nir: add pass to lower fb reads

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>

src/compiler/nir/meson.build

@@ -123,6 +123,7 @@ files_libnir = files(
   'nir_lower_constant_initializers.c',
   'nir_lower_double_ops.c',
   'nir_lower_drawpixels.c',
+  'nir_lower_fb_read.c',
   'nir_lower_fragcoord_wtrans.c',
   'nir_lower_frexp.c',
   'nir_lower_global_vars_to_local.c',

src/compiler/nir/nir.h

@@ -1444,7 +1444,8 @@ typedef enum {
    nir_texop_txl,                /**< Texture look-up with explicit LOD */
    nir_texop_txd,                /**< Texture look-up with partial derivatives */
    nir_texop_txf,                /**< Texel fetch with explicit LOD */
-   nir_texop_txf_ms,                /**< Multisample texture fetch */
+   nir_texop_txf_ms,             /**< Multisample texture fetch */
+   nir_texop_txf_ms_fb,          /**< Multisample texture fetch from framebuffer */
    nir_texop_txf_ms_mcs,         /**< Multisample compression value fetch */
    nir_texop_txs,                /**< Texture size */
    nir_texop_lod,                /**< Texture lod query */
@@ -1579,6 +1580,7 @@ nir_tex_instr_is_query(const nir_tex_instr *instr)
    case nir_texop_txd:
    case nir_texop_txf:
    case nir_texop_txf_ms:
+   case nir_texop_txf_ms_fb:
    case nir_texop_tg4:
       return false;
    default:
@@ -1618,6 +1620,7 @@ nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src)
       switch (instr->op) {
       case nir_texop_txf:
       case nir_texop_txf_ms:
+      case nir_texop_txf_ms_fb:
       case nir_texop_txf_ms_mcs:
       case nir_texop_samples_identical:
          return nir_type_int;
@@ -3394,6 +3397,8 @@ bool nir_lower_wpos_ytransform(nir_shader *shader,
                                const nir_lower_wpos_ytransform_options *options);
 bool nir_lower_wpos_center(nir_shader *shader, const bool for_sample_shading);
 
+bool nir_lower_fb_read(nir_shader *shader);
+
 typedef struct nir_lower_drawpixels_options {
    gl_state_index16 texcoord_state_tokens[STATE_LENGTH];
    gl_state_index16 scale_state_tokens[STATE_LENGTH];

src/compiler/nir/nir_lower_fb_read.c

@@ -0,0 +1,121 @@
+/*
+ * Copyright © 2019 Google, Inc.
+ *
+ * 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 "nir.h"
+#include "nir_builder.h"
+
+/* Lowing for fragment shader load_output.
+ *
+ * This pass supports the blend_equation_advanced, where a fragment
+ * shader loads the output (fragcolor) to read the current framebuffer.
+ * It does this by lowering the output read to a txf_ms_fb instruction.
+ * This instruction works similarly to a normal txf_ms except without
+ * taking a texture source argument.  (The driver backend is expected
+ * to wire this up to a free texture slot which is configured to read
+ * from the framebuffer.)
+ *
+ * This should be run after lower_wpos_ytransform, because the tex
+ * coordinates should be the physical fragcoord, not the logical
+ * y-flipped coord.
+ *
+ * Note that this pass explicitly does *not* add a sampler uniform
+ * (as txf_ms_fb does not reference a texture).  The driver backend
+ * is going to want nif->info.num_textures to include the count of
+ * number of textures *not* including the one it inserts to sample
+ * from the framebuffer, so it more easily knows where to insert the
+ * hidden texture to read from the fb.
+ */
+
+static void
+lower_fb_read(nir_builder *b, nir_intrinsic_instr *intr)
+{
+   b->cursor = nir_before_instr(&intr->instr);
+
+   nir_ssa_def *fragcoord = nir_load_frag_coord(b);
+   nir_ssa_def *sampid = nir_load_sample_id(b);
+
+   fragcoord = nir_f2i32(b, fragcoord);
+
+   nir_tex_instr *tex = nir_tex_instr_create(b->shader, 2);
+   tex->op = nir_texop_txf_ms_fb;
+   tex->sampler_dim = GLSL_SAMPLER_DIM_2D;
+   tex->coord_components = 2;
+   tex->dest_type = nir_type_float;
+   tex->src[0].src_type = nir_tex_src_coord;
+   tex->src[0].src = nir_src_for_ssa(nir_channels(b, fragcoord, 0x3));
+   tex->src[1].src_type = nir_tex_src_ms_index;
+   tex->src[1].src = nir_src_for_ssa(sampid);
+
+   nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, NULL);
+   nir_builder_instr_insert(b, &tex->instr);
+
+   nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(&tex->dest.ssa));
+}
+
+bool
+nir_lower_fb_read(nir_shader *shader)
+{
+   bool progress = false;
+
+   assert(shader->info.stage == MESA_SHADER_FRAGMENT);
+
+   nir_foreach_function(function, shader) {
+      nir_function_impl *impl = function->impl;
+
+      if (!impl)
+         continue;
+
+      nir_foreach_block(block, impl) {
+         nir_foreach_instr_safe(instr, block) {
+            if (instr->type != nir_instr_type_intrinsic)
+               continue;
+
+            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+            if (intr->intrinsic != nir_intrinsic_load_output)
+               continue;
+
+            /* TODO KHR_blend_equation_advanced is limited to non-MRT
+             * scenarios.. but possible there are other extensions
+             * where this pass would be useful that do support MRT?
+             *
+             * I guess for now I'll leave that as an exercise for the
+             * reader.
+             */
+            if (nir_intrinsic_base(intr) != 0 ||
+                nir_src_as_uint(intr->src[0]) != 0)
+               continue;
+
+            nir_builder b;
+            nir_builder_init(&b, impl);
+            lower_fb_read(&b, intr);
+            progress = true;
+         }
+      }
+
+      nir_metadata_preserve(impl, nir_metadata_block_index |
+                                  nir_metadata_dominance);
+
+   }
+
+   return progress;
+}

src/compiler/nir/nir_print.c

@@ -883,6 +883,9 @@ print_tex_instr(nir_tex_instr *instr, print_state *state)
    case nir_texop_txf_ms:
       fprintf(fp, "txf_ms ");
       break;
+   case nir_texop_txf_ms_fb:
+      fprintf(fp, "txf_ms_fb ");
+      break;
    case nir_texop_txf_ms_mcs:
       fprintf(fp, "txf_ms_mcs ");
       break;
@@ -995,12 +998,14 @@ print_tex_instr(nir_tex_instr *instr, print_state *state)
       fprintf(fp, " } (offsets)");
    }
 
-   if (!has_texture_deref) {
-      fprintf(fp, ", %u (texture)", instr->texture_index);
-   }
+   if (instr->op != nir_texop_txf_ms_fb) {
+      if (!has_texture_deref) {
+         fprintf(fp, ", %u (texture)", instr->texture_index);
+      }
 
-   if (!has_sampler_deref) {
-      fprintf(fp, ", %u (sampler)", instr->sampler_index);
+      if (!has_sampler_deref) {
+         fprintf(fp, ", %u (sampler)", instr->sampler_index);
+      }
    }
 }
 

src/compiler/spirv/spirv_to_nir.c

@@ -2200,6 +2200,9 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    case nir_texop_samples_identical:
       /* These don't */
       break;
+   case nir_texop_txf_ms_fb:
+      vtn_fail("unexpected nir_texop_txf_ms_fb");
+      break;
    case nir_texop_txf_ms_mcs:
       vtn_fail("unexpected nir_texop_txf_ms_mcs");
    }