freedreno/ir3/ir3_a4xx.c

/*
 * Copyright (C) 2017-2018 Rob Clark <robclark@freedesktop.org>
 *
 * 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.
 *
 * Authors:
 *    Rob Clark <robclark@freedesktop.org>
 */

/* 500 gets us LDIB but doesn't change any other a4xx instructions */
#define GPU 500

#include "ir3_context.h"
#include "ir3_image.h"

/* SSBO data is available at this CB address, addressed like regular consts
 * containing the following data in each vec4:
 *
 * [ base address, pitch, array_pitch, cpp ]
 *
 * These mirror the values uploaded to A4XX_SSBO_0 state. For A5XX, these are
 * uploaded manually by the driver.
 */
#define A4XX_SSBO_CB_BASE(i) (0x700 + ((i) << 2))

/*
 * Handlers for instructions changed/added in a4xx:
 */

/* Convert byte offset to address of appropriate width for GPU */
static struct ir3_instruction *
byte_offset_to_address(struct ir3_context *ctx,
      nir_src *ssbo,
      struct ir3_instruction *byte_offset)
{
   struct ir3_block *b = ctx->block;

   if (ctx->compiler->gen == 4) {
      uint32_t index = nir_src_as_uint(*ssbo);
      unsigned cb = A4XX_SSBO_CB_BASE(index);
      byte_offset = ir3_ADD_U(b, create_uniform(b, cb), 0, byte_offset, 0);
   }

   if (ctx->compiler->is_64bit) {
      return ir3_collect(b, byte_offset, create_immed(b, 0));
   } else {
      return byte_offset;
   }
}

/* src[] = { buffer_index, offset }. No const_index */
static void
emit_intrinsic_load_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
                         struct ir3_instruction **dst)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *ldgb, *src0, *src1, *byte_offset, *offset;

   struct ir3_instruction *ssbo = ir3_ssbo_to_ibo(ctx, intr->src[0]);

   byte_offset = ir3_get_src(ctx, &intr->src[1])[0];
   offset = ir3_get_src(ctx, &intr->src[2])[0];

   /* src0 is uvec2(offset*4, 0), src1 is offset.. nir already *= 4: */
   src0 = byte_offset_to_address(ctx, &intr->src[0], byte_offset);
   src1 = offset;

   ldgb = ir3_LDGB(b, ssbo, 0, src0, 0, src1, 0);
   ldgb->dsts[0]->wrmask = MASK(intr->num_components);
   ldgb->cat6.iim_val = intr->num_components;
   ldgb->cat6.d = 4;
   ldgb->cat6.type = TYPE_U32;
   ldgb->barrier_class = IR3_BARRIER_BUFFER_R;
   ldgb->barrier_conflict = IR3_BARRIER_BUFFER_W;

   ir3_split_dest(b, dst, ldgb, 0, intr->num_components);
}

/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
static void
emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *stgb, *src0, *src1, *src2, *byte_offset, *offset;
   unsigned wrmask = nir_intrinsic_write_mask(intr);
   unsigned ncomp = ffs(~wrmask) - 1;

   assert(wrmask == BITFIELD_MASK(intr->num_components));

   struct ir3_instruction *ssbo = ir3_ssbo_to_ibo(ctx, intr->src[1]);

   byte_offset = ir3_get_src(ctx, &intr->src[2])[0];
   offset = ir3_get_src(ctx, &intr->src[3])[0];

   /* src0 is value, src1 is offset, src2 is uvec2(offset*4, 0)..
    * nir already *= 4:
    */
   src0 = ir3_create_collect(b, ir3_get_src(ctx, &intr->src[0]), ncomp);
   src1 = offset;
   src2 = byte_offset_to_address(ctx, &intr->src[1], byte_offset);

   stgb = ir3_STGB(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
   stgb->cat6.iim_val = ncomp;
   stgb->cat6.d = 4;
   stgb->cat6.type = TYPE_U32;
   stgb->barrier_class = IR3_BARRIER_BUFFER_W;
   stgb->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;

   array_insert(b, b->keeps, stgb);
}

static struct ir3_instruction *
emit_atomic(struct ir3_block *b,
            nir_atomic_op op,
            struct ir3_instruction *bo,
            struct ir3_instruction *data,
            struct ir3_instruction *offset,
            struct ir3_instruction *byte_offset)
{
   switch (op) {
   case nir_atomic_op_iadd:
      return ir3_ATOMIC_S_ADD(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_imin:
      return ir3_ATOMIC_S_MIN(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_umin:
      return ir3_ATOMIC_S_MIN(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_imax:
      return ir3_ATOMIC_S_MAX(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_umax:
      return ir3_ATOMIC_S_MAX(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_iand:
      return ir3_ATOMIC_S_AND(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_ior:
      return ir3_ATOMIC_S_OR(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_ixor:
      return ir3_ATOMIC_S_XOR(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_xchg:
      return ir3_ATOMIC_S_XCHG(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   case nir_atomic_op_cmpxchg:
      return ir3_ATOMIC_S_CMPXCHG(b, bo, 0, data, 0, offset, 0, byte_offset, 0);
   default:
      unreachable("boo");
   }
}

/*
 * SSBO atomic intrinsics
 *
 * All of the SSBO atomic memory operations read a value from memory,
 * compute a new value using one of the operations below, write the new
 * value to memory, and return the original value read.
 *
 * All operations take 3 sources except CompSwap that takes 4. These
 * sources represent:
 *
 * 0: The SSBO buffer index.
 * 1: The byte offset into the SSBO buffer of the variable that the atomic
 *    operation will operate on.
 * 2: The data parameter to the atomic function (i.e. the value to add
 *    in, etc).
 * 3: CompSwap: the second data parameter.
 *    Non-CompSwap: The dword offset into the SSBO buffer variable.
 * 4: CompSwap: The dword offset into the SSBO buffer variable.
 *
 * We use custom ssbo_*_ir3 intrinsics generated by ir3_nir_lower_io_offsets()
 * so we can have the dword offset generated in NIR.
 */
static struct ir3_instruction *
emit_intrinsic_atomic_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
   struct ir3_block *b = ctx->block;
   nir_atomic_op op = nir_intrinsic_atomic_op(intr);
   type_t type = nir_atomic_op_type(op) == nir_type_int ? TYPE_S32 : TYPE_U32;

   struct ir3_instruction *ssbo = ir3_ssbo_to_ibo(ctx, intr->src[0]);

   struct ir3_instruction *data = ir3_get_src(ctx, &intr->src[2])[0];
   /* 64b byte offset */
   struct ir3_instruction *byte_offset =
      byte_offset_to_address(ctx, &intr->src[0], ir3_get_src(ctx, &intr->src[1])[0]);
   /* dword offset for everything but cmpxchg */
   struct ir3_instruction *src3 = ir3_get_src(ctx, &intr->src[3])[0];

   if (op == nir_atomic_op_cmpxchg) {
      /* for cmpxchg, src0 is [ui]vec2(data, compare): */
      data = ir3_collect(b, src3, data);
      src3 = ir3_get_src(ctx, &intr->src[4])[0];
   }

   struct ir3_instruction *atomic =
      emit_atomic(b, op, ssbo, data, src3, byte_offset);

   atomic->cat6.iim_val = 1;
   atomic->cat6.d = 4;
   atomic->cat6.type = type;
   atomic->barrier_class = IR3_BARRIER_BUFFER_W;
   atomic->barrier_conflict = IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;

   /* even if nothing consume the result, we can't DCE the instruction: */
   array_insert(b, b->keeps, atomic);

   return atomic;
}

static struct ir3_instruction *
get_image_offset(struct ir3_context *ctx, const nir_intrinsic_instr *instr,
                 struct ir3_instruction *const *coords, bool byteoff)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *offset;
   unsigned index = nir_src_as_uint(instr->src[0]);
   unsigned ncoords = ir3_get_image_coords(instr, NULL);

   /* to calculate the byte offset (yes, uggg) we need (up to) three
    * const values to know the bytes per pixel, and y and z stride:
    */
   unsigned cb;
   if (ctx->compiler->gen > 4) {
      const struct ir3_const_state *const_state = ir3_const_state(ctx->so);
      assert(const_state->image_dims.mask & (1 << index));

      cb = regid(const_state->offsets.image_dims, 0) +
         const_state->image_dims.off[index];
   } else {
      index += ctx->s->info.num_ssbos;
      cb = A4XX_SSBO_CB_BASE(index);
   }

   /* offset = coords.x * bytes_per_pixel: */
   if (ctx->compiler->gen == 4)
      offset = ir3_MUL_S24(b, coords[0], 0, create_uniform(b, cb + 3), 0);
   else
      offset = ir3_MUL_S24(b, coords[0], 0, create_uniform(b, cb + 0), 0);
   if (ncoords > 1) {
      /* offset += coords.y * y_pitch: */
      offset =
         ir3_MAD_S24(b, create_uniform(b, cb + 1), 0, coords[1], 0, offset, 0);
   }
   if (ncoords > 2) {
      /* offset += coords.z * z_pitch: */
      offset =
         ir3_MAD_S24(b, create_uniform(b, cb + 2), 0, coords[2], 0, offset, 0);
   }

   /* a4xx: must add in the base address: */
   if (ctx->compiler->gen == 4)
      offset = ir3_ADD_U(b, offset, 0, create_uniform(b, cb + 0), 0);

   if (!byteoff) {
      /* Some cases, like atomics, seem to use dword offset instead
       * of byte offsets.. blob just puts an extra shr.b in there
       * in those cases:
       */
      offset = ir3_SHR_B(b, offset, 0, create_immed(b, 2), 0);
   }

   if (ctx->compiler->is_64bit)
      return ir3_collect(b, offset, create_immed(b, 0));
   else
      return offset;
}

/* src[] = { deref, coord, sample_index }. const_index[] = {} */
static void
emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
                          struct ir3_instruction **dst)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
   struct ir3_instruction *ibo = ir3_image_to_ibo(ctx, intr->src[0]);
   struct ir3_instruction *offset = get_image_offset(ctx, intr, coords, true);
   unsigned ncoords = ir3_get_image_coords(intr, NULL);
   unsigned ncomp =
      ir3_get_num_components_for_image_format(nir_intrinsic_format(intr));

   struct ir3_instruction *ldib;
   /* At least A420 does not have LDIB. Use LDGB and perform conversion
    * ourselves.
    *
    * TODO: Actually do the conversion. ES 3.1 only requires this for
    * single-component 32-bit types anyways.
    */
   if (ctx->compiler->gen > 4) {
      ldib = ir3_LDIB(
            b, ibo, 0, offset, 0, ir3_create_collect(b, coords, ncoords), 0);
   } else {
      ldib = ir3_LDGB(
            b, ibo, 0, offset, 0, ir3_create_collect(b, coords, ncoords), 0);
      switch (nir_intrinsic_format(intr)) {
      case PIPE_FORMAT_R32_UINT:
      case PIPE_FORMAT_R32_SINT:
      case PIPE_FORMAT_R32_FLOAT:
         break;
      default:
         /* For some reason even more 32-bit components don't work. */
         assert(0);
         break;
      }
   }
   ldib->dsts[0]->wrmask = MASK(intr->num_components);
   ldib->cat6.iim_val = ncomp;
   ldib->cat6.d = ncoords;
   ldib->cat6.type = ir3_get_type_for_image_intrinsic(intr);
   ldib->cat6.typed = true;
   ldib->barrier_class = IR3_BARRIER_IMAGE_R;
   ldib->barrier_conflict = IR3_BARRIER_IMAGE_W;

   ir3_split_dest(b, dst, ldib, 0, intr->num_components);
}

/* src[] = { index, coord, sample_index, value }. const_index[] = {} */
static void
emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *stib, *offset;
   struct ir3_instruction *const *value = ir3_get_src(ctx, &intr->src[3]);
   struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
   struct ir3_instruction *ibo = ir3_image_to_ibo(ctx, intr->src[0]);
   unsigned ncoords = ir3_get_image_coords(intr, NULL);
   unsigned ncomp =
      ir3_get_num_components_for_image_format(nir_intrinsic_format(intr));

   /* src0 is value
    * src1 is coords
    * src2 is 64b byte offset
    */

   offset = get_image_offset(ctx, intr, coords, true);

   /* NOTE: stib seems to take byte offset, but stgb.typed can be used
    * too and takes a dword offset.. not quite sure yet why blob uses
    * one over the other in various cases.
    */

   stib = ir3_STIB(b, ibo, 0, ir3_create_collect(b, value, ncomp), 0,
                   ir3_create_collect(b, coords, ncoords), 0, offset, 0);
   stib->cat6.iim_val = ncomp;
   stib->cat6.d = ncoords;
   stib->cat6.type = ir3_get_type_for_image_intrinsic(intr);
   stib->cat6.typed = true;
   stib->barrier_class = IR3_BARRIER_IMAGE_W;
   stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;

   array_insert(b, b->keeps, stib);
}

/* src[] = { deref, coord, sample_index, value, compare }. const_index[] = {} */
static struct ir3_instruction *
emit_intrinsic_atomic_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
   struct ir3_block *b = ctx->block;
   struct ir3_instruction *atomic, *src0, *src1, *src2;
   struct ir3_instruction *const *coords = ir3_get_src(ctx, &intr->src[1]);
   struct ir3_instruction *image = ir3_image_to_ibo(ctx, intr->src[0]);
   unsigned ncoords = ir3_get_image_coords(intr, NULL);
   nir_atomic_op op = nir_intrinsic_atomic_op(intr);

   /* src0 is value (or uvec2(value, compare))
    * src1 is coords
    * src2 is 64b byte offset
    */
   src0 = ir3_get_src(ctx, &intr->src[3])[0];
   src1 = ir3_create_collect(b, coords, ncoords);
   src2 = get_image_offset(ctx, intr, coords, ctx->compiler->gen == 4);

   if (op == nir_atomic_op_cmpxchg)
      src0 = ir3_collect(b, ir3_get_src(ctx, &intr->src[4])[0], src0);

   atomic = emit_atomic(b, op, image, src0, src1, src2);
   atomic->cat6.iim_val = 1;
   atomic->cat6.d = ncoords;
   atomic->cat6.type = ir3_get_type_for_image_intrinsic(intr);
   atomic->cat6.typed = ctx->compiler->gen == 5;
   atomic->barrier_class = IR3_BARRIER_IMAGE_W;
   atomic->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;

   /* even if nothing consume the result, we can't DCE the instruction: */
   array_insert(b, b->keeps, atomic);

   return atomic;
}

static struct ir3_instruction *
emit_intrinsic_atomic_global(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
   unreachable("Global atomic are unimplemented on A5xx");
}

const struct ir3_context_funcs ir3_a4xx_funcs = {
   .emit_intrinsic_load_ssbo = emit_intrinsic_load_ssbo,
   .emit_intrinsic_store_ssbo = emit_intrinsic_store_ssbo,
   .emit_intrinsic_atomic_ssbo = emit_intrinsic_atomic_ssbo,
   .emit_intrinsic_load_image = emit_intrinsic_load_image,
   .emit_intrinsic_store_image = emit_intrinsic_store_image,
   .emit_intrinsic_atomic_image = emit_intrinsic_atomic_image,
   .emit_intrinsic_image_size = emit_intrinsic_image_size_tex,
   .emit_intrinsic_load_global_ir3 = NULL,
   .emit_intrinsic_store_global_ir3 = NULL,
   .emit_intrinsic_atomic_global = emit_intrinsic_atomic_global,
};