Xenia-Canary/src/xenia/gpu/shaders/texture_load.xesli

/**
 ******************************************************************************
 * Xenia : Xbox 360 Emulator Research Project                                 *
 ******************************************************************************
 * Copyright 2022 Ben Vanik. All rights reserved.                             *
 * Released under the BSD license - see LICENSE in the root for more details. *
 ******************************************************************************
 */

#ifndef XENIA_GPU_SHADERS_TEXTURE_LOAD_XESLI_
#define XENIA_GPU_SHADERS_TEXTURE_LOAD_XESLI_

#include "endian.xesli"
#include "texture_address.xesli"

xesl_cbuffer_begin(xe_texture_load_constants, set=2, binding=0, b0, space0)
  uint xe_texture_load_is_tiled_3d_endian_scale;
  // Base offset in bytes, resolution-scaled.
  uint xe_texture_load_guest_offset;
  // For tiled textures - row pitch in guest blocks, aligned to 32, unscaled.
  // For linear textures - row pitch in bytes.
  uint xe_texture_load_guest_pitch_aligned;
  // For 3D textures only (ignored otherwise) - aligned to 32, unscaled.
  uint xe_texture_load_guest_z_stride_block_rows_aligned;

  // - std140 vector boundary -

  // If this is a packed mip tail, this is aligned to tile dimensions.
  // Resolution-scaled.
  xesl_uint3 xe_texture_load_size_blocks;
  // Base offset in bytes.
  uint xe_texture_load_host_offset;

  // - std140 vector boundary -

  uint xe_texture_load_host_pitch;
  uint xe_texture_load_height_texels;
xesl_cbuffer_end(xe_texture_load_constants)

bool XeTextureLoadIsTiled() {
  #ifdef XE_TEXTURE_LOAD_RESOLUTION_SCALED
    // Only resolved textures can be resolution-scaled, and resolving is only
    // possible to a tiled destination.
    return true;
  #else
    return (xesl_constant(xe_texture_load_constants,
                          xe_texture_load_is_tiled_3d_endian_scale) & 1u) != 0u;
  #endif
}

bool XeTextureLoadIs3D() {
  return (xesl_constant(xe_texture_load_constants,
                        xe_texture_load_is_tiled_3d_endian_scale) & (1u << 1u))
         != 0u;
}

uint XeTextureLoadEndian32() {
  return (xesl_constant(xe_texture_load_constants,
                        xe_texture_load_is_tiled_3d_endian_scale) >> 2u) & 3u;
}

xesl_uint2 XeTextureLoadResolutionScale() {
  #ifdef XE_TEXTURE_LOAD_RESOLUTION_SCALED
    return (xesl_constant(xe_texture_load_constants,
                          xe_texture_load_is_tiled_3d_endian_scale.xx) >>
            xesl_uint2(4u, 6u)) & 3u;
  #else
    return xesl_uint2(1u, 1u);
  #endif
}

xesl_uint3 XeTextureLoadSizeBlocks() {
  return xesl_constant(xe_texture_load_constants, xe_texture_load_size_blocks);
}

uint XeTextureLoadHostOffset() {
  return xesl_constant(xe_texture_load_constants, xe_texture_load_host_offset);
}

uint XeTextureLoadHostPitch() {
  return xesl_constant(xe_texture_load_constants, xe_texture_load_host_pitch);
}

uint XeTextureLoadHeightTexels() {
  return xesl_constant(xe_texture_load_constants,
                       xe_texture_load_height_texels);
}

// bpb and bpb_log2 are separate because bpb may be not a power of 2 (like 96).
uint XeTextureLoadGuestBlockOffset(xesl_uint3 block_index, uint bpb,
                                   uint bpb_log2) {
  #ifdef XE_TEXTURE_LOAD_RESOLUTION_SCALED
    // Only resolved textures can be resolution-scaled, and resolving is only
    // possible to a tiled destination.
    return
        xesl_constant(xe_texture_load_constants, xe_texture_load_guest_offset) +
        XeTextureScaledTiledOffset(
            XeTextureLoadIs3D(), block_index,
            xesl_constant(xe_texture_load_constants,
                          xe_texture_load_guest_pitch_aligned),
            xesl_constant(xe_texture_load_constants,
                          xe_texture_load_guest_z_stride_block_rows_aligned),
            bpb_log2, XeTextureLoadResolutionScale());
  #else
    int block_offset_guest;
    xesl_dont_flatten if (XeTextureLoadIsTiled()) {
      xesl_dont_flatten if (XeTextureLoadIs3D()) {
        block_offset_guest = XeTextureTiledOffset3D(
            xesl_int3(block_index),
            xesl_constant(xe_texture_load_constants,
                          xe_texture_load_guest_pitch_aligned),
            xesl_constant(xe_texture_load_constants,
                          xe_texture_load_guest_z_stride_block_rows_aligned),
            bpb_log2);
      } else {
        block_offset_guest = XeTextureTiledOffset2D(
            xesl_int2(block_index.xy),
            xesl_constant(xe_texture_load_constants,
                          xe_texture_load_guest_pitch_aligned), bpb_log2);
      }
    } else {
      block_offset_guest = XeTextureGuestLinearOffset(
          xesl_int3(block_index),
          xesl_constant(xe_texture_load_constants,
                        xe_texture_load_guest_pitch_aligned),
          xesl_constant(xe_texture_load_constants,
                        xe_texture_load_guest_z_stride_block_rows_aligned),
          bpb);
    }
    return uint(int(xesl_constant(xe_texture_load_constants,
                                  xe_texture_load_guest_offset)) +
           block_offset_guest);
  #endif
}

// Offset of the beginning of the odd R32G32/R32G32B32A32 load address from the
// address of the even load, for power-of-two-sized textures.
uint XeTextureLoadRightConsecutiveBlocksOffset(uint block_x, uint bpb_log2) {
  #ifdef XE_TEXTURE_LOAD_RESOLUTION_SCALED
    return XeTextureScaledRightSubUnitOffsetInConsecutivePair(
               block_x, bpb_log2, XeTextureLoadResolutionScale());
  #else
    uint offset;
    uint consecutive_blocks_log2 =
        XeTextureTiledConsecutiveBlocksLog2(bpb_log2);
    xesl_dont_flatten if (XeTextureLoadIsTiled()) {
      offset = XeTextureTiledOddConsecutiveBlocksOffset(bpb_log2);
    } else {
      offset = 1u << (consecutive_blocks_log2 + bpb_log2);
    }
    return offset;
  #endif
}

#endif  // XENIA_GPU_SHADERS_TEXTURE_LOAD_XESLI_