intel/compiler/brw_opt_bank_conflicts.cpp

/*
 * Copyright © 2017 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.
 */

/** @file
 *
 * This file contains a GRF bank conflict mitigation pass.  The pass is
 * intended to be run after register allocation and works by rearranging the
 * layout of the GRF space (without altering the semantics of the program) in
 * a way that minimizes the number of GRF bank conflicts incurred by ternary
 * instructions.
 *
 * Unfortunately there is close to no information about bank conflicts in the
 * hardware spec, but experimentally on Gfx7-Gfx9 ternary instructions seem to
 * incur an average bank conflict penalty of one cycle per SIMD8 op whenever
 * the second and third source are stored in the same GRF bank (\sa bank_of()
 * for the exact bank layout) which cannot be fetched during the same cycle by
 * the EU, unless the EU logic manages to optimize out the read cycle of a
 * duplicate source register (\sa is_conflict_optimized_out()).
 *
 * The asymptotic run-time of the algorithm is dominated by the
 * shader_conflict_weight_matrix() computation below, which is O(n) on the
 * number of instructions in the program, however for small and medium-sized
 * programs the run-time is likely to be dominated by
 * optimize_reg_permutation() which is O(m^3) on the number of GRF atoms of
 * the program (\sa partitioning), which is bounded (since the program uses a
 * bounded number of registers post-regalloc) and of the order of 100.  For
 * that reason optimize_reg_permutation() is vectorized in order to keep the
 * cubic term within reasonable bounds for m close to its theoretical maximum.
 */

#include "brw_fs.h"
#include "brw_cfg.h"

#ifdef __SSE2__

#include <emmintrin.h>

/**
 * Thin layer around vector intrinsics so they can be easily replaced with
 * e.g. the fall-back scalar path, an implementation with different vector
 * width or using different SIMD architectures (AVX-512?!).
 *
 * This implementation operates on pairs of independent SSE2 integer vectors à
 * la SIMD16 for somewhat improved throughput.  SSE2 is supported by virtually
 * all platforms that care about bank conflicts, so this path should almost
 * always be available in practice.
 */
namespace {
   /**
    * SIMD integer vector data type.
    */
   struct vector_type {
      __m128i v[2];
   };

   /**
    * Scalar data type matching the representation of a single component of \p
    * vector_type.
    */
   typedef int16_t scalar_type;

   /**
    * Maximum integer value representable as a \p scalar_type.
    */
   const scalar_type max_scalar = INT16_MAX;

   /**
    * Number of components of a \p vector_type.
    */
   const unsigned vector_width = 2 * sizeof(__m128i) / sizeof(scalar_type);

   /**
    * Set the i-th component of vector \p v to \p x.
    */
   void
   set(vector_type &v, unsigned i, scalar_type x)
   {
      assert(i < vector_width);
      memcpy((char *)v.v + i * sizeof(x), &x, sizeof(x));
   }

   /**
    * Get the i-th component of vector \p v.
    */
   scalar_type
   get(const vector_type &v, unsigned i)
   {
      assert(i < vector_width);
      scalar_type x;
      memcpy(&x, (char *)v.v + i * sizeof(x), sizeof(x));
      return x;
   }

   /**
    * Add two vectors with saturation.
    */
   vector_type
   adds(const vector_type &v, const vector_type &w)
   {
      const vector_type u = {{
            _mm_adds_epi16(v.v[0], w.v[0]),
            _mm_adds_epi16(v.v[1], w.v[1])
         }};
      return u;
   }

   /**
    * Subtract two vectors with saturation.
    */
   vector_type
   subs(const vector_type &v, const vector_type &w)
   {
      const vector_type u = {{
            _mm_subs_epi16(v.v[0], w.v[0]),
            _mm_subs_epi16(v.v[1], w.v[1])
         }};
      return u;
   }

   /**
    * Compute the bitwise conjunction of two vectors.
    */
   vector_type
   mask(const vector_type &v, const vector_type &w)
   {
      const vector_type u = {{
            _mm_and_si128(v.v[0], w.v[0]),
            _mm_and_si128(v.v[1], w.v[1])
         }};
      return u;
   }

   /**
    * Reduce the components of a vector using saturating addition.
    */
   scalar_type
   sums(const vector_type &v)
   {
      const __m128i v8 = _mm_adds_epi16(v.v[0], v.v[1]);
      const __m128i v4 = _mm_adds_epi16(v8, _mm_shuffle_epi32(v8, 0x4e));
      const __m128i v2 = _mm_adds_epi16(v4, _mm_shuffle_epi32(v4, 0xb1));
      const __m128i v1 = _mm_adds_epi16(v2, _mm_shufflelo_epi16(v2, 0xb1));
      return _mm_extract_epi16(v1, 0);
   }
}

#else

/**
 * Thin layer around vector intrinsics so they can be easily replaced with
 * e.g. the fall-back scalar path, an implementation with different vector
 * width or using different SIMD architectures (AVX-512?!).
 *
 * This implementation operates on scalar values and doesn't rely on
 * any vector extensions.  This is mainly intended for debugging and
 * to keep this file building on exotic platforms.
 */
namespace {
   /**
    * SIMD integer vector data type.
    */
   typedef int16_t vector_type;

   /**
    * Scalar data type matching the representation of a single component of \p
    * vector_type.
    */
   typedef int16_t scalar_type;

   /**
    * Maximum integer value representable as a \p scalar_type.
    */
   const scalar_type max_scalar = INT16_MAX;

   /**
    * Number of components of a \p vector_type.
    */
   const unsigned vector_width = 1;

   /**
    * Set the i-th component of vector \p v to \p x.
    */
   void
   set(vector_type &v, unsigned i, scalar_type x)
   {
      assert(i < vector_width);
      v = x;
   }

   /**
    * Get the i-th component of vector \p v.
    */
   scalar_type
   get(const vector_type &v, unsigned i)
   {
      assert(i < vector_width);
      return v;
   }

   /**
    * Add two vectors with saturation.
    */
   vector_type
   adds(vector_type v, vector_type w)
   {
      return MAX2(INT16_MIN, MIN2(INT16_MAX, int(v) + w));
   }

   /**
    * Subtract two vectors with saturation.
    */
   vector_type
   subs(vector_type v, vector_type w)
   {
      return MAX2(INT16_MIN, MIN2(INT16_MAX, int(v) - w));
   }

   /**
    * Compute the bitwise conjunction of two vectors.
    */
   vector_type
   mask(vector_type v, vector_type w)
   {
      return v & w;
   }

   /**
    * Reduce the components of a vector using saturating addition.
    */
   scalar_type
   sums(vector_type v)
   {
      return v;
   }
}

#endif

/**
 * Swap \p x and \p y.
 */
#define SWAP(x, y) do {                          \
      __typeof(y) _swap_tmp = y;                 \
      y = x;                                     \
      x = _swap_tmp;                             \
   } while (0)

namespace {
   /**
    * Variable-length vector type intended to represent cycle-count costs for
    * arbitrary atom-to-bank assignments.  It's indexed by a pair of integers
    * (i, p), where i is an atom index and p in {0, 1} indicates the parity of
    * the conflict (respectively, whether the cost is incurred whenever the
    * atoms are assigned the same bank b or opposite-parity banks b and b^1).
    * \sa shader_conflict_weight_matrix()
    */
   struct weight_vector_type {
      weight_vector_type() : v(NULL), size(0) {}

      weight_vector_type(unsigned n) : v(alloc(n)), size(n) {}

      weight_vector_type(const weight_vector_type &u) :
         v(alloc(u.size)), size(u.size)
      {
         memcpy(v, u.v,
                DIV_ROUND_UP(u.size, vector_width) * sizeof(vector_type));
      }

      ~weight_vector_type()
      {
         free(v);
      }

      weight_vector_type &
      operator=(weight_vector_type u)
      {
         SWAP(v, u.v);
         SWAP(size, u.size);
         return *this;
      }

      vector_type *v;
      unsigned size;

   private:
      static vector_type *
      alloc(unsigned n)
      {
         const unsigned align = MAX2(sizeof(void *), __alignof__(vector_type));
         const unsigned size = DIV_ROUND_UP(n, vector_width) * sizeof(vector_type);
         void *p;
         if (posix_memalign(&p, align, size))
            return NULL;
         memset(p, 0, size);
         return reinterpret_cast<vector_type *>(p);
      }
   };

   /**
    * Set the (i, p)-th component of weight vector \p v to \p x.
    */
   void
   set(weight_vector_type &v, unsigned i, unsigned p, scalar_type x)
   {
      set(v.v[(2 * i + p) / vector_width], (2 * i + p) % vector_width, x);
   }

   /**
    * Get the (i, p)-th component of weight vector \p v.
    */
   scalar_type
   get(const weight_vector_type &v, unsigned i, unsigned p)
   {
      return get(v.v[(2 * i + p) / vector_width], (2 * i + p) % vector_width);
   }

   /**
    * Swap the (i, p)-th and (j, q)-th components of weight vector \p v.
    */
   void
   swap(weight_vector_type &v,
        unsigned i, unsigned p,
        unsigned j, unsigned q)
   {
      const scalar_type tmp = get(v, i, p);
      set(v, i, p, get(v, j, q));
      set(v, j, q, tmp);
   }
}

namespace {
   /**
    * Object that represents the partitioning of an arbitrary register space
    * into indivisible units (referred to as atoms below) that can potentially
    * be rearranged independently from other registers.  The partitioning is
    * inferred from a number of contiguity requirements specified using
    * require_contiguous().  This allows efficient look-up of the atom index a
    * given register address belongs to, or conversely the range of register
    * addresses that belong to a given atom.
    */
   struct partitioning {
      /**
       * Create a (for the moment unrestricted) partitioning of a register
       * file of size \p n.  The units are arbitrary.
       */
      partitioning(unsigned n) :
         max_reg(n),
         offsets(new unsigned[n + num_terminator_atoms]),
         atoms(new unsigned[n + num_terminator_atoms])
      {
         for (unsigned i = 0; i < n + num_terminator_atoms; i++) {
            offsets[i] = i;
            atoms[i] = i;
         }
      }

      partitioning(const partitioning &p) :
         max_reg(p.max_reg),
         offsets(new unsigned[p.num_atoms() + num_terminator_atoms]),
         atoms(new unsigned[p.max_reg + num_terminator_atoms])
      {
         memcpy(offsets, p.offsets,
                sizeof(unsigned) * (p.num_atoms() + num_terminator_atoms));
         memcpy(atoms, p.atoms,
                sizeof(unsigned) * (p.max_reg + num_terminator_atoms));
      }

      ~partitioning()
      {
         delete[] offsets;
         delete[] atoms;
      }

      partitioning &
      operator=(partitioning p)
      {
         SWAP(max_reg, p.max_reg);
         SWAP(offsets, p.offsets);
         SWAP(atoms, p.atoms);
         return *this;
      }

      /**
       * Require register range [reg, reg + n[ to be considered part of the
       * same atom.
       */
      void
      require_contiguous(unsigned reg, unsigned n)
      {
         unsigned r = atoms[reg];

         /* Renumber atoms[reg...] = { r... } and their offsets[r...] for the
          * case that the specified contiguity requirement leads to the fusion
          * (yay) of one or more existing atoms.
          */
         for (unsigned reg1 = reg + 1; reg1 <= max_reg; reg1++) {
            if (offsets[atoms[reg1]] < reg + n) {
               atoms[reg1] = r;
            } else {
               if (offsets[atoms[reg1 - 1]] != offsets[atoms[reg1]])
                  r++;

               offsets[r] = offsets[atoms[reg1]];
               atoms[reg1] = r;
            }
         }
      }

      /**
       * Get the atom index register address \p reg belongs to.
       */
      unsigned
      atom_of_reg(unsigned reg) const
      {
         return atoms[reg];
      }

      /**
       * Get the base register address that belongs to atom \p r.
       */
      unsigned
      reg_of_atom(unsigned r) const
      {
         return offsets[r];
      }

      /**
       * Get the size of atom \p r in register address units.
       */
      unsigned
      size_of_atom(unsigned r) const
      {
         assert(r < num_atoms());
         return reg_of_atom(r + 1) - reg_of_atom(r);
      }

      /**
       * Get the number of atoms the whole register space is partitioned into.
       */
      unsigned
      num_atoms() const
      {
         return atoms[max_reg];
      }

   private:
      /**
       * Number of trailing atoms inserted for convenience so among other
       * things we don't need to special-case the last element in
       * size_of_atom().
       */
      static const unsigned num_terminator_atoms = 1;
      unsigned max_reg;
      unsigned *offsets;
      unsigned *atoms;
   };

   /**
    * Only GRF sources (whether they have been register-allocated or not) can
    * possibly incur bank conflicts.
    */
   bool
   is_grf(const brw_reg &r)
   {
      return r.file == VGRF || r.file == FIXED_GRF;
   }

   /**
    * Register offset of \p r in GRF units.  Useful because the representation
    * of GRFs post-register allocation is somewhat inconsistent and depends on
    * whether the register already had a fixed GRF offset prior to register
    * allocation or whether it was part of a VGRF allocation.
    */
   unsigned
   reg_of(const brw_reg &r)
   {
      assert(is_grf(r));
      if (r.file == VGRF)
         return r.nr + r.offset / REG_SIZE;
      else
         return reg_offset(r) / REG_SIZE;
   }

   /**
    * Calculate the finest partitioning of the GRF space compatible with the
    * register contiguity requirements derived from all instructions part of
    * the program.
    */
   partitioning
   shader_reg_partitioning(const fs_visitor *v)
   {
      partitioning p(BRW_MAX_GRF);

      foreach_block_and_inst(block, fs_inst, inst, v->cfg) {
         if (is_grf(inst->dst))
            p.require_contiguous(reg_of(inst->dst), regs_written(inst));

         for (int i = 0; i < inst->sources; i++) {
            if (is_grf(inst->src[i]))
               p.require_contiguous(reg_of(inst->src[i]),
                                    regs_read(v->devinfo, inst, i));
         }
      }

      return p;
   }

   /**
    * Return the set of GRF atoms that should be left untouched at their
    * original location to avoid violating hardware or software assumptions.
    */
   bool *
   shader_reg_constraints(const fs_visitor *v, const partitioning &p)
   {
      bool *constrained = new bool[p.num_atoms()]();

      /* These are read implicitly by some send-message instructions without
       * any indication at the IR level.  Assume they are unsafe to move
       * around.
       */
      for (unsigned reg = 0; reg < 2; reg++)
         constrained[p.atom_of_reg(reg)] = true;

      /* At Intel Broadwell PRM, vol 07, section "Instruction Set Reference",
       * subsection "EUISA Instructions", Send Message (page 990):
       *
       * "r127 must not be used for return address when there is a src and
       * dest overlap in send instruction."
       *
       * Register allocation ensures that, so don't move 127 around to avoid
       * breaking that property.
       */
      constrained[p.atom_of_reg(127)] = true;

      foreach_block_and_inst(block, fs_inst, inst, v->cfg) {
         /* Assume that anything referenced via fixed GRFs is baked into the
          * hardware's fixed-function logic and may be unsafe to move around.
          * Also take into account the source GRF restrictions of EOT
          * send-message instructions.
          */
         if (inst->dst.file == FIXED_GRF)
            constrained[p.atom_of_reg(reg_of(inst->dst))] = true;

         for (int i = 0; i < inst->sources; i++) {
            if (inst->src[i].file == FIXED_GRF ||
                (is_grf(inst->src[i]) && inst->eot))
               constrained[p.atom_of_reg(reg_of(inst->src[i]))] = true;
         }
      }

      return constrained;
   }

   /**
    * Return whether the hardware will be able to prevent a bank conflict by
    * optimizing out the read cycle of a source register.  The formula was
    * found experimentally.
    */
   bool
   is_conflict_optimized_out(const intel_device_info *devinfo,
                             const fs_inst *inst)
   {
      return
         (is_grf(inst->src[0]) && (reg_of(inst->src[0]) == reg_of(inst->src[1]) ||
                                   reg_of(inst->src[0]) == reg_of(inst->src[2]))) ||
          reg_of(inst->src[1]) == reg_of(inst->src[2]);
   }

   /**
    * Return a matrix that allows reasonably efficient computation of the
    * cycle-count cost of bank conflicts incurred throughout the whole program
    * for any given atom-to-bank assignment.
    *
    * More precisely, if C_r_s_p is the result of this function, the total
    * cost of all bank conflicts involving any given atom r can be readily
    * recovered as follows:
    *
    *  S(B) = Sum_s_p(d_(p^B_r)_(B_s) * C_r_s_p)
    *
    * where d_i_j is the Kronecker delta, and B_r indicates the bank
    * assignment of r.  \sa delta_conflicts() for a vectorized implementation
    * of the expression above.
    *
    * FINISHME: Teach this about the Gfx10+ bank conflict rules, which are
    *           somewhat more relaxed than on previous generations.  In the
    *           meantime optimizing based on Gfx9 weights is likely to be more
    *           helpful than not optimizing at all.
    */
   weight_vector_type *
   shader_conflict_weight_matrix(const fs_visitor *v, const partitioning &p)
   {
      weight_vector_type *conflicts = new weight_vector_type[p.num_atoms()];
      for (unsigned r = 0; r < p.num_atoms(); r++)
         conflicts[r] = weight_vector_type(2 * p.num_atoms());

      /* Crude approximation of the number of times the current basic block
       * will be executed at run-time.
       */
      unsigned block_scale = 1;

      foreach_block_and_inst(block, fs_inst, inst, v->cfg) {
         if (inst->opcode == BRW_OPCODE_DO) {
            block_scale *= 10;

         } else if (inst->opcode == BRW_OPCODE_WHILE) {
            block_scale /= 10;

         } else if (inst->is_3src(v->compiler) &&
                    is_grf(inst->src[1]) && is_grf(inst->src[2])) {
            const unsigned r = p.atom_of_reg(reg_of(inst->src[1]));
            const unsigned s = p.atom_of_reg(reg_of(inst->src[2]));

            /* Estimate of the cycle-count cost of incurring a bank conflict
             * for this instruction.  This is only true on the average, for a
             * sequence of back-to-back ternary instructions, since the EU
             * front-end only seems to be able to issue a new instruction at
             * an even cycle.  The cost of a bank conflict incurred by an
             * isolated ternary instruction may be higher.
             */
            const unsigned exec_size = inst->dst.component_size(inst->exec_size);
            const unsigned cycle_scale = block_scale * DIV_ROUND_UP(exec_size,
                                                                    REG_SIZE);

            /* Neglect same-atom conflicts (since they're either trivial or
             * impossible to avoid without splitting the atom), and conflicts
             * known to be optimized out by the hardware.
             */
            if (r != s && !is_conflict_optimized_out(v->devinfo, inst)) {
               /* Calculate the parity of the sources relative to the start of
                * their respective atoms.  If their parity is the same (and
                * none of the atoms straddle the 2KB mark), the instruction
                * will incur a conflict iff both atoms are assigned the same
                * bank b.  If their parity is opposite, the instruction will
                * incur a conflict iff they are assigned opposite banks (b and
                * b^1).
                */
               const bool p_r = 1 & (reg_of(inst->src[1]) - p.reg_of_atom(r));
               const bool p_s = 1 & (reg_of(inst->src[2]) - p.reg_of_atom(s));
               const unsigned p = p_r ^ p_s;

               /* Calculate the updated cost of a hypothetical conflict
                * between atoms r and s.  Note that the weight matrix is
                * symmetric with respect to indices r and s by construction.
                */
               const scalar_type w = MIN2(unsigned(max_scalar),
                                          get(conflicts[r], s, p) + cycle_scale);
               set(conflicts[r], s, p, w);
               set(conflicts[s], r, p, w);
            }
         }
      }

      return conflicts;
   }

   /**
    * Return the set of GRF atoms that could potentially lead to bank
    * conflicts if laid out unfavorably in the GRF space according to
    * the specified \p conflicts matrix (\sa
    * shader_conflict_weight_matrix()).
    */
   bool *
   have_any_conflicts(const partitioning &p,
                      const weight_vector_type *conflicts)
   {
      bool *any_conflicts = new bool[p.num_atoms()]();

      for (unsigned r = 0; r < p.num_atoms(); r++) {
         const unsigned m = DIV_ROUND_UP(conflicts[r].size, vector_width);
         for (unsigned s = 0; s < m; s++)
            any_conflicts[r] |= sums(conflicts[r].v[s]);
      }

      return any_conflicts;
   }

   /**
    * Calculate the difference between two S(B) cost estimates as defined
    * above (\sa shader_conflict_weight_matrix()).  This represents the
    * (partial) cycle-count benefit from moving an atom r from bank p to n.
    * The respective bank assignments Bp and Bn are encoded as the \p
    * bank_mask_p and \p bank_mask_n bitmasks for efficient computation,
    * according to the formula:
    *
    *  bank_mask(B)_s_p = -d_(p^B_r)_(B_s)
    *
    * Notice the similarity with the delta function in the S(B) expression
    * above, and how bank_mask(B) can be precomputed for every possible
    * selection of r since bank_mask(B) only depends on it via B_r that may
    * only assume one of four different values, so the caller can keep every
    * possible bank_mask(B) vector in memory without much hassle (\sa
    * bank_characteristics()).
    */
   int
   delta_conflicts(const weight_vector_type &bank_mask_p,
                   const weight_vector_type &bank_mask_n,
                   const weight_vector_type &conflicts)
   {
      const unsigned m = DIV_ROUND_UP(conflicts.size, vector_width);
      vector_type s_p = {}, s_n = {};

      for (unsigned r = 0; r < m; r++) {
         s_p = adds(s_p, mask(bank_mask_p.v[r], conflicts.v[r]));
         s_n = adds(s_n, mask(bank_mask_n.v[r], conflicts.v[r]));
      }

      return sums(subs(s_p, s_n));
   }

   /**
    * Register atom permutation, represented as the start GRF offset each atom
    * is mapped into.
    */
   struct permutation {
      permutation() : v(NULL), size(0) {}

      permutation(unsigned n) :
         v(new unsigned[n]()), size(n) {}

      permutation(const permutation &p) :
         v(new unsigned[p.size]), size(p.size)
      {
         memcpy(v, p.v, p.size * sizeof(unsigned));
      }

      ~permutation()
      {
         delete[] v;
      }

      permutation &
      operator=(permutation p)
      {
         SWAP(v, p.v);
         SWAP(size, p.size);
         return *this;
      }

      unsigned *v;
      unsigned size;
   };

   /**
    * Return an identity permutation of GRF atoms.
    */
   permutation
   identity_reg_permutation(const partitioning &p)
   {
      permutation map(p.num_atoms());

      for (unsigned r = 0; r < map.size; r++)
         map.v[r] = p.reg_of_atom(r);

      return map;
   }

   /**
    * Return the bank index of GRF address \p reg, numbered according to the
    * table:
    *        Even Odd
    *    Lo    0   1
    *    Hi    2   3
    */
   unsigned
   bank_of(unsigned reg)
   {
      return (reg & 0x40) >> 5 | (reg & 1);
   }

   /**
    * Return bitmasks suitable for use as bank mask arguments for the
    * delta_conflicts() computation.  Note that this is just the (negative)
    * characteristic function of each bank, if you regard it as a set
    * containing all atoms assigned to it according to the \p map array.
    */
   weight_vector_type *
   bank_characteristics(const permutation &map)
   {
      weight_vector_type *banks = new weight_vector_type[4];

      for (unsigned b = 0; b < 4; b++) {
         banks[b] = weight_vector_type(2 * map.size);

         for (unsigned j = 0; j < map.size; j++) {
            for (unsigned p = 0; p < 2; p++)
               set(banks[b], j, p,
                   (b ^ p) == bank_of(map.v[j]) ? -1 : 0);
         }
      }

      return banks;
   }

   /**
    * Return an improved permutation of GRF atoms based on \p map attempting
    * to reduce the total cycle-count cost of bank conflicts greedily.
    *
    * Note that this doesn't attempt to merge multiple atoms into one, which
    * may allow it to do a better job in some cases -- It simply reorders
    * existing atoms in the GRF space without affecting their identity.
    */
   permutation
   optimize_reg_permutation(const partitioning &p,
                            const bool *constrained,
                            const weight_vector_type *conflicts,
                            permutation map)
   {
      const bool *any_conflicts = have_any_conflicts(p, conflicts);
      weight_vector_type *banks = bank_characteristics(map);

      for (unsigned r = 0; r < map.size; r++) {
         const unsigned bank_r = bank_of(map.v[r]);

         if (!constrained[r]) {
            unsigned best_s = r;
            int best_benefit = 0;

            for (unsigned s = 0; s < map.size; s++) {
               const unsigned bank_s = bank_of(map.v[s]);

               if (bank_r != bank_s && !constrained[s] &&
                   p.size_of_atom(r) == p.size_of_atom(s) &&
                   (any_conflicts[r] || any_conflicts[s])) {
                  const int benefit =
                     delta_conflicts(banks[bank_r], banks[bank_s], conflicts[r]) +
                     delta_conflicts(banks[bank_s], banks[bank_r], conflicts[s]);

                  if (benefit > best_benefit) {
                     best_s = s;
                     best_benefit = benefit;
                  }
               }
            }

            if (best_s != r) {
               for (unsigned b = 0; b < 4; b++) {
                  for (unsigned p = 0; p < 2; p++)
                     swap(banks[b], r, p, best_s, p);
               }

               SWAP(map.v[r], map.v[best_s]);
            }
         }
      }

      delete[] banks;
      delete[] any_conflicts;
      return map;
   }

   /**
    * Apply the GRF atom permutation given by \p map to register \p r and
    * return the result.
    */
   brw_reg
   transform(const partitioning &p, const permutation &map, brw_reg r)
   {
      if (r.file == VGRF) {
         const unsigned reg = reg_of(r);
         const unsigned s = p.atom_of_reg(reg);
         r.nr = map.v[s] + reg - p.reg_of_atom(s);
         r.offset = r.offset % REG_SIZE;
      }

      return r;
   }
}

bool
brw_opt_bank_conflicts(fs_visitor &s)
{
   assert(s.grf_used || !"Must be called after register allocation");

   /* TODO: Re-work this pass for Gfx20+. */
   if (s.devinfo->ver >= 20)
      return false;

   const partitioning p = shader_reg_partitioning(&s);
   const bool *constrained = shader_reg_constraints(&s, p);
   const weight_vector_type *conflicts =
      shader_conflict_weight_matrix(&s, p);
   const permutation map =
      optimize_reg_permutation(p, constrained, conflicts,
                               identity_reg_permutation(p));

   foreach_block_and_inst(block, fs_inst, inst, s.cfg) {
      inst->dst = transform(p, map, inst->dst);

      for (int i = 0; i < inst->sources; i++)
         inst->src[i] = transform(p, map, inst->src[i]);
   }

   delete[] conflicts;
   delete[] constrained;
   return true;
}

/**
 * Return whether the instruction incurs GRF bank conflict cycles.
 *
 * Note that this is only accurate after register allocation because otherwise
 * we don't know which bank each VGRF is going to end up aligned to.
 */
bool
has_bank_conflict(const struct brw_isa_info *isa, const fs_inst *inst)
{
   return is_3src(isa, inst->opcode) &&
          is_grf(inst->src[1]) && is_grf(inst->src[2]) &&
          bank_of(reg_of(inst->src[1])) == bank_of(reg_of(inst->src[2])) &&
          !is_conflict_optimized_out(isa->devinfo, inst);
}