nir: save IO semantics in lowered IO intrinsics

This enables drivers and utils to get all IO information from intrinsics,
so that they don't have to walk the complex types of NIR variables to find
out other information about IO intrinsics.

NIR in/out variables can be removed after nir_lower_io. We could remove
the variables in the pass, but for now I just decided to remove
the variables in radeonsi before shaders are returned to st/mesa.
(st/mesa just needs adjustments to work without NIR in/out variables)

Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6442>

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 49d61484d73e2d24936909e1e7a1eae460c703a2..8371ae9e6dc34cb6c3e3219c99d2c16c58256b55 100644 (file)
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -1499,7 +1499,7 @@ typedef struct {
 
 #include "nir_intrinsics.h"
 
-#define NIR_INTRINSIC_MAX_CONST_INDEX 4
+#define NIR_INTRINSIC_MAX_CONST_INDEX 5
 
 /** Represents an intrinsic
  *
@@ -1744,10 +1744,24 @@ typedef enum {
     */
    NIR_INTRINSIC_EXECUTION_SCOPE,
 
+   /**
+    * Value of nir_io_semantics.
+    */
+   NIR_INTRINSIC_IO_SEMANTICS,
+
    NIR_INTRINSIC_NUM_INDEX_FLAGS,
 
 } nir_intrinsic_index_flag;
 
+typedef struct {
+   unsigned location:7; /* gl_vert_attrib, gl_varying_slot, or gl_frag_result */
+   unsigned num_slots:6;  /* max 32, may be pessimistic with const indexing */
+   unsigned dual_source_blend_index:1;
+   unsigned fb_fetch_output:1; /* for GL_KHR_blend_equation_advanced */
+   unsigned gs_streams:8; /* xxyyzzww: 2-bit stream index for each component */
+   unsigned _pad:9;
+} nir_io_semantics;
+
 #define NIR_INTRINSIC_MAX_INPUTS 5
 
 typedef struct {
@@ -1917,6 +1931,30 @@ nir_intrinsic_align(const nir_intrinsic_instr *intrin)
    return align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
 }
 
+static inline void
+nir_intrinsic_set_io_semantics(nir_intrinsic_instr *intrin,
+                               nir_io_semantics semantics)
+{
+   const nir_intrinsic_info *info = &nir_intrinsic_infos[intrin->intrinsic];
+   assert(info->index_map[NIR_INTRINSIC_IO_SEMANTICS] > 0);
+   STATIC_ASSERT(sizeof(nir_io_semantics) == sizeof(intrin->const_index[0]));
+   semantics._pad = 0; /* clear padding bits */
+   memcpy(&intrin->const_index[info->index_map[NIR_INTRINSIC_IO_SEMANTICS] - 1],
+          &semantics, sizeof(semantics));
+}
+
+static inline nir_io_semantics
+nir_intrinsic_io_semantics(const nir_intrinsic_instr *intrin)
+{
+   const nir_intrinsic_info *info = &nir_intrinsic_infos[intrin->intrinsic];
+   assert(info->index_map[NIR_INTRINSIC_IO_SEMANTICS] > 0);
+   nir_io_semantics semantics;
+   memcpy(&semantics,
+          &intrin->const_index[info->index_map[NIR_INTRINSIC_IO_SEMANTICS] - 1],
+          sizeof(semantics));
+   return semantics;
+}
+
 unsigned
 nir_image_intrinsic_coord_components(const nir_intrinsic_instr *instr);
 

diff --git a/src/compiler/nir/nir_intrinsics.py b/src/compiler/nir/nir_intrinsics.py
index 8d00446a2a24c18d70fe37403a8824b52d203008..e9cc0da6472150e98f28b6675cea13f218b11320 100644 (file)
--- a/src/compiler/nir/nir_intrinsics.py
+++ b/src/compiler/nir/nir_intrinsics.py
@@ -134,6 +134,7 @@ MEMORY_MODES = "NIR_INTRINSIC_MEMORY_MODES"
 MEMORY_SCOPE = "NIR_INTRINSIC_MEMORY_SCOPE"
 # Scope of a control barrier
 EXECUTION_SCOPE = "NIR_INTRINSIC_EXECUTION_SCOPE"
+IO_SEMANTICS = "NIR_INTRINSIC_IO_SEMANTICS"
 
 #
 # Possible flags:
@@ -738,22 +739,22 @@ load("ubo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], flags=[CAN_ELIMINATE, CA
 # src[] = { buffer_index, offset in vec4 units }
 load("ubo_vec4", [-1, 1], [ACCESS, COMPONENT], flags=[CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
-load("input", [1], [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
+load("input", [1], [BASE, COMPONENT, TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { vertex_id, offset }.
-load("input_vertex", [1, 1], [BASE, COMPONENT, TYPE], [CAN_ELIMINATE, CAN_REORDER])
+load("input_vertex", [1, 1], [BASE, COMPONENT, TYPE, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { vertex, offset }.
-load("per_vertex_input", [1, 1], [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
+load("per_vertex_input", [1, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { barycoord, offset }.
-load("interpolated_input", [2, 1], [BASE, COMPONENT], [CAN_ELIMINATE, CAN_REORDER])
+load("interpolated_input", [2, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE, CAN_REORDER])
 
 # src[] = { buffer_index, offset }.
 load("ssbo", [-1, 1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { buffer_index }
 load("ssbo_address", [1], [], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
-load("output", [1], [BASE, COMPONENT], flags=[CAN_ELIMINATE])
+load("output", [1], [BASE, COMPONENT, IO_SEMANTICS], flags=[CAN_ELIMINATE])
 # src[] = { vertex, offset }.
-load("per_vertex_output", [1, 1], [BASE, COMPONENT], [CAN_ELIMINATE])
+load("per_vertex_output", [1, 1], [BASE, COMPONENT, IO_SEMANTICS], [CAN_ELIMINATE])
 # src[] = { offset }.
 load("shared", [1], [BASE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { offset }.
@@ -777,9 +778,9 @@ def store(name, srcs, indices=[], flags=[]):
     intrinsic("store_" + name, [0] + srcs, indices=indices, flags=flags)
 
 # src[] = { value, offset }.
-store("output", [1], [BASE, WRMASK, COMPONENT, TYPE])
+store("output", [1], [BASE, WRMASK, COMPONENT, TYPE, IO_SEMANTICS])
 # src[] = { value, vertex, offset }.
-store("per_vertex_output", [1, 1], [BASE, WRMASK, COMPONENT])
+store("per_vertex_output", [1, 1], [BASE, WRMASK, COMPONENT, IO_SEMANTICS])
 # src[] = { value, block_index, offset }
 store("ssbo", [-1, 1], [WRMASK, ACCESS, ALIGN_MUL, ALIGN_OFFSET])
 # src[] = { value, offset }.

diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c
index 2e17ad70b36b2464d7820d89e40387de7ab050de..7f20b96fb0c07dfe2f07475c2f54103149e207fa 100644 (file)
--- a/src/compiler/nir/nir_lower_io.c
+++ b/src/compiler/nir/nir_lower_io.c
@@ -159,6 +159,19 @@ nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage)
    return false;
 }
 
+static unsigned get_number_of_slots(struct lower_io_state *state,
+                                    const nir_variable *var)
+{
+   const struct glsl_type *type = var->type;
+
+   if (nir_is_per_vertex_io(var, state->builder.shader->info.stage)) {
+      assert(glsl_type_is_array(type));
+      type = glsl_get_array_element(type);
+   }
+
+   return state->type_size(type, var->data.bindless);
+}
+
 static nir_ssa_def *
 get_io_offset(nir_builder *b, nir_deref_instr *deref,
               nir_ssa_def **vertex_index,
@@ -291,6 +304,14 @@ emit_load(struct lower_io_state *state,
        load->intrinsic == nir_intrinsic_load_uniform)
       nir_intrinsic_set_type(load, type);
 
+   if (load->intrinsic != nir_intrinsic_load_uniform) {
+      nir_io_semantics semantics = {0};
+      semantics.location = var->data.location;
+      semantics.num_slots = get_number_of_slots(state, var);
+      semantics.fb_fetch_output = var->data.fb_fetch_output;
+      nir_intrinsic_set_io_semantics(load, semantics);
+   }
+
    if (vertex_index) {
       load->src[0] = nir_src_for_ssa(vertex_index);
       load->src[1] = nir_src_for_ssa(offset);
@@ -393,6 +414,25 @@ emit_store(struct lower_io_state *state, nir_ssa_def *data,
 
    store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
 
+   unsigned gs_streams = 0;
+   if (state->builder.shader->info.stage == MESA_SHADER_GEOMETRY) {
+      if (var->data.stream & NIR_STREAM_PACKED) {
+         gs_streams = var->data.stream & ~NIR_STREAM_PACKED;
+      } else {
+         assert(var->data.stream < 4);
+         gs_streams = 0;
+         for (unsigned i = 0; i < num_components; ++i)
+            gs_streams |= var->data.stream << (2 * i);
+      }
+   }
+
+   nir_io_semantics semantics = {0};
+   semantics.location = var->data.location;
+   semantics.num_slots = get_number_of_slots(state, var);
+   semantics.dual_source_blend_index = var->data.index;
+   semantics.gs_streams = gs_streams;
+   nir_intrinsic_set_io_semantics(store, semantics);
+
    nir_builder_instr_insert(b, &store->instr);
 }
 
@@ -519,6 +559,11 @@ lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
    nir_intrinsic_set_base(load, var->data.driver_location);
    nir_intrinsic_set_component(load, component);
 
+   nir_io_semantics semantics = {0};
+   semantics.location = var->data.location;
+   semantics.num_slots = get_number_of_slots(state, var);
+   nir_intrinsic_set_io_semantics(load, semantics);
+
    load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
    load->src[1] = nir_src_for_ssa(offset);