
+   nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+   unsigned vec_stride = glsl_get_explicit_stride(deref->type);
+   unsigned scalar_size = type_scalar_size_bytes(deref->type);
+   assert(vec_stride == 0 || glsl_type_is_vector(deref->type));
+   assert(vec_stride == 0 || vec_stride >= scalar_size);
+
+   nir_ssa_def *addr = &deref->dest.ssa;
+   if (intrin->intrinsic == nir_intrinsic_load_deref) {
+      nir_ssa_def *value;
+      if (vec_stride > scalar_size) {
+         nir_ssa_def *comps[4] = { NULL, };
+         for (unsigned i = 0; i < intrin->num_components; i++) {
+            nir_ssa_def *comp_addr = build_addr_iadd_imm(b, addr, addr_format,
+                                                         vec_stride * i);
+            comps[i] = build_explicit_io_load(b, intrin, comp_addr,
+                                              addr_format, 1);
+         }
+         value = nir_vec(b, comps, intrin->num_components);
+      } else {
+         value = build_explicit_io_load(b, intrin, addr, addr_format,
+                                        intrin->num_components);
+      }
+      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(value));
+   } else if (intrin->intrinsic == nir_intrinsic_store_deref) {
+      assert(intrin->src[1].is_ssa);
+      nir_ssa_def *value = intrin->src[1].ssa;
+      nir_component_mask_t write_mask = nir_intrinsic_write_mask(intrin);
+      if (vec_stride > scalar_size) {
+         for (unsigned i = 0; i < intrin->num_components; i++) {
+            if (!(write_mask & (1 << i)))
+               continue;
+
+            nir_ssa_def *comp_addr = build_addr_iadd_imm(b, addr, addr_format,
+                                                         vec_stride * i);
+            build_explicit_io_store(b, intrin, comp_addr, addr_format,
+                                    nir_channel(b, value, i), 1);
+         }
+      } else {
+         build_explicit_io_store(b, intrin, addr, addr_format,
+                                 value, write_mask);
+      }
+   } else {
+      nir_ssa_def *value =
+         build_explicit_io_atomic(b, intrin, addr, addr_format);
+      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(value));
+   }
+
+   nir_instr_remove(&intrin->instr);
+}
+
+static bool
+nir_lower_explicit_io_impl(nir_function_impl *impl, nir_variable_mode modes,
+                           nir_address_format addr_format)
+{
+   bool progress = false;
+
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   /* Walk in reverse order so that we can see the full deref chain when we
+    * lower the access operations.  We lower them assuming that the derefs
+    * will be turned into address calculations later.
+    */
+   nir_foreach_block_reverse(block, impl) {
+      nir_foreach_instr_reverse_safe(instr, block) {
+         switch (instr->type) {
+         case nir_instr_type_deref: {
+            nir_deref_instr *deref = nir_instr_as_deref(instr);
+            if (deref->mode & modes) {
+               lower_explicit_io_deref(&b, deref, addr_format);
+               progress = true;
+            }
+            break;
+         }
+
+         case nir_instr_type_intrinsic: {
+            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+            switch (intrin->intrinsic) {
+            case nir_intrinsic_load_deref:
+            case nir_intrinsic_store_deref:
+            case nir_intrinsic_deref_atomic_add:
+            case nir_intrinsic_deref_atomic_imin:
+            case nir_intrinsic_deref_atomic_umin:
+            case nir_intrinsic_deref_atomic_imax:
+            case nir_intrinsic_deref_atomic_umax:
+            case nir_intrinsic_deref_atomic_and:
+            case nir_intrinsic_deref_atomic_or:
+            case nir_intrinsic_deref_atomic_xor:
+            case nir_intrinsic_deref_atomic_exchange:
+            case nir_intrinsic_deref_atomic_comp_swap:
+            case nir_intrinsic_deref_atomic_fadd:
+            case nir_intrinsic_deref_atomic_fmin:
+            case nir_intrinsic_deref_atomic_fmax:
+            case nir_intrinsic_deref_atomic_fcomp_swap: {
+               nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+               if (deref->mode & modes) {
+                  lower_explicit_io_access(&b, intrin, addr_format);
+                  progress = true;
+               }
+               break;
+            }
+
+            default:
+               break;
+            }
+            break;
+         }
+
+         default:
+            /* Nothing to do */
+            break;
+         }
+      }
+   }
+
+   if (progress) {
+      nir_metadata_preserve(impl, nir_metadata_block_index |
+                                  nir_metadata_dominance);
+   }
+
+   return progress;
+}
+
+bool
+nir_lower_explicit_io(nir_shader *shader, nir_variable_mode modes,
+                      nir_address_format addr_format)
+{
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl &&
+          nir_lower_explicit_io_impl(function->impl, modes, addr_format))
+         progress = true;
+   }
+
+   return progress;
+}
+
 /**
  * Return the offset source for a load/store intrinsic.
  */