mesa/src/intel/compiler/intel_nir_opt_peephole_ffma.c

/*
 * Copyright © 2014 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 "intel_nir.h"
#include "compiler/nir/nir_builder.h"

/*
 * Implements a small peephole optimization that looks for a multiply that
 * is only ever used in an add and replaces both with an fma.
 */

static inline bool
are_all_uses_fadd(nir_def *def)
{
   nir_foreach_use_including_if(use_src, def) {
      if (nir_src_is_if(use_src))
         return false;

      nir_instr *use_instr = nir_src_parent_instr(use_src);
      if (use_instr->type != nir_instr_type_alu)
         return false;

      nir_alu_instr *use_alu = nir_instr_as_alu(use_instr);
      switch (use_alu->op) {
      case nir_op_fadd:
         break; /* This one's ok */

      case nir_op_mov:
      case nir_op_fneg:
      case nir_op_fabs:
         if (!are_all_uses_fadd(&use_alu->def))
            return false;
         break;

      default:
         return false;
      }
   }

   return true;
}

static nir_alu_instr *
get_mul_for_src(nir_alu_src *src, unsigned num_components,
                uint8_t *swizzle, bool *negate, bool *abs)
{
   uint8_t swizzle_tmp[NIR_MAX_VEC_COMPONENTS];

   nir_instr *instr = src->src.ssa->parent_instr;
   if (instr->type != nir_instr_type_alu)
      return NULL;

   nir_alu_instr *alu = nir_instr_as_alu(instr);

   /* We want to bail if any of the other ALU operations involved is labeled
    * exact.  One reason for this is that, while the value that is changing is
    * actually the result of the add and not the multiply, the intention of
    * the user when they specify an exact multiply is that they want *that*
    * value and what they don't care about is the add.  Another reason is that
    * SPIR-V explicitly requires this behaviour.
    */
   if (alu->exact)
      return NULL;

   switch (alu->op) {
   case nir_op_mov:
      alu = get_mul_for_src(&alu->src[0], alu->def.num_components,
                            swizzle, negate, abs);
      break;

   case nir_op_fneg:
      alu = get_mul_for_src(&alu->src[0], alu->def.num_components,
                            swizzle, negate, abs);
      *negate = !*negate;
      break;

   case nir_op_fabs:
      alu = get_mul_for_src(&alu->src[0], alu->def.num_components,
                            swizzle, negate, abs);
      *negate = false;
      *abs = true;
      break;

   case nir_op_fmul:
      /* Only absorb a fmul into a ffma if the fmul is only used in fadd
       * operations.  This prevents us from being too aggressive with our
       * fusing which can actually lead to more instructions.
       */
      if (!are_all_uses_fadd(&alu->def))
         return NULL;
      break;

   default:
      return NULL;
   }

   if (!alu)
      return NULL;

   /* Copy swizzle data before overwriting it to avoid setting a wrong swizzle.
    *
    * Example:
    *   Former swizzle[] = xyzw
    *   src->swizzle[] = zyxx
    *
    *   Expected output swizzle = zyxx
    *   If we reuse swizzle in the loop, then output swizzle would be zyzz.
    */
   memcpy(swizzle_tmp, swizzle, NIR_MAX_VEC_COMPONENTS*sizeof(uint8_t));
   for (int i = 0; i < num_components; i++)
      swizzle[i] = swizzle_tmp[src->swizzle[i]];

   return alu;
}

/**
 * Given a list of (at least two) nir_alu_src's, tells if any of them is a
 * constant value and is used only once.
 */
static bool
any_alu_src_is_a_constant(nir_alu_src srcs[])
{
   for (unsigned i = 0; i < 2; i++) {
      if (srcs[i].src.ssa->parent_instr->type == nir_instr_type_load_const) {
         nir_load_const_instr *load_const =
            nir_instr_as_load_const (srcs[i].src.ssa->parent_instr);

         if (list_is_singular(&load_const->def.uses))
            return true;
      }
   }

   return false;
}

static bool
intel_nir_opt_peephole_ffma_instr(nir_builder *b,
                                  nir_instr *instr,
                                  UNUSED void *cb_data)
{
   if (instr->type != nir_instr_type_alu)
      return false;

   nir_alu_instr *add = nir_instr_as_alu(instr);
   if (add->op != nir_op_fadd)
      return false;

   if (add->exact)
      return false;


   /* This, is the case a + a.  We would rather handle this with an
    * algebraic reduction than fuse it.  Also, we want to only fuse
    * things where the multiply is used only once and, in this case,
    * it would be used twice by the same instruction.
    */
   if (add->src[0].src.ssa == add->src[1].src.ssa)
      return false;

   nir_alu_instr *mul;
   uint8_t add_mul_src, swizzle[NIR_MAX_VEC_COMPONENTS];
   bool negate, abs;
   for (add_mul_src = 0; add_mul_src < 2; add_mul_src++) {
      for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++)
         swizzle[i] = i;

      negate = false;
      abs = false;

      mul = get_mul_for_src(&add->src[add_mul_src],
                            add->def.num_components,
                            swizzle, &negate, &abs);

      if (mul != NULL)
         break;
   }

   if (mul == NULL)
      return false;

   unsigned bit_size = add->def.bit_size;

   nir_def *mul_src[2];
   mul_src[0] = mul->src[0].src.ssa;
   mul_src[1] = mul->src[1].src.ssa;

   /* If any of the operands of the fmul and any of the fadd is a constant,
    * we bypass because it will be more efficient as the constants will be
    * propagated as operands, potentially saving two load_const instructions.
    */
   if (any_alu_src_is_a_constant(mul->src) &&
       any_alu_src_is_a_constant(add->src)) {
      return false;
   }

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

   if (abs) {
      for (unsigned i = 0; i < 2; i++)
         mul_src[i] = nir_fabs(b, mul_src[i]);
   }

   if (negate)
      mul_src[0] = nir_fneg(b, mul_src[0]);

   nir_alu_instr *ffma = nir_alu_instr_create(b->shader, nir_op_ffma);

   for (unsigned i = 0; i < 2; i++) {
      ffma->src[i].src = nir_src_for_ssa(mul_src[i]);
      for (unsigned j = 0; j < add->def.num_components; j++)
         ffma->src[i].swizzle[j] = mul->src[i].swizzle[swizzle[j]];
   }
   nir_alu_src_copy(&ffma->src[2], &add->src[1 - add_mul_src]);

   nir_def_init(&ffma->instr, &ffma->def,
                add->def.num_components, bit_size);
   nir_def_rewrite_uses(&add->def, &ffma->def);

   nir_builder_instr_insert(b, &ffma->instr);
   assert(list_is_empty(&add->def.uses));
   nir_instr_remove(&add->instr);

   return true;
}

bool
intel_nir_opt_peephole_ffma(nir_shader *shader)
{
   return nir_shader_instructions_pass(shader, intel_nir_opt_peephole_ffma_instr,
                                       nir_metadata_block_index |
                                       nir_metadata_dominance,
                                       NULL);
}