// SPDX-License-Identifier: Apache-2.0

// ----------------------------------------------------------------------------

// Copyright 2011-2026 Arm Limited

//

// Licensed under the Apache License, Version 2.0 (the "License"); you may not

// use this file except in compliance with the License. You may obtain a copy

// of the License at:

//

// http://www.apache.org/licenses/LICENSE-2.0

//

// Unless required by applicable law or agreed to in writing, software

// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT

// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the

// License for the specific language governing permissions and limitations

// under the License.

// ----------------------------------------------------------------------------

/**

* @brief Functions for the library entrypoint.

*/

#include <array>

#include <cstring>

#include <new>

#include "astcenc.h"

#include "astcenc_internal_entry.h"

#include "astcenc_diagnostic_trace.h"

/**

* @brief Record of the quality tuning parameter values.

*

* See the @c astcenc_config structure for detailed parameter documentation.

*

* Note that the mse_overshoot entries are scaling factors relative to the base MSE to hit db_limit.

* A 20% overshoot is harder to hit for a higher base db_limit, so we may actually use lower ratios

* for the more through search presets because the underlying db_limit is so much higher.

*/

struct astcenc_preset_config

{

float quality;

unsigned int tune_partition_count_limit;

unsigned int tune_2partition_index_limit;

unsigned int tune_3partition_index_limit;

unsigned int tune_4partition_index_limit;

unsigned int tune_block_mode_limit;

unsigned int tune_refinement_limit;

unsigned int tune_candidate_limit;

unsigned int tune_2partitioning_candidate_limit;

unsigned int tune_3partitioning_candidate_limit;

unsigned int tune_4partitioning_candidate_limit;

float tune_db_limit_a_base;

float tune_db_limit_b_base;

float tune_mse_overshoot;

float tune_2partition_early_out_limit_factor;

float tune_3partition_early_out_limit_factor;

float tune_2plane_early_out_limit_correlation;

float tune_search_mode0_enable;

};

/**

* @brief The static presets for high bandwidth encodings (x < 25 texels per block).

*/

static const std::array<astcenc_preset_config, 6> preset_configs_high {{

{

ASTCENC_PRE_FASTEST,

2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f, 0.0f

}, {

ASTCENC_PRE_FAST,

3, 18, 10, 8, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.90f, 0.0f

}, {

ASTCENC_PRE_MEDIUM,

4, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 2.5f, 1.1f, 1.05f, 0.95f, 0.0f

}, {

ASTCENC_PRE_THOROUGH,

4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.35f, 1.15f, 0.97f, 0.0f

}, {

ASTCENC_PRE_VERYTHOROUGH,

4, 256, 128, 64, 98, 4, 6, 8, 6, 4, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 0.0f

}, {

ASTCENC_PRE_EXHAUSTIVE,

4, 512, 512, 512, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 0.0f

}

}};

/**

* @brief The static presets for medium bandwidth encodings (25 <= x < 64 texels per block).

*/

static const std::array<astcenc_preset_config, 6> preset_configs_mid {{

{

ASTCENC_PRE_FASTEST,

2, 10, 6, 4, 43, 2, 2, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.80f, 1.0f

}, {

ASTCENC_PRE_FAST,

3, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.2f, 63.2f, 3.5f, 1.0f, 1.0f, 0.85f, 1.0f

}, {

ASTCENC_PRE_MEDIUM,

3, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.0f, 1.1f, 1.05f, 0.90f, 1.0f

}, {

ASTCENC_PRE_THOROUGH,

4, 82, 60, 30, 94, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.4f, 1.2f, 0.95f, 0.0f

}, {

ASTCENC_PRE_VERYTHOROUGH,

4, 256, 128, 64, 98, 4, 6, 8, 6, 3, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 0.0f

}, {

ASTCENC_PRE_EXHAUSTIVE,

4, 256, 256, 256, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 0.0f

}

}};

/**

* @brief The static presets for low bandwidth encodings (64 <= x texels per block).

*/

static const std::array<astcenc_preset_config, 6> preset_configs_low {{

{

ASTCENC_PRE_FASTEST,

2, 10, 6, 4, 40, 2, 2, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.80f, 1.0f

}, {

ASTCENC_PRE_FAST,

2, 18, 12, 10, 55, 3, 3, 2, 2, 2, 85.0f, 63.0f, 3.5f, 1.0f, 1.0f, 0.85f, 1.0f

}, {

ASTCENC_PRE_MEDIUM,

3, 34, 28, 16, 77, 3, 3, 2, 2, 2, 95.0f, 70.0f, 3.5f, 1.1f, 1.05f, 0.90f, 1.0f

}, {

ASTCENC_PRE_THOROUGH,

4, 82, 60, 30, 93, 4, 4, 3, 2, 2, 105.0f, 77.0f, 10.0f, 1.3f, 1.2f, 0.97f, 1.0f

}, {

ASTCENC_PRE_VERYTHOROUGH,

4, 256, 128, 64, 98, 4, 6, 8, 5, 2, 200.0f, 200.0f, 10.0f, 1.6f, 1.4f, 0.98f, 1.0f

}, {

ASTCENC_PRE_EXHAUSTIVE,

4, 256, 256, 256, 100, 4, 8, 8, 8, 8, 200.0f, 200.0f, 10.0f, 2.0f, 2.0f, 0.99f, 1.0f

}

}};

/**

* @brief Validate CPU floating point meets assumptions made in the codec.

*

* The codec is written with the assumption that float bit patterns are valid

* IEEE754 values that are stored and reloaded with round-to-nearest rounding.

* This is always the case in an IEEE-754 compliant system, however not every

* system or compilation mode is actually IEEE-754 compliant. This normally

* fails if the code is compiled with fast math enabled, for example.

*

* @return Return @c ASTCENC_SUCCESS if validated, an error on failure.

*/

static astcenc_error validate_cpu_float()

{

volatile float xprec_testval = 2.51f;

float store = xprec_testval + 12582912.0f;

float q = store - 12582912.0f;

if (q != 3.0f)

{

return ASTCENC_ERR_BAD_CPU_FLOAT;

}

return ASTCENC_SUCCESS;

}

/**

* @brief Validate config profile.

*

* @param profile The profile to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_profile(

astcenc_profile profile

) {

// Values in this enum are from an external user, so not guaranteed to be

// bounded to the enum values

switch (static_cast<int>(profile))

{

case ASTCENC_PRF_LDR_SRGB:

case ASTCENC_PRF_LDR:

case ASTCENC_PRF_HDR_RGB_LDR_A:

case ASTCENC_PRF_HDR:

return ASTCENC_SUCCESS;

default:

return ASTCENC_ERR_BAD_PROFILE;

}

}

/**

* @brief Validate block size.

*

* @param block_x The block x dimensions.

* @param block_y The block y dimensions.

* @param block_z The block z dimensions.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_block_size(

unsigned int block_x,

unsigned int block_y,

unsigned int block_z

) {

// Test if this is a legal block size at all

bool is_legal = (((block_z <= 1) && is_legal_2d_block_size(block_x, block_y)) ||

((block_z >= 2) && is_legal_3d_block_size(block_x, block_y, block_z)));

if (!is_legal)

{

return ASTCENC_ERR_BAD_BLOCK_SIZE;

}

// Test if this build has sufficient capacity for this block size

bool have_capacity = (block_x * block_y * block_z) <= BLOCK_MAX_TEXELS;

if (!have_capacity)

{

return ASTCENC_ERR_NOT_IMPLEMENTED;

}

return ASTCENC_SUCCESS;

}

/**

* @brief Validate flags.

*

* @param profile The profile to check.

* @param flags The flags to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_flags(

astcenc_profile profile,

unsigned int flags

) {

// Flags field must not contain any unknown flag bits

unsigned int exMask = ~ASTCENC_ALL_FLAGS;

if (popcount(flags & exMask) != 0)

{

return ASTCENC_ERR_BAD_FLAGS;

}

// Flags field must only contain at most a single map type

exMask = ASTCENC_FLG_MAP_NORMAL

| ASTCENC_FLG_MAP_RGBM;

if (popcount(flags & exMask) > 1)

{

return ASTCENC_ERR_BAD_FLAGS;

}

// Decode_unorm8 must only be used with an LDR profile

bool is_unorm8 = flags & ASTCENC_FLG_USE_DECODE_UNORM8;

bool is_hdr = (profile == ASTCENC_PRF_HDR) || (profile == ASTCENC_PRF_HDR_RGB_LDR_A);

if (is_unorm8 && is_hdr)

{

return ASTCENC_ERR_BAD_DECODE_MODE;

}

return ASTCENC_SUCCESS;

}

#if !defined(ASTCENC_DECOMPRESS_ONLY)

/**

* @brief Validate single channel compression swizzle.

*

* @param swizzle The swizzle to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_compression_swz(

astcenc_swz swizzle

) {

// Not all enum values are handled; SWZ_Z is invalid for compression

switch (static_cast<int>(swizzle))

{

case ASTCENC_SWZ_R:

case ASTCENC_SWZ_G:

case ASTCENC_SWZ_B:

case ASTCENC_SWZ_A:

case ASTCENC_SWZ_0:

case ASTCENC_SWZ_1:

return ASTCENC_SUCCESS;

default:

return ASTCENC_ERR_BAD_SWIZZLE;

}

}

/**

* @brief Validate overall compression swizzle.

*

* @param swizzle The swizzle to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_compression_swizzle(

const astcenc_swizzle& swizzle

) {

if (validate_compression_swz(swizzle.r) ||

validate_compression_swz(swizzle.g) ||

validate_compression_swz(swizzle.b) ||

validate_compression_swz(swizzle.a))

{

return ASTCENC_ERR_BAD_SWIZZLE;

}

return ASTCENC_SUCCESS;

}

#endif

/**

* @brief Validate single channel decompression swizzle.

*

* @param swizzle The swizzle to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_decompression_swz(

astcenc_swz swizzle

) {

// Values in this enum are from an external user, so not guaranteed to be

// bounded to the enum values

switch (static_cast<int>(swizzle))

{

case ASTCENC_SWZ_R:

case ASTCENC_SWZ_G:

case ASTCENC_SWZ_B:

case ASTCENC_SWZ_A:

case ASTCENC_SWZ_0:

case ASTCENC_SWZ_1:

case ASTCENC_SWZ_Z:

return ASTCENC_SUCCESS;

default:

return ASTCENC_ERR_BAD_SWIZZLE;

}

}

/**

* @brief Validate overall decompression swizzle.

*

* @param swizzle The swizzle to check.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_decompression_swizzle(

const astcenc_swizzle& swizzle

) {

if (validate_decompression_swz(swizzle.r) ||

validate_decompression_swz(swizzle.g) ||

validate_decompression_swz(swizzle.b) ||

validate_decompression_swz(swizzle.a))

{

return ASTCENC_ERR_BAD_SWIZZLE;

}

return ASTCENC_SUCCESS;

}

/**

* Validate that an incoming configuration is in-spec.

*

* This function can respond in two ways:

*

* * Numerical inputs that have valid ranges are clamped to those valid ranges. No error is thrown

* for out-of-range inputs in this case.

* * Numerical inputs and logic inputs are are logically invalid and which make no sense

* algorithmically will return an error.

*

* @param[in,out] config The input compressor configuration.

*

* @return Return @c ASTCENC_SUCCESS if validated, otherwise an error on failure.

*/

static astcenc_error validate_config(

astcenc_config &config

) {

astcenc_error status;

status = validate_profile(config.profile);

if (status != ASTCENC_SUCCESS)

{

return status;

}

status = validate_flags(config.profile, config.flags);

if (status != ASTCENC_SUCCESS)

{

return status;

}

status = validate_block_size(config.block_x, config.block_y, config.block_z);

if (status != ASTCENC_SUCCESS)

{

return status;

}

#if defined(ASTCENC_DECOMPRESS_ONLY)

// Decompress-only builds only support decompress-only contexts

if (!(config.flags & ASTCENC_FLG_DECOMPRESS_ONLY))

{

return ASTCENC_ERR_BAD_PARAM;

}

#endif

config.rgbm_m_scale = astc::max(config.rgbm_m_scale, 1.0f);

config.tune_partition_count_limit = astc::clamp(config.tune_partition_count_limit, 1u, 4u);

config.tune_2partition_index_limit = astc::clamp(config.tune_2partition_index_limit, 1u, BLOCK_MAX_PARTITIONINGS);

config.tune_3partition_index_limit = astc::clamp(config.tune_3partition_index_limit, 1u, BLOCK_MAX_PARTITIONINGS);

config.tune_4partition_index_limit = astc::clamp(config.tune_4partition_index_limit, 1u, BLOCK_MAX_PARTITIONINGS);

config.tune_block_mode_limit = astc::clamp(config.tune_block_mode_limit, 1u, 100u);

config.tune_refinement_limit = astc::max(config.tune_refinement_limit, 1u);

config.tune_candidate_limit = astc::clamp(config.tune_candidate_limit, 1u, TUNE_MAX_TRIAL_CANDIDATES);

config.tune_2partitioning_candidate_limit = astc::clamp(config.tune_2partitioning_candidate_limit, 1u, TUNE_MAX_PARTITIONING_CANDIDATES);

config.tune_3partitioning_candidate_limit = astc::clamp(config.tune_3partitioning_candidate_limit, 1u, TUNE_MAX_PARTITIONING_CANDIDATES);

config.tune_4partitioning_candidate_limit = astc::clamp(config.tune_4partitioning_candidate_limit, 1u, TUNE_MAX_PARTITIONING_CANDIDATES);

config.tune_db_limit = astc::max(config.tune_db_limit, 0.0f);

config.tune_mse_overshoot = astc::max(config.tune_mse_overshoot, 1.0f);

config.tune_2partition_early_out_limit_factor = astc::max(config.tune_2partition_early_out_limit_factor, 0.0f);

config.tune_3partition_early_out_limit_factor = astc::max(config.tune_3partition_early_out_limit_factor, 0.0f);

config.tune_2plane_early_out_limit_correlation = astc::max(config.tune_2plane_early_out_limit_correlation, 0.0f);

// Specifying a zero weight color component is not allowed; force to small value

float max_weight = astc::max(astc::max(config.cw_r_weight, config.cw_g_weight),

astc::max(config.cw_b_weight, config.cw_a_weight));

if (max_weight > 0.0f)

{

max_weight /= 1000.0f;

config.cw_r_weight = astc::max(config.cw_r_weight, max_weight);

config.cw_g_weight = astc::max(config.cw_g_weight, max_weight);

config.cw_b_weight = astc::max(config.cw_b_weight, max_weight);

config.cw_a_weight = astc::max(config.cw_a_weight, max_weight);

}

// If all color components error weights are zero then return an error

else

{

return ASTCENC_ERR_BAD_PARAM;

}

return ASTCENC_SUCCESS;

}

/* See header for documentation. */

astcenc_error astcenc_config_init(

astcenc_profile profile,

unsigned int block_x,

unsigned int block_y,

unsigned int block_z,

float quality,

unsigned int flags,

astcenc_config* configp

) {

astcenc_error status;

status = validate_cpu_float();

if (status != ASTCENC_SUCCESS)

{

return status;

}

// Zero init all config fields; although most of will be over written

astcenc_config& config = *configp;

std::memset(&config, 0, sizeof(config));

// Process the block size

block_z = astc::max(block_z, 1u); // For 2D blocks Z==0 is accepted, but convert to 1

status = validate_block_size(block_x, block_y, block_z);

if (status != ASTCENC_SUCCESS)

{

return status;

}

config.block_x = block_x;

config.block_y = block_y;

config.block_z = block_z;

float texels = static_cast<float>(block_x * block_y * block_z);

float ltexels = logf(texels) / logf(10.0f);

// Process the performance quality level or preset; note that this must be done before we

// process any additional settings, such as color profile and flags, which may replace some of

// these settings with more use case tuned values

if (quality < ASTCENC_PRE_FASTEST ||

quality > ASTCENC_PRE_EXHAUSTIVE)

{

return ASTCENC_ERR_BAD_QUALITY;

}

static const std::array<astcenc_preset_config, 6>* preset_configs;

int texels_int = block_x * block_y * block_z;

if (texels_int < 25)

{

preset_configs = &preset_configs_high;

}

else if (texels_int < 64)

{

preset_configs = &preset_configs_mid;

}

else

{

preset_configs = &preset_configs_low;

}

// Determine which preset to use, or which pair to interpolate

size_t start;

size_t end;

for (end = 0; end < preset_configs->size(); end++)

{

if ((*preset_configs)[end].quality >= quality)

{

break;

}

}

start = end == 0 ? 0 : end - 1;

// Start and end node are the same - so just transfer the values.

if (start == end)

{

config.tune_partition_count_limit = (*preset_configs)[start].tune_partition_count_limit;

config.tune_2partition_index_limit = (*preset_configs)[start].tune_2partition_index_limit;

config.tune_3partition_index_limit = (*preset_configs)[start].tune_3partition_index_limit;

config.tune_4partition_index_limit = (*preset_configs)[start].tune_4partition_index_limit;

config.tune_block_mode_limit = (*preset_configs)[start].tune_block_mode_limit;

config.tune_refinement_limit = (*preset_configs)[start].tune_refinement_limit;

config.tune_candidate_limit = (*preset_configs)[start].tune_candidate_limit;

config.tune_2partitioning_candidate_limit = (*preset_configs)[start].tune_2partitioning_candidate_limit;

config.tune_3partitioning_candidate_limit = (*preset_configs)[start].tune_3partitioning_candidate_limit;

config.tune_4partitioning_candidate_limit = (*preset_configs)[start].tune_4partitioning_candidate_limit;

config.tune_db_limit = astc::max((*preset_configs)[start].tune_db_limit_a_base - 35 * ltexels,

(*preset_configs)[start].tune_db_limit_b_base - 19 * ltexels);

config.tune_mse_overshoot = (*preset_configs)[start].tune_mse_overshoot;

config.tune_2partition_early_out_limit_factor = (*preset_configs)[start].tune_2partition_early_out_limit_factor;

config.tune_3partition_early_out_limit_factor = (*preset_configs)[start].tune_3partition_early_out_limit_factor;

config.tune_2plane_early_out_limit_correlation = (*preset_configs)[start].tune_2plane_early_out_limit_correlation;

config.tune_search_mode0_enable = (*preset_configs)[start].tune_search_mode0_enable;

}

// Start and end node are not the same - so interpolate between them

else

{

auto& node_a = (*preset_configs)[start];

auto& node_b = (*preset_configs)[end];

float wt_range = node_b.quality - node_a.quality;

assert(wt_range > 0);

// Compute interpolation factors

float wt_node_a = (node_b.quality - quality) / wt_range;

float wt_node_b = (quality - node_a.quality) / wt_range;

#define LERP(param) ((node_a.param * wt_node_a) + (node_b.param * wt_node_b))

#define LERPI(param) astc::flt2int_rtn(\

(static_cast<float>(node_a.param) * wt_node_a) + \

(static_cast<float>(node_b.param) * wt_node_b))

#define LERPUI(param) static_cast<unsigned int>(LERPI(param))

config.tune_partition_count_limit = LERPI(tune_partition_count_limit);

config.tune_2partition_index_limit = LERPI(tune_2partition_index_limit);

config.tune_3partition_index_limit = LERPI(tune_3partition_index_limit);

config.tune_4partition_index_limit = LERPI(tune_4partition_index_limit);

config.tune_block_mode_limit = LERPI(tune_block_mode_limit);

config.tune_refinement_limit = LERPI(tune_refinement_limit);

config.tune_candidate_limit = LERPUI(tune_candidate_limit);

config.tune_2partitioning_candidate_limit = LERPUI(tune_2partitioning_candidate_limit);

config.tune_3partitioning_candidate_limit = LERPUI(tune_3partitioning_candidate_limit);

config.tune_4partitioning_candidate_limit = LERPUI(tune_4partitioning_candidate_limit);

config.tune_db_limit = astc::max(LERP(tune_db_limit_a_base) - 35 * ltexels,

LERP(tune_db_limit_b_base) - 19 * ltexels);

config.tune_mse_overshoot = LERP(tune_mse_overshoot);

config.tune_2partition_early_out_limit_factor = LERP(tune_2partition_early_out_limit_factor);

config.tune_3partition_early_out_limit_factor = LERP(tune_3partition_early_out_limit_factor);

config.tune_2plane_early_out_limit_correlation = LERP(tune_2plane_early_out_limit_correlation);

config.tune_search_mode0_enable = LERP(tune_search_mode0_enable);

#undef LERP

#undef LERPI

#undef LERPUI

}

// Set heuristics to the defaults for each color profile

config.cw_r_weight = 1.0f;

config.cw_g_weight = 1.0f;

config.cw_b_weight = 1.0f;

config.cw_a_weight = 1.0f;

config.a_scale_radius = 0;

config.rgbm_m_scale = 0.0f;

config.profile = profile;

// Values in this enum are from an external user, so not guaranteed to be

// bounded to the enum values

switch (static_cast<int>(profile))

{

case ASTCENC_PRF_LDR:

case ASTCENC_PRF_LDR_SRGB:

break;

case ASTCENC_PRF_HDR_RGB_LDR_A:

case ASTCENC_PRF_HDR:

config.tune_db_limit = 999.0f;

config.tune_search_mode0_enable = 0.0f;

break;

default:

return ASTCENC_ERR_BAD_PROFILE;

}

// Flags field must not contain any unknown flag bits

status = validate_flags(profile, flags);

if (status != ASTCENC_SUCCESS)

{

return status;

}

if (flags & ASTCENC_FLG_MAP_NORMAL)

{

// Normal map encoding uses L+A blocks, so allow one more partitioning

// than normal. We need need fewer bits for endpoints, so more likely

// to be able to use more partitions than an RGB/RGBA block

config.tune_partition_count_limit = astc::min(config.tune_partition_count_limit + 1u, 4u);

config.cw_g_weight = 0.0f;

config.cw_b_weight = 0.0f;

config.tune_2partition_early_out_limit_factor *= 1.5f;

config.tune_3partition_early_out_limit_factor *= 1.5f;

config.tune_2plane_early_out_limit_correlation = 0.99f;

// Normals are prone to blocking artifacts on smooth curves

// so force compressor to try harder here ...

config.tune_db_limit *= 1.03f;

}

else if (flags & ASTCENC_FLG_MAP_RGBM)

{

config.rgbm_m_scale = 5.0f;

config.cw_a_weight = 2.0f * config.rgbm_m_scale;

}

else // (This is color data)

{

// This is a very basic perceptual metric for RGB color data, which weights error

// significance by the perceptual luminance contribution of each color channel. For

// luminance the usual weights to compute luminance from a linear RGB value are as

// follows:

//

// l = r * 0.3 + g * 0.59 + b * 0.11

//

// ... but we scale these up to keep a better balance between color and alpha. Note

// that if the content is using alpha we'd recommend using the -a option to weight

// the color contribution by the alpha transparency.

if (flags & ASTCENC_FLG_USE_PERCEPTUAL)

{

config.cw_r_weight = 0.30f * 2.25f;

config.cw_g_weight = 0.59f * 2.25f;

config.cw_b_weight = 0.11f * 2.25f;

}

}

config.flags = flags;

return ASTCENC_SUCCESS;

}

/* See header for documentation. */

astcenc_error astcenc_context_alloc(

const astcenc_config* configp,

unsigned int thread_count,

astcenc_context** context,

const astcenc_context* parent_context

) {

astcenc_error status;

status = validate_cpu_float();

if (status != ASTCENC_SUCCESS)

{

return status;

}

if (thread_count == 0)

{

return ASTCENC_ERR_BAD_PARAM;

}

#if defined(ASTCENC_DIAGNOSTICS)

// Force single threaded compressor use in diagnostic mode

if (thread_count != 1)

{

return ASTCENC_ERR_BAD_PARAM;

}

#endif

// Exactly one of config or parent_context must be set

bool has_config = configp != nullptr;

bool has_parent = parent_context != nullptr;

if (!(has_config ^ has_parent))

{

return ASTCENC_ERR_BAD_PARAM;

}

if (has_parent)

{

configp = &parent_context->context.config;

}

const astcenc_config& config = *configp;

astcenc_context* ctxo = new astcenc_context;

astcenc_contexti* ctx = &ctxo->context;

ctx->thread_count = thread_count;

ctx->config = *configp;

ctx->working_buffers = nullptr;

// These are allocated per-compress, as they depend on image size

ctx->input_alpha_averages = nullptr;

// Copy the config first and validate the copy (we may modify it)

status = validate_config(ctx->config);

if (status != ASTCENC_SUCCESS)

{

delete ctxo;

return status;

}

if (!parent_context)

{

block_size_descriptor* bsd = aligned_malloc<block_size_descriptor>(sizeof(block_size_descriptor), ASTCENC_VECALIGN);

if (!bsd)

{

delete ctxo;

return ASTCENC_ERR_OUT_OF_MEM;

}

bool can_omit_modes = static_cast<bool>(config.flags & ASTCENC_FLG_SELF_DECOMPRESS_ONLY);

init_block_size_descriptor(config.block_x, config.block_y, config.block_z,

can_omit_modes,

config.tune_partition_count_limit,

static_cast<float>(config.tune_block_mode_limit) / 100.0f,

*bsd);

ctx->owns_bsd = true;

ctx->bsd = bsd;

}

else

{

ctx->owns_bsd = false;

ctx->bsd = parent_context->context.bsd;

}

#if !defined(ASTCENC_DECOMPRESS_ONLY)

// Do setup only needed by compression

if (!(ctx->config.flags & ASTCENC_FLG_DECOMPRESS_ONLY))

{

// Turn a dB limit into a per-texel error for faster use later

if ((ctx->config.profile == ASTCENC_PRF_LDR) || (ctx->config.profile == ASTCENC_PRF_LDR_SRGB))

{

ctx->config.tune_db_limit = astc::pow(0.1f, ctx->config.tune_db_limit * 0.1f) * 65535.0f * 65535.0f;

}

else

{

ctx->config.tune_db_limit = 0.0f;

}

size_t worksize = sizeof(compression_working_buffers) * thread_count;

ctx->working_buffers = aligned_malloc<compression_working_buffers>(worksize, ASTCENC_VECALIGN);

static_assert((ASTCENC_VECALIGN == 0) || ((sizeof(compression_working_buffers) % ASTCENC_VECALIGN) == 0),

"compression_working_buffers size must be multiple of vector alignment");

if (!ctx->working_buffers)

{

if (ctx->owns_bsd)

{

aligned_free<const block_size_descriptor>(ctx->bsd);

}

delete ctxo;

*context = nullptr;

return ASTCENC_ERR_OUT_OF_MEM;

}

}

#endif

#if defined(ASTCENC_DIAGNOSTICS)

ctx->trace_log = new TraceLog(ctx->config.trace_file_path);

if (!ctx->trace_log->m_file)

{

return ASTCENC_ERR_DTRACE_FAILURE;

}

trace_add_data("block_x", config.block_x);

trace_add_data("block_y", config.block_y);

trace_add_data("block_z", config.block_z);

#endif

*context = ctxo;

#if !defined(ASTCENC_DECOMPRESS_ONLY)

prepare_angular_tables();

#endif

return ASTCENC_SUCCESS;

}

/* See header dor documentation. */

void astcenc_context_free(

astcenc_context* ctxo

) {

if (ctxo)

{

astcenc_contexti* ctx = &ctxo->context;

aligned_free<compression_working_buffers>(ctx->working_buffers);

if (ctx->owns_bsd)

{

aligned_free<const block_size_descriptor>(ctx->bsd);

}

#if defined(ASTCENC_DIAGNOSTICS)

delete ctx->trace_log;

#endif

delete ctxo;

}

}

#if !defined(ASTCENC_DECOMPRESS_ONLY)

/**

* @brief Compress an image, after any preflight has completed.

*

* @param[out] ctxo The compressor context.

* @param thread_index The thread index.

* @param image The intput image.

* @param swizzle The input swizzle.

* @param[out] buffer The output array for the compressed data.

*/

static void compress_image(

astcenc_context& ctxo,

unsigned int thread_index,

const astcenc_image& image,

const astcenc_swizzle& swizzle,

uint8_t* buffer

) {

astcenc_contexti& ctx = ctxo.context;

const block_size_descriptor& bsd = *ctx.bsd;

astcenc_profile decode_mode = ctx.config.profile;

image_block blk;

int block_x = bsd.xdim;

int block_y = bsd.ydim;

int block_z = bsd.zdim;

blk.texel_count = static_cast<uint8_t>(block_x * block_y * block_z);

int dim_x = image.dim_x;

int dim_y = image.dim_y;

int dim_z = image.dim_z;

int xblocks = (dim_x + block_x - 1) / block_x;

int yblocks = (dim_y + block_y - 1) / block_y;

int zblocks = (dim_z + block_z - 1) / block_z;

int block_count = zblocks * yblocks * xblocks;

int row_blocks = xblocks;

int plane_blocks = xblocks * yblocks;

blk.decode_unorm8 = ctxo.context.config.flags & ASTCENC_FLG_USE_DECODE_UNORM8;

// Populate the block channel weights

blk.channel_weight = vfloat4(ctx.config.cw_r_weight,

ctx.config.cw_g_weight,

ctx.config.cw_b_weight,

ctx.config.cw_a_weight);

// Use preallocated scratch buffer

auto& temp_buffers = ctx.working_buffers[thread_index];

// Only the first thread actually runs the initializer

ctxo.manage_compress.init(block_count, ctx.config.progress_callback);

// Determine if we can use an optimized load function

bool needs_swz = (swizzle.r != ASTCENC_SWZ_R) || (swizzle.g != ASTCENC_SWZ_G) ||

(swizzle.b != ASTCENC_SWZ_B) || (swizzle.a != ASTCENC_SWZ_A);

bool needs_hdr = (decode_mode == ASTCENC_PRF_HDR) ||

(decode_mode == ASTCENC_PRF_HDR_RGB_LDR_A);

bool use_fast_load = !needs_swz && !needs_hdr &&

block_z == 1 && image.data_type == ASTCENC_TYPE_U8;

auto load_func = load_image_block;

if (use_fast_load)

{

load_func = load_image_block_fast_ldr;

}

// All threads run this processing loop until there is no work remaining

while (true)

{

unsigned int count;

unsigned int base = ctxo.manage_compress.get_task_assignment(16, count);

if (!count)

{

break;

}

for (unsigned int i = base; i < base + count; i++)

{

// Decode i into x, y, z block indices

int z = i / plane_blocks;

unsigned int rem = i - (z * plane_blocks);

int y = rem / row_blocks;

int x = rem - (y * row_blocks);

// Test if we can apply some basic alpha-scale RDO

bool use_full_block = true;

if (ctx.config.a_scale_radius != 0 && block_z == 1)

{

int start_x = x * block_x;

int end_x = astc::min(dim_x, start_x + block_x);

int start_y = y * block_y;

int end_y = astc::min(dim_y, start_y + block_y);

// SATs accumulate error, so don't test exactly zero. Test for

// less than 1 alpha in the expanded block footprint that

// includes the alpha radius.

int x_footprint = block_x + 2 * (ctx.config.a_scale_radius - 1);

int y_footprint = block_y + 2 * (ctx.config.a_scale_radius - 1);

float footprint = static_cast<float>(x_footprint * y_footprint);

float threshold = 0.9f / (255.0f * footprint);

// Do we have any alpha values?

use_full_block = false;

for (int ay = start_y; ay < end_y; ay++)

{

for (int ax = start_x; ax < end_x; ax++)

{

float a_avg = ctx.input_alpha_averages[ay * dim_x + ax];

if (a_avg > threshold)

{

use_full_block = true;

ax = end_x;

ay = end_y;

}

}

}

}

// Fetch the full block for compression

if (use_full_block)

{

load_func(decode_mode, image, blk, bsd, x * block_x, y * block_y, z * block_z, swizzle);

// Scale RGB error contribution by the maximum alpha in the block

// This encourages preserving alpha accuracy in regions with high

// transparency, and can buy up to 0.5 dB PSNR.

if (ctx.config.flags & ASTCENC_FLG_USE_ALPHA_WEIGHT)

{

float alpha_scale = blk.data_max.lane<3>() * (1.0f / 65535.0f);

blk.channel_weight = vfloat4(ctx.config.cw_r_weight * alpha_scale,

ctx.config.cw_g_weight * alpha_scale,

ctx.config.cw_b_weight * alpha_scale,

ctx.config.cw_a_weight);

}

}

// Apply alpha scale RDO - substitute constant color block

else

{

blk.origin_texel = vfloat4::zero();

blk.data_min = vfloat4::zero();

blk.data_mean = vfloat4::zero();

blk.data_max = vfloat4::zero();

blk.grayscale = true;

}

int offset = ((z * yblocks + y) * xblocks + x) * 16;

uint8_t *bp = buffer + offset;

compress_block(ctx, blk, bp, temp_buffers);

}

ctxo.manage_compress.complete_task_assignment(count);

}

}

/**

* @brief Compute regional averages in an image.

*

* This function can be called by multiple threads, but only after a single

* thread calls the setup function @c init_compute_averages().

*

* Results are written back into @c img->input_alpha_averages.

*

* @param[out] ctx The context.

* @param ag The average and variance arguments created during setup.

*/

static void compute_averages(

astcenc_context& ctx,

const avg_args &ag

) {

pixel_region_args arg = ag.arg;

arg.work_memory = new vfloat4[ag.work_memory_size];

int size_x = ag.img_size_x;

int size_y = ag.img_size_y;

int size_z = ag.img_size_z;

int step_xy = ag.blk_size_xy;

int step_z = ag.blk_size_z;

int y_tasks = (size_y + step_xy - 1) / step_xy;

// All threads run this processing loop until there is no work remaining