This repository has been archived on 2023-10-09. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
blender-archive/source/blender/imbuf/intern/util_gpu.c
Jeroen Bakker 37533cd6cb ImBuf: Optimize GPU memory by using 1 component format for grayscale images
This is done by checking the number of bitplanes from the image buffer.
We assume that for float buffer to use the same bitplanes as it was a
byte buffer.

Then, the data of the image buffer is packed at the start of the `rect` or
`float_rect` before upload.

**Statistics - einar.v004.blend **

Note that not all grayscale textures have been stored as BW images so the
amount of memory that can be reduced would be more.

Without patch
```
104 Textures - 3294.99 MB (3294.47 MB over 32x32), 37 RTs - 192.52 MB.
Avg. tex dimension: 2201.88x1253.51 (2283.53x2202.13 over 32x32)
464 Buffers - 25.01 MB total 1.24 MB IBs 23.50 MB VBs.
3512.52 MB - Grand total GPU buffer + texture load
```

Patch applied
```
104 Textures - 2917.66 MB (2917.14 MB over 32x32), 39 RTs - 215.45 MB.
Avg. tex dimension: 2221.38x1252.75 (2323.28x2253.47 over 32x32)
467 Buffers - 25.01 MB total 1.24 MB IBs 23.51 MB VBs.
3158.13 MB - Grand total GPU buffer + texture load.
```

Reviewed By: fclem

Differential Revision: https://developer.blender.org/D15484
2022-08-23 14:35:39 +02:00

371 lines
13 KiB
C

/* SPDX-License-Identifier: GPL-2.0-or-later
* Copyright 2001-2002 NaN Holding BV. All rights reserved. */
/** \file
* \ingroup imbuf
*/
#include "imbuf.h"
#include "BLI_math.h"
#include "BLI_utildefines.h"
#include "MEM_guardedalloc.h"
#include "BKE_global.h"
#include "GPU_capabilities.h"
#include "GPU_state.h"
#include "GPU_texture.h"
#include "IMB_colormanagement.h"
#include "IMB_imbuf.h"
#include "IMB_imbuf_types.h"
/* gpu ibuf utils */
static bool imb_is_grayscale_texture_format_compatible(const ImBuf *ibuf)
{
if (ibuf->planes > 8) {
return false;
}
/* Only imbufs with colorspace that do not modify the chrominance of the texture data relative
* to the scene color space can be uploaded as single channel textures. */
if (IMB_colormanagement_space_is_data(ibuf->rect_colorspace) ||
IMB_colormanagement_space_is_srgb(ibuf->rect_colorspace) ||
IMB_colormanagement_space_is_scene_linear(ibuf->rect_colorspace)) {
return true;
};
return false;
}
static void imb_gpu_get_format(const ImBuf *ibuf,
bool high_bitdepth,
bool use_grayscale,
eGPUDataFormat *r_data_format,
eGPUTextureFormat *r_texture_format)
{
const bool float_rect = (ibuf->rect_float != NULL);
const bool is_grayscale = use_grayscale && imb_is_grayscale_texture_format_compatible(ibuf);
if (float_rect) {
/* Float. */
const bool use_high_bitdepth = (!(ibuf->flags & IB_halffloat) && high_bitdepth);
*r_data_format = GPU_DATA_FLOAT;
*r_texture_format = is_grayscale ? (use_high_bitdepth ? GPU_R32F : GPU_R16F) :
(use_high_bitdepth ? GPU_RGBA32F : GPU_RGBA16F);
}
else {
if (IMB_colormanagement_space_is_data(ibuf->rect_colorspace) ||
IMB_colormanagement_space_is_scene_linear(ibuf->rect_colorspace)) {
/* Non-color data or scene linear, just store buffer as is. */
*r_data_format = GPU_DATA_UBYTE;
*r_texture_format = (is_grayscale) ? GPU_R8 : GPU_RGBA8;
}
else if (IMB_colormanagement_space_is_srgb(ibuf->rect_colorspace)) {
/* sRGB, store as byte texture that the GPU can decode directly. */
*r_data_format = (is_grayscale) ? GPU_DATA_FLOAT : GPU_DATA_UBYTE;
*r_texture_format = (is_grayscale) ? GPU_R16F : GPU_SRGB8_A8;
}
else {
/* Other colorspace, store as half float texture to avoid precision loss. */
*r_data_format = GPU_DATA_FLOAT;
*r_texture_format = (is_grayscale) ? GPU_R16F : GPU_RGBA16F;
}
}
}
static const char *imb_gpu_get_swizzle(const ImBuf *ibuf)
{
return imb_is_grayscale_texture_format_compatible(ibuf) ? "rrra" : "rgba";
}
/* Return false if no suitable format was found. */
#ifdef WITH_DDS
static bool IMB_gpu_get_compressed_format(const ImBuf *ibuf, eGPUTextureFormat *r_texture_format)
{
/* For DDS we only support data, scene linear and sRGB. Converting to
* different colorspace would break the compression. */
const bool use_srgb = (!IMB_colormanagement_space_is_data(ibuf->rect_colorspace) &&
!IMB_colormanagement_space_is_scene_linear(ibuf->rect_colorspace));
if (ibuf->dds_data.fourcc == FOURCC_DXT1) {
*r_texture_format = (use_srgb) ? GPU_SRGB8_A8_DXT1 : GPU_RGBA8_DXT1;
}
else if (ibuf->dds_data.fourcc == FOURCC_DXT3) {
*r_texture_format = (use_srgb) ? GPU_SRGB8_A8_DXT3 : GPU_RGBA8_DXT3;
}
else if (ibuf->dds_data.fourcc == FOURCC_DXT5) {
*r_texture_format = (use_srgb) ? GPU_SRGB8_A8_DXT5 : GPU_RGBA8_DXT5;
}
else {
return false;
}
return true;
}
#endif
/**
* Apply colormanagement and scale buffer if needed.
* `*r_freedata` is set to true if the returned buffer need to be manually freed.
*/
static void *imb_gpu_get_data(const ImBuf *ibuf,
const bool do_rescale,
const int rescale_size[2],
const bool store_premultiplied,
bool *r_freedata)
{
bool is_float_rect = (ibuf->rect_float != NULL);
const bool is_grayscale = imb_is_grayscale_texture_format_compatible(ibuf);
void *data_rect = (is_float_rect) ? (void *)ibuf->rect_float : (void *)ibuf->rect;
bool freedata = false;
if (is_float_rect) {
/* Float image is already in scene linear colorspace or non-color data by
* convention, no colorspace conversion needed. But we do require 4 channels
* currently. */
if (ibuf->channels != 4 || !store_premultiplied) {
data_rect = MEM_mallocN(sizeof(float[4]) * ibuf->x * ibuf->y, __func__);
*r_freedata = freedata = true;
if (data_rect == NULL) {
return NULL;
}
IMB_colormanagement_imbuf_to_float_texture(
(float *)data_rect, 0, 0, ibuf->x, ibuf->y, ibuf, store_premultiplied);
}
}
else {
/* Byte image is in original colorspace from the file, and may need conversion.
*
* We must also convert to premultiplied for correct texture interpolation
* and consistency with float images. */
if (IMB_colormanagement_space_is_data(ibuf->rect_colorspace)) {
/* Non-color data, just store buffer as is. */
}
else if (IMB_colormanagement_space_is_srgb(ibuf->rect_colorspace) ||
IMB_colormanagement_space_is_scene_linear(ibuf->rect_colorspace)) {
/* sRGB or scene linear, store as byte texture that the GPU can decode directly. */
data_rect = MEM_mallocN(
(is_grayscale ? sizeof(float[4]) : sizeof(uchar[4])) * ibuf->x * ibuf->y, __func__);
*r_freedata = freedata = true;
if (data_rect == NULL) {
return NULL;
}
/* Texture storage of images is defined by the alpha mode of the image. The
* downside of this is that there can be artifacts near alpha edges. However,
* this allows us to use sRGB texture formats and preserves color values in
* zero alpha areas, and appears generally closer to what game engines that we
* want to be compatible with do. */
if (is_grayscale) {
/* Convert to byte buffer to then pack as half floats reducing the buffer size by half. */
IMB_colormanagement_imbuf_to_float_texture(
(float *)data_rect, 0, 0, ibuf->x, ibuf->y, ibuf, store_premultiplied);
is_float_rect = true;
}
else {
IMB_colormanagement_imbuf_to_byte_texture(
(uchar *)data_rect, 0, 0, ibuf->x, ibuf->y, ibuf, store_premultiplied);
}
}
else {
/* Other colorspace, store as float texture to avoid precision loss. */
data_rect = MEM_mallocN(sizeof(float[4]) * ibuf->x * ibuf->y, __func__);
*r_freedata = freedata = true;
if (data_rect == NULL) {
return NULL;
}
/* Texture storage of images is defined by the alpha mode of the image. The
* downside of this is that there can be artifacts near alpha edges. However,
* this allows us to use sRGB texture formats and preserves color values in
* zero alpha areas, and appears generally closer to what game engines that we
* want to be compatible with do. */
IMB_colormanagement_imbuf_to_float_texture(
(float *)data_rect, 0, 0, ibuf->x, ibuf->y, ibuf, store_premultiplied);
}
}
if (do_rescale) {
uint *rect = (is_float_rect) ? NULL : (uint *)data_rect;
float *rect_float = (is_float_rect) ? (float *)data_rect : NULL;
ImBuf *scale_ibuf = IMB_allocFromBuffer(rect, rect_float, ibuf->x, ibuf->y, 4);
IMB_scaleImBuf(scale_ibuf, UNPACK2(rescale_size));
if (freedata) {
MEM_freeN(data_rect);
}
data_rect = (is_float_rect) ? (void *)scale_ibuf->rect_float : (void *)scale_ibuf->rect;
*r_freedata = freedata = true;
/* Steal the rescaled buffer to avoid double free. */
scale_ibuf->rect_float = NULL;
scale_ibuf->rect = NULL;
IMB_freeImBuf(scale_ibuf);
}
/* Pack first channel data manually at the start of the buffer. */
if (is_grayscale) {
void *src_rect = data_rect;
if (freedata == false) {
data_rect = MEM_mallocN((is_float_rect ? sizeof(float) : sizeof(uchar)) * ibuf->x * ibuf->y,
__func__);
*r_freedata = freedata = true;
}
if (data_rect == NULL) {
return NULL;
}
if (is_float_rect) {
for (uint64_t i = 0; i < ibuf->x * ibuf->y; i++) {
((float *)data_rect)[i] = ((float *)src_rect)[i * 4];
}
}
else {
for (uint64_t i = 0; i < ibuf->x * ibuf->y; i++) {
((uchar *)data_rect)[i] = ((uchar *)src_rect)[i * 4];
}
}
}
return data_rect;
}
GPUTexture *IMB_touch_gpu_texture(const char *name,
ImBuf *ibuf,
int w,
int h,
int layers,
bool use_high_bitdepth,
bool use_grayscale)
{
eGPUDataFormat data_format;
eGPUTextureFormat tex_format;
imb_gpu_get_format(ibuf, use_high_bitdepth, use_grayscale, &data_format, &tex_format);
GPUTexture *tex;
if (layers > 0) {
tex = GPU_texture_create_2d_array(name, w, h, layers, 9999, tex_format, NULL);
}
else {
tex = GPU_texture_create_2d(name, w, h, 9999, tex_format, NULL);
}
GPU_texture_swizzle_set(tex, imb_gpu_get_swizzle(ibuf));
GPU_texture_anisotropic_filter(tex, true);
return tex;
}
void IMB_update_gpu_texture_sub(GPUTexture *tex,
ImBuf *ibuf,
int x,
int y,
int z,
int w,
int h,
bool use_high_bitdepth,
bool use_grayscale,
bool use_premult)
{
const bool do_rescale = (ibuf->x != w || ibuf->y != h);
const int size[2] = {w, h};
eGPUDataFormat data_format;
eGPUTextureFormat tex_format;
imb_gpu_get_format(ibuf, use_high_bitdepth, use_grayscale, &data_format, &tex_format);
bool freebuf = false;
void *data = imb_gpu_get_data(ibuf, do_rescale, size, use_premult, &freebuf);
/* Update Texture. */
GPU_texture_update_sub(tex, data_format, data, x, y, z, w, h, 1);
if (freebuf) {
MEM_freeN(data);
}
}
GPUTexture *IMB_create_gpu_texture(const char *name,
ImBuf *ibuf,
bool use_high_bitdepth,
bool use_premult)
{
GPUTexture *tex = NULL;
int size[2] = {GPU_texture_size_with_limit(ibuf->x), GPU_texture_size_with_limit(ibuf->y)};
bool do_rescale = (ibuf->x != size[0]) || (ibuf->y != size[1]);
#ifdef WITH_DDS
if (ibuf->ftype == IMB_FTYPE_DDS) {
eGPUTextureFormat compressed_format;
if (!IMB_gpu_get_compressed_format(ibuf, &compressed_format)) {
fprintf(stderr, "Unable to find a suitable DXT compression,");
}
else if (do_rescale) {
fprintf(stderr, "Unable to load DXT image resolution,");
}
else if (!is_power_of_2_i(ibuf->x) || !is_power_of_2_i(ibuf->y)) {
fprintf(stderr, "Unable to load non-power-of-two DXT image resolution,");
}
else {
tex = GPU_texture_create_compressed_2d(name,
ibuf->x,
ibuf->y,
ibuf->dds_data.nummipmaps,
compressed_format,
ibuf->dds_data.data);
if (tex != NULL) {
return tex;
}
fprintf(stderr, "ST3C support not found,");
}
/* Fallback to uncompressed texture. */
fprintf(stderr, " falling back to uncompressed.\n");
}
#endif
eGPUDataFormat data_format;
eGPUTextureFormat tex_format;
imb_gpu_get_format(ibuf, use_high_bitdepth, true, &data_format, &tex_format);
bool freebuf = false;
/* Create Texture. */
tex = GPU_texture_create_2d(name, UNPACK2(size), 9999, tex_format, NULL);
if (tex == NULL) {
size[0] = max_ii(1, size[0] / 2);
size[1] = max_ii(1, size[1] / 2);
tex = GPU_texture_create_2d(name, UNPACK2(size), 9999, tex_format, NULL);
do_rescale = true;
}
BLI_assert(tex != NULL);
void *data = imb_gpu_get_data(ibuf, do_rescale, size, use_premult, &freebuf);
GPU_texture_update(tex, data_format, data);
GPU_texture_swizzle_set(tex, imb_gpu_get_swizzle(ibuf));
GPU_texture_anisotropic_filter(tex, true);
if (freebuf) {
MEM_freeN(data);
}
return tex;
}
eGPUTextureFormat IMB_gpu_get_texture_format(const ImBuf *ibuf, bool high_bitdepth, bool use_grayscale)
{
eGPUTextureFormat gpu_texture_format;
eGPUDataFormat gpu_data_format;
imb_gpu_get_format(ibuf, high_bitdepth, use_grayscale, &gpu_data_format, &gpu_texture_format);
return gpu_texture_format;
}