panfrost/compiler/bi_opt_mod_props.c

/*
 * Copyright (C) 2021 Collabora, Ltd.
 * Copyright (C) 2021 Alyssa Rosenzweig <[email protected]>
 *
 * 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 "bi_builder.h"
#include "compiler.h"

/*
 * Due to a Bifrost encoding restriction, some instructions cannot have an abs
 * modifier on both sources. Check if adding a fabs modifier to a given source
 * of a binary instruction would cause this restriction to be hit.
 */
static bool
bi_would_impact_abs(unsigned arch, bi_instr *I, bi_index repl, unsigned s)
{
   return (arch <= 8) && I->src[1 - s].abs &&
          bi_is_word_equiv(I->src[1 - s], repl);
}

static bool
bi_takes_fabs(unsigned arch, bi_instr *I, bi_index repl, unsigned s)
{
   switch (I->op) {
   case BI_OPCODE_FCMP_V2F16:
   case BI_OPCODE_FMAX_V2F16:
   case BI_OPCODE_FMIN_V2F16:
      return !bi_would_impact_abs(arch, I, repl, s);
   case BI_OPCODE_FADD_V2F16:
      /*
       * For FADD.v2f16, the FMA pipe has the abs encoding hazard,
       * while the FADD pipe cannot encode a clamp. Either case in
       * isolation can be worked around in the scheduler, but both
       * together is impossible to encode. Avoid the hazard.
       */
      return !(I->clamp && bi_would_impact_abs(arch, I, repl, s));
   case BI_OPCODE_V2F32_TO_V2F16:
      /* TODO: Needs both match or lower */
      return false;
   case BI_OPCODE_FLOG_TABLE_F32:
      /* TODO: Need to check mode */
      return false;
   default:
      return bi_opcode_props[I->op].abs & BITFIELD_BIT(s);
   }
}

static bool
bi_takes_fneg(unsigned arch, bi_instr *I, unsigned s)
{
   switch (I->op) {
   case BI_OPCODE_CUBE_SSEL:
   case BI_OPCODE_CUBE_TSEL:
   case BI_OPCODE_CUBEFACE:
      /* TODO: Bifrost encoding restriction: need to match or lower */
      return arch >= 9;
   case BI_OPCODE_FREXPE_F32:
   case BI_OPCODE_FREXPE_V2F16:
   case BI_OPCODE_FLOG_TABLE_F32:
      /* TODO: Need to check mode */
      return false;
   default:
      return bi_opcode_props[I->op].neg & BITFIELD_BIT(s);
   }
}

static bool
bi_is_fabsneg(enum bi_opcode op, enum bi_size size)
{
   return (size == BI_SIZE_32 && op == BI_OPCODE_FABSNEG_F32) ||
          (size == BI_SIZE_16 && op == BI_OPCODE_FABSNEG_V2F16);
}

static enum bi_swizzle
bi_compose_swizzle_16(enum bi_swizzle a, enum bi_swizzle b)
{
   assert(a <= BI_SWIZZLE_H11);
   assert(b <= BI_SWIZZLE_H11);

   bool al = (a & BI_SWIZZLE_H10);
   bool ar = (a & BI_SWIZZLE_H01);
   bool bl = (b & BI_SWIZZLE_H10);
   bool br = (b & BI_SWIZZLE_H01);

   return ((al ? br : bl) ? BI_SWIZZLE_H10 : 0) |
          ((ar ? br : bl) ? BI_SWIZZLE_H01 : 0);
}

/* Like bi_replace_index, but composes instead of overwrites */

static inline bi_index
bi_compose_float_index(bi_index old, bi_index repl)
{
   /* abs(-x) = abs(+x) so ignore repl.neg if old.abs is set, otherwise
    * -(-x) = x but -(+x) = +(-x) so need to exclusive-or the negates */
   repl.neg = old.neg ^ (repl.neg && !old.abs);

   /* +/- abs(+/- abs(x)) = +/- abs(x), etc so just or the two */
   repl.abs |= old.abs;

   /* Use the old swizzle to select from the replacement swizzle */
   repl.swizzle = bi_compose_swizzle_16(old.swizzle, repl.swizzle);

   return repl;
}

/* DISCARD.b32(FCMP.f(x, y)) --> DISCARD.f(x, y) */

static inline bool
bi_fuse_discard_fcmp(bi_context *ctx, bi_instr *I, bi_instr *mod)
{
   if (!mod)
      return false;
   if (I->op != BI_OPCODE_DISCARD_B32)
      return false;
   if (mod->op != BI_OPCODE_FCMP_F32 && mod->op != BI_OPCODE_FCMP_V2F16)
      return false;
   if (mod->cmpf >= BI_CMPF_GTLT)
      return false;

   /* result_type doesn't matter */

   /* .abs and .neg modifiers allowed on Valhall DISCARD but not Bifrost */
   bool absneg = mod->src[0].neg || mod->src[0].abs;
   absneg |= mod->src[1].neg || mod->src[1].abs;

   if (ctx->arch <= 8 && absneg)
      return false;

   enum bi_swizzle r = I->src[0].swizzle;

   bi_builder b = bi_init_builder(ctx, bi_before_instr(I));
   I = bi_discard_f32(&b, mod->src[0], mod->src[1], mod->cmpf);

   if (mod->op == BI_OPCODE_FCMP_V2F16) {
      I->src[0].swizzle = bi_compose_swizzle_16(r, I->src[0].swizzle);
      I->src[1].swizzle = bi_compose_swizzle_16(r, I->src[1].swizzle);
   }

   return true;
}

/*
 * S32_TO_F32(S8_TO_S32(x)) -> S8_TO_F32 and friends. Round modes don't matter
 * because all 8-bit and 16-bit integers may be represented exactly as fp32.
 */
struct {
   enum bi_opcode inner;
   enum bi_opcode outer;
   enum bi_opcode replacement;
} bi_small_int_patterns[] = {
   {BI_OPCODE_S8_TO_S32, BI_OPCODE_S32_TO_F32, BI_OPCODE_S8_TO_F32},
   {BI_OPCODE_U8_TO_U32, BI_OPCODE_U32_TO_F32, BI_OPCODE_U8_TO_F32},
   {BI_OPCODE_U8_TO_U32, BI_OPCODE_S32_TO_F32, BI_OPCODE_U8_TO_F32},
   {BI_OPCODE_S16_TO_S32, BI_OPCODE_S32_TO_F32, BI_OPCODE_S16_TO_F32},
   {BI_OPCODE_U16_TO_U32, BI_OPCODE_U32_TO_F32, BI_OPCODE_U16_TO_F32},
   {BI_OPCODE_U16_TO_U32, BI_OPCODE_S32_TO_F32, BI_OPCODE_U16_TO_F32},
};

static inline void
bi_fuse_small_int_to_f32(bi_instr *I, bi_instr *mod)
{
   for (unsigned i = 0; i < ARRAY_SIZE(bi_small_int_patterns); ++i) {
      if (I->op != bi_small_int_patterns[i].outer)
         continue;
      if (mod->op != bi_small_int_patterns[i].inner)
         continue;

      assert(I->src[0].swizzle == BI_SWIZZLE_H01);
      I->src[0] = mod->src[0];
      I->round = BI_ROUND_NONE;
      I->op = bi_small_int_patterns[i].replacement;
   }
}

void
bi_opt_mod_prop_forward(bi_context *ctx)
{
   bi_instr **lut = calloc(sizeof(bi_instr *), ctx->ssa_alloc);

   bi_foreach_instr_global_safe(ctx, I) {
      /* Try fusing FCMP into DISCARD.b32, building a new DISCARD.f32
       * instruction. As this is the only optimization DISCARD is
       * involved in, this shortcircuits other processing.
       */
      if (I->op == BI_OPCODE_DISCARD_B32) {
         if (bi_is_ssa(I->src[0]) &&
             bi_fuse_discard_fcmp(ctx, I, lut[I->src[0].value])) {
            bi_remove_instruction(I);
         }

         continue;
      }

      bi_foreach_dest(I, d) {
         lut[I->dest[d].value] = I;
      }

      bi_foreach_ssa_src(I, s) {
         bi_instr *mod = lut[I->src[s].value];

         if (!mod)
            continue;

         unsigned size = bi_opcode_props[I->op].size;

         bi_fuse_small_int_to_f32(I, mod);

         if (bi_is_fabsneg(mod->op, size)) {
            if (mod->src[0].abs && !bi_takes_fabs(ctx->arch, I, mod->src[0], s))
               continue;

            if (mod->src[0].neg && !bi_takes_fneg(ctx->arch, I, s))
               continue;

            I->src[s] = bi_compose_float_index(I->src[s], mod->src[0]);
         }
      }
   }

   free(lut);
}

/* RSCALE has restrictions on how the clamp may be used, only used for
 * specialized transcendental sequences that set the clamp explicitly anyway */

static bool
bi_takes_clamp(bi_instr *I)
{
   switch (I->op) {
   case BI_OPCODE_FMA_RSCALE_F32:
   case BI_OPCODE_FMA_RSCALE_V2F16:
   case BI_OPCODE_FADD_RSCALE_F32:
      return false;
   case BI_OPCODE_FADD_V2F16:
      /* Encoding restriction */
      return !(I->src[0].abs && I->src[1].abs &&
               bi_is_word_equiv(I->src[0], I->src[1]));
   default:
      return bi_opcode_props[I->op].clamp;
   }
}

static bool
bi_is_fclamp(enum bi_opcode op, enum bi_size size)
{
   return (size == BI_SIZE_32 && op == BI_OPCODE_FCLAMP_F32) ||
          (size == BI_SIZE_16 && op == BI_OPCODE_FCLAMP_V2F16);
}

static bool
bi_optimizer_clamp(bi_instr *I, bi_instr *use)
{
   if (!bi_is_fclamp(use->op, bi_opcode_props[I->op].size))
      return false;
   if (!bi_takes_clamp(I))
      return false;

   /* Clamps are bitfields (clamp_m1_1/clamp_0_inf) so composition is OR */
   I->clamp |= use->clamp;
   I->dest[0] = use->dest[0];
   return true;
}

static enum bi_opcode
bi_sized_mux_op(unsigned size)
{
   switch (size) {
   case 8:
      return BI_OPCODE_MUX_V4I8;
   case 16:
      return BI_OPCODE_MUX_V2I16;
   case 32:
      return BI_OPCODE_MUX_I32;
   default:
      unreachable("invalid size");
   }
}

static bool
bi_is_fixed_mux(bi_instr *I, unsigned size, bi_index v1)
{
   return I->op == bi_sized_mux_op(size) &&
          bi_is_value_equiv(I->src[0], bi_zero()) &&
          bi_is_value_equiv(I->src[1], v1);
}

static bool
bi_takes_int_result_type(enum bi_opcode op)
{
   switch (op) {
   case BI_OPCODE_ICMP_I32:
   case BI_OPCODE_ICMP_S32:
   case BI_OPCODE_ICMP_U32:
   case BI_OPCODE_ICMP_V2I16:
   case BI_OPCODE_ICMP_V2S16:
   case BI_OPCODE_ICMP_V2U16:
   case BI_OPCODE_ICMP_V4I8:
   case BI_OPCODE_ICMP_V4S8:
   case BI_OPCODE_ICMP_V4U8:
   case BI_OPCODE_FCMP_F32:
   case BI_OPCODE_FCMP_V2F16:
      return true;
   default:
      return false;
   }
}

static bool
bi_takes_float_result_type(enum bi_opcode op)
{
   return (op == BI_OPCODE_FCMP_F32) || (op == BI_OPCODE_FCMP_V2F16);
}

/* CMP+MUX -> CMP with result type */
static bool
bi_optimizer_result_type(bi_instr *I, bi_instr *mux)
{
   if (bi_opcode_props[I->op].size != bi_opcode_props[mux->op].size)
      return false;

   if (bi_is_fixed_mux(mux, 32, bi_imm_f32(1.0)) ||
       bi_is_fixed_mux(mux, 16, bi_imm_f16(1.0))) {

      if (!bi_takes_float_result_type(I->op))
         return false;

      I->result_type = BI_RESULT_TYPE_F1;
   } else if (bi_is_fixed_mux(mux, 32, bi_imm_u32(1)) ||
              bi_is_fixed_mux(mux, 16, bi_imm_u16(1)) ||
              bi_is_fixed_mux(mux, 8, bi_imm_u8(1))) {

      if (!bi_takes_int_result_type(I->op))
         return false;

      I->result_type = BI_RESULT_TYPE_I1;
   } else {
      return false;
   }

   I->dest[0] = mux->dest[0];
   return true;
}

static bool
bi_is_var_tex(bi_instr *var, bi_instr *tex)
{
   return (var->op == BI_OPCODE_LD_VAR_IMM) &&
          (tex->op == BI_OPCODE_TEXS_2D_F16 ||
           tex->op == BI_OPCODE_TEXS_2D_F32) &&
          (var->register_format == BI_REGISTER_FORMAT_F32) &&
          ((var->sample == BI_SAMPLE_CENTER &&
            var->update == BI_UPDATE_STORE) ||
           (var->sample == BI_SAMPLE_NONE &&
            var->update == BI_UPDATE_RETRIEVE)) &&
          (tex->texture_index == tex->sampler_index) &&
          (tex->texture_index < 4) && (var->index < 8);
}

static bool
bi_optimizer_var_tex(bi_context *ctx, bi_instr *var, bi_instr *tex)
{
   if (!bi_is_var_tex(var, tex))
      return false;

   /* Construct the corresponding VAR_TEX intruction */
   bi_builder b = bi_init_builder(ctx, bi_after_instr(var));

   bi_instr *I = bi_var_tex_f32_to(&b, tex->dest[0], tex->lod_mode, var->sample,
                                   var->update, tex->texture_index, var->index);
   I->skip = tex->skip;

   if (tex->op == BI_OPCODE_TEXS_2D_F16)
      I->op = BI_OPCODE_VAR_TEX_F16;

   /* Dead code elimination will clean up for us */
   return true;
}

void
bi_opt_mod_prop_backward(bi_context *ctx)
{
   unsigned count = ctx->ssa_alloc;
   bi_instr **uses = calloc(count, sizeof(*uses));
   BITSET_WORD *multiple = calloc(BITSET_WORDS(count), sizeof(*multiple));

   bi_foreach_instr_global_rev(ctx, I) {
      bi_foreach_ssa_src(I, s) {
         unsigned v = I->src[s].value;

         if (uses[v] && uses[v] != I)
            BITSET_SET(multiple, v);
         else
            uses[v] = I;
      }

      if (!I->nr_dests)
         continue;

      bi_instr *use = uses[I->dest[0].value];

      if (!use || BITSET_TEST(multiple, I->dest[0].value))
         continue;

      /* Destination has a single use, try to propagate */
      bool propagated =
         bi_optimizer_clamp(I, use) || bi_optimizer_result_type(I, use);

      if (!propagated && I->op == BI_OPCODE_LD_VAR_IMM &&
          use->op == BI_OPCODE_SPLIT_I32) {
         /* Need to see through the split in a
          * ld_var_imm/split/var_tex  sequence
          */
         bi_instr *tex = uses[use->dest[0].value];

         if (!tex || BITSET_TEST(multiple, use->dest[0].value))
            continue;

         use = tex;
         propagated = bi_optimizer_var_tex(ctx, I, use);
      }

      if (propagated) {
         bi_remove_instruction(use);
         continue;
      }
   }

   free(uses);
   free(multiple);
}

/*
 * Lower pseudo instructions that exist to simplify the optimizer. Returns the
 * replacement instruction, or NULL if no replacement is needed.
 */
static bool
bi_lower_opt_instruction_helper(bi_builder *b, bi_instr *I)
{
   bi_instr *repl;

   switch (I->op) {
   case BI_OPCODE_FABSNEG_F32:
   case BI_OPCODE_FCLAMP_F32:
      repl = bi_fadd_f32_to(b, I->dest[0], I->src[0], bi_negzero());
      repl->clamp = I->clamp;
      return true;

   case BI_OPCODE_FABSNEG_V2F16:
   case BI_OPCODE_FCLAMP_V2F16:
      repl = bi_fadd_v2f16_to(b, I->dest[0], I->src[0], bi_negzero());
      repl->clamp = I->clamp;
      return true;

   case BI_OPCODE_DISCARD_B32:
      bi_discard_f32(b, I->src[0], bi_zero(), BI_CMPF_NE);
      return true;

   default:
      return false;
   }
}

void
bi_lower_opt_instructions(bi_context *ctx)
{
   bi_foreach_instr_global_safe(ctx, I) {
      bi_builder b = bi_init_builder(ctx, bi_before_instr(I));

      if (bi_lower_opt_instruction_helper(&b, I))
         bi_remove_instruction(I);
   }
}