src/compiler/nir/nir_lower_wrmasks.c

/*
 * Copyright © 2020 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"

/* A pass to split intrinsics with discontinuous writemasks into ones
 * with contiguous writemasks starting with .x, ie:
 *
 *   vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
 *   intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (2, 0, 4, 0) // wrmask=y
 *
 * is turned into:
 *
 *   vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
 *   vec1 32 ssa_107 = load_const (0x00000001)
 *   vec1 32 ssa_108 = iadd ssa_106, ssa_107
 *   vec1 32 ssa_109 = mov ssa_76.y
 *   intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x
 *
 * and likewise:
 *
 *   vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
 *   intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (15, 0, 4, 0) // wrmask=xzw
 *
 * is split into:
 *
 *   // .x component:
 *   vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35
 *   vec1 32 ssa_107 = load_const (0x00000000)
 *   vec1 32 ssa_108 = iadd ssa_106, ssa_107
 *   vec1 32 ssa_109 = mov ssa_76.x
 *   intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x
 *   // .zw components:
 *   vec1 32 ssa_110 = load_const (0x00000002)
 *   vec1 32 ssa_111 = iadd ssa_106, ssa_110
 *   vec2 32 ssa_112 = mov ssa_76.zw
 *   intrinsic store_ssbo (ssa_112, ssa_105, ssa_111) (3, 0, 4, 0) // wrmask=xy
 */

static int
value_src(nir_intrinsic_op intrinsic)
{
   switch (intrinsic) {
   case nir_intrinsic_store_output:
   case nir_intrinsic_store_per_vertex_output:
   case nir_intrinsic_store_ssbo:
   case nir_intrinsic_store_shared:
   case nir_intrinsic_store_global:
   case nir_intrinsic_store_scratch:
      return 0;
   default:
      return -1;
   }
}

static int
offset_src(nir_intrinsic_op intrinsic)
{
   switch (intrinsic) {
   case nir_intrinsic_store_output:
   case nir_intrinsic_store_shared:
   case nir_intrinsic_store_global:
   case nir_intrinsic_store_scratch:
      return 1;
   case nir_intrinsic_store_per_vertex_output:
   case nir_intrinsic_store_ssbo:
      return 2;
   default:
      return -1;
   }
}

static void
split_wrmask(nir_builder *b, nir_intrinsic_instr *intr)
{
   const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];

   b->cursor = nir_before_instr(&intr->instr);

   assert(!info->has_dest);  /* expecting only store intrinsics */

   unsigned num_srcs   = info->num_srcs;
   unsigned value_idx  = value_src(intr->intrinsic);
   unsigned offset_idx = offset_src(intr->intrinsic);
   unsigned num_comp   = nir_intrinsic_src_components(intr, value_idx);

   unsigned wrmask = nir_intrinsic_write_mask(intr);
   while (wrmask) {
      unsigned first_component = ffs(wrmask) - 1;
      unsigned length = ffs(~(wrmask >> first_component)) - 1;

      nir_ssa_def *value  = nir_ssa_for_src(b, intr->src[value_idx], num_comp);
      nir_ssa_def *offset = nir_ssa_for_src(b, intr->src[offset_idx], 1);

      /* swizzle out the consecutive components that we'll store
       * in this iteration:
       */
      unsigned cur_mask = (BITFIELD_MASK(length) << first_component);
      value = nir_channels(b, value, cur_mask);

      /* and create the replacement intrinsic: */
      nir_intrinsic_instr *new_intr =
            nir_intrinsic_instr_create(b->shader, intr->intrinsic);

      nir_intrinsic_copy_const_indices(new_intr, intr);
      nir_intrinsic_set_write_mask(new_intr, BITFIELD_MASK(length));

      const int offset_units = value->bit_size / 8;

      if (nir_intrinsic_has_align_mul(intr)) {
         assert(nir_intrinsic_has_align_offset(intr));
         unsigned align_mul = nir_intrinsic_align_mul(intr);
         unsigned align_off = nir_intrinsic_align_offset(intr);

         align_off += offset_units * first_component;
         align_off = align_off % align_mul;

         nir_intrinsic_set_align(new_intr, align_mul, align_off);
      }

      /* if the instruction has a BASE, fold the offset adjustment
       * into that instead of adding alu instructions, otherwise add
       * instructions
       */
      unsigned offset_adj = offset_units * first_component;
      if (nir_intrinsic_has_base(intr)) {
         nir_intrinsic_set_base(new_intr,
               nir_intrinsic_base(intr) + offset_adj);
      } else {
         offset = nir_iadd(b, offset,
               nir_imm_intN_t(b, offset_adj, offset->bit_size));
      }

      new_intr->num_components = length;

      /* Copy the sources, replacing value/offset, and passing everything
       * else through to the new instrution:
       */
      for (unsigned i = 0; i < num_srcs; i++) {
         if (i == value_idx) {
            new_intr->src[i] = nir_src_for_ssa(value);
         } else if (i == offset_idx) {
            new_intr->src[i] = nir_src_for_ssa(offset);
         } else {
            new_intr->src[i] = intr->src[i];
         }
      }

      nir_builder_instr_insert(b, &new_intr->instr);

      /* Clear the bits in the writemask that we just wrote, then try
       * again to see if more channels are left.
       */
      wrmask &= ~cur_mask;
   }

   /* Finally remove the original intrinsic. */
   nir_instr_remove(&intr->instr);
}

struct nir_lower_wrmasks_state {
   nir_instr_filter_cb cb;
   const void *data;
};

static bool
nir_lower_wrmasks_instr(nir_builder *b, nir_instr *instr, void *data)
{
   struct nir_lower_wrmasks_state *state = data;

   if (instr->type != nir_instr_type_intrinsic)
      return false;

   nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);

   /* if no wrmask, then skip it: */
   if (!nir_intrinsic_has_write_mask(intr))
      return false;

   /* if wrmask is already contiguous, then nothing to do: */
   if (nir_intrinsic_write_mask(intr) == BITFIELD_MASK(intr->num_components))
      return false;

   /* do we know how to lower this instruction? */
   if (value_src(intr->intrinsic) < 0)
      return false;

   assert(offset_src(intr->intrinsic) >= 0);

   /* does backend need us to lower this intrinsic? */
   if (state->cb && !state->cb(instr, state->data))
      return false;

   split_wrmask(b, intr);

   return true;
}

bool
nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data)
{
   struct nir_lower_wrmasks_state state = {
      .cb = cb,
      .data = data,
   };

   return nir_shader_instructions_pass(shader,
                                       nir_lower_wrmasks_instr,
                                       nir_metadata_block_index |
                                       nir_metadata_dominance,
                                       &state);
}