vk_data_conversion.hh

Go to the documentation of this file.

/* SPDX-FileCopyrightText: 2023 Blender Authors
 *
 * SPDX-License-Identifier: GPL-2.0-or-later */
 
#pragma once
 
#include "BLI_math_vector_types.hh"
 
#include "gpu_texture_private.hh"
 
#include "vk_common.hh"
 
namespace blender::gpu {
struct VKWorkarounds;
 
void convert_host_to_device(void *dst_buffer,
                            const void *src_buffer,
                            size_t buffer_size,
                            eGPUDataFormat host_format,
                            eGPUTextureFormat host_texture_format,
                            eGPUTextureFormat device_format);
 
void convert_device_to_host(void *dst_buffer,
                            const void *src_buffer,
                            size_t buffer_size,
                            eGPUDataFormat host_format,
                            eGPUTextureFormat host_texture_format,
                            eGPUTextureFormat device_format);
 
bool conversion_needed(const GPUVertFormat &vertex_format);
 
void convert_in_place(void *data, const GPUVertFormat &vertex_format, const uint vertex_len);
 
/* -------------------------------------------------------------------- */
struct VertexFormatConverter {
 private:
  const GPUVertFormat *source_format_ = nullptr;
 
  const GPUVertFormat *device_format_ = nullptr;
 
  bool needs_conversion_ = false;
 
  GPUVertFormat converted_format_;
 
 public:
  bool is_initialized() const;
 
  void init(const GPUVertFormat *vertex_format, const VKWorkarounds &workarounds);
 
  const GPUVertFormat &device_format_get() const;
 
  bool needs_conversion() const;
 
  void convert(void *device_data, const void *src_data, const uint vertex_len) const;
 
  void reset();
 
 private:
  void update_conversion_flags(const GPUVertFormat &vertex_format,
                               const VKWorkarounds &workarounds);
  void update_conversion_flags(const GPUVertAttr &vertex_attribute,
                               const VKWorkarounds &workarounds);
 
  void init_device_format(const VKWorkarounds &workarounds);
  void make_device_compatible(GPUVertAttr &vertex_attribute,
                              const VKWorkarounds &workarounds,
                              bool &needs_repack) const;
 
  void convert_row(void *device_row_data, const void *source_row_data) const;
  void convert_attribute(void *device_row_data,
                         const void *source_row_data,
                         const GPUVertAttr &device_attribute,
                         const GPUVertAttr &source_attribute) const;
};
 
/* \} */
 
/* -------------------------------------------------------------------- */
template<bool HasSignBit, uint8_t MantissaBitLen, uint8_t ExponentBitLen>
class FloatingPointFormat {
 public:
  static constexpr bool HAS_SIGN = HasSignBit;
  static constexpr uint8_t SIGN_SHIFT = MantissaBitLen + ExponentBitLen;
  static constexpr uint32_t SIGN_MASK = HasSignBit ? 1 : 0;
  static constexpr uint8_t MANTISSA_LEN = MantissaBitLen;
  static constexpr uint8_t MANTISSA_SHIFT = 0;
  static constexpr uint32_t MANTISSA_MASK = (1 << MantissaBitLen) - 1;
  static constexpr uint32_t MANTISSA_NAN_MASK = MANTISSA_MASK;
  static constexpr uint8_t EXPONENT_SHIFT = MantissaBitLen;
  static constexpr uint8_t EXPONENT_LEN = ExponentBitLen;
  static constexpr uint32_t EXPONENT_MASK = (1 << ExponentBitLen) - 1;
  static constexpr int32_t EXPONENT_BIAS = (1 << (ExponentBitLen - 1)) - 1;
  static constexpr int32_t EXPONENT_SPECIAL_MASK = EXPONENT_MASK;
 
  static uint32_t get_mantissa(uint32_t floating_point_number)
  {
    return (floating_point_number >> MANTISSA_SHIFT) & MANTISSA_MASK;
  }
  static uint32_t clear_mantissa(uint32_t floating_point_number)
  {
    return floating_point_number & ~(MANTISSA_MASK << MANTISSA_SHIFT);
  }
  static uint32_t set_mantissa(uint32_t mantissa, uint32_t floating_point_number)
  {
    uint32_t result = clear_mantissa(floating_point_number);
    result |= mantissa << MANTISSA_SHIFT;
    return result;
  }
 
  static uint32_t get_exponent(uint32_t floating_point_number)
  {
    return ((floating_point_number >> EXPONENT_SHIFT) & EXPONENT_MASK);
  }
  static uint32_t clear_exponent(uint32_t floating_point_number)
  {
    return floating_point_number & ~(EXPONENT_MASK << EXPONENT_SHIFT);
  }
  static uint32_t set_exponent(uint32_t exponent, uint32_t floating_point_number)
  {
    uint32_t result = clear_exponent(floating_point_number);
    result |= (exponent) << EXPONENT_SHIFT;
    return result;
  }
 
  static bool is_signed(uint32_t floating_point_number)
  {
    if constexpr (HasSignBit) {
      return (floating_point_number >> SIGN_SHIFT) & SIGN_MASK;
    }
    return false;
  }
  static uint32_t clear_sign(uint32_t floating_point_number)
  {
    return floating_point_number & ~(1 << SIGN_SHIFT);
  }
 
  static uint32_t set_sign(bool sign, uint32_t floating_point_number)
  {
    if constexpr (!HasSignBit) {
      return floating_point_number;
    }
    uint32_t result = clear_sign(floating_point_number);
    result |= uint32_t(sign) << SIGN_SHIFT;
    return result;
  }
};
 
using FormatF32 = FloatingPointFormat<true, 23, 8>;
using FormatF11 = FloatingPointFormat<false, 6, 5>;
using FormatF10 = FloatingPointFormat<false, 5, 5>;
 
template<
    typename DestinationFormat,
 
    typename SourceFormat,
 
    bool ClampNegativeToZero = true>
uint32_t convert_float_formats(uint32_t value)
{
  bool is_signed = SourceFormat::is_signed(value);
  uint32_t mantissa = SourceFormat::get_mantissa(value);
  int32_t exponent = SourceFormat::get_exponent(value);
 
  const bool is_nan = (exponent == SourceFormat::EXPONENT_SPECIAL_MASK) && mantissa;
  const bool is_inf = (exponent == SourceFormat::EXPONENT_SPECIAL_MASK) && (mantissa == 0);
  const bool is_zero = (exponent == 0 && mantissa == 0);
 
  /* Sign conversion */
  if constexpr (!DestinationFormat::HAS_SIGN && ClampNegativeToZero) {
    if (is_signed && !is_nan) {
      return 0;
    }
  }
  if (is_zero) {
    return 0;
  }
 
  if (is_inf) {
    exponent = DestinationFormat::EXPONENT_SPECIAL_MASK;
  }
  else if (is_nan) {
    exponent = DestinationFormat::EXPONENT_SPECIAL_MASK;
    mantissa = DestinationFormat::MANTISSA_NAN_MASK;
  }
  else {
    /* Exponent conversion */
    exponent -= SourceFormat::EXPONENT_BIAS;
    /* Clamping when destination has lower precision. */
    if constexpr (SourceFormat::EXPONENT_LEN > DestinationFormat::EXPONENT_LEN) {
      if (exponent > DestinationFormat::EXPONENT_BIAS) {
        exponent = 0;
        mantissa = SourceFormat::MANTISSA_MASK;
      }
      else if (exponent < -DestinationFormat::EXPONENT_BIAS) {
        return 0;
      }
    }
    exponent += DestinationFormat::EXPONENT_BIAS;
 
    /* Mantissa conversion */
    if constexpr (SourceFormat::MANTISSA_LEN > DestinationFormat::MANTISSA_LEN) {
      mantissa = mantissa >> (SourceFormat::MANTISSA_LEN - DestinationFormat::MANTISSA_LEN);
    }
    else if constexpr (SourceFormat::MANTISSA_LEN < DestinationFormat::MANTISSA_LEN) {
      mantissa = mantissa << (DestinationFormat::MANTISSA_LEN - SourceFormat::MANTISSA_LEN);
    }
  }
 
  uint32_t result = 0;
  result = DestinationFormat::set_sign(is_signed, result);
  result = DestinationFormat::set_exponent(exponent, result);
  result = DestinationFormat::set_mantissa(mantissa, result);
  return result;
}
 
/* \} */
 
};  // namespace blender::gpu