/**
* Copy from a source image into a destination image of the specified
* format and check the result.
*
* If \a strict_layout_qualifiers is false, uniform layout qualifiers
* will be omitted where allowed by the spec. If \a
* strict_access_qualifiers is false, the "readonly" and "writeonly"
* qualifiers will be omitted. If \a strict_binding is false, the
* image will be bound as READ_WRITE, otherwise only the required
* access type will be used.
*/
static bool
run_test(const struct image_format_info *format,
bool strict_layout_qualifiers,
bool strict_access_qualifiers,
bool strict_binding)
{
const struct grid_info grid =
grid_info(GL_FRAGMENT_SHADER,
image_base_internal_format(format), W, H);
const struct image_info img =
image_info(GL_TEXTURE_2D, format->format, W, H);
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(img, ""),
test_hunk(strict_layout_qualifiers,
strict_access_qualifiers),
hunk("SRC_IMAGE_Q uniform IMAGE_BARE_T src_img;\n"
"DST_IMAGE_Q uniform IMAGE_BARE_T dst_img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" imageStore(dst_img, IMAGE_ADDR(idx),"
" imageLoad(src_img, IMAGE_ADDR(idx)));\n"
" return x;\n"
"}\n"), NULL));
bool ret = prog && init_fb(grid) &&
init_image(img, 0, strict_binding) &&
init_image(img, 1, strict_binding) &&
set_uniform_int(prog, "src_img", 0) &&
set_uniform_int(prog, "dst_img", 1) &&
draw_grid(grid, prog) &&
check(grid, img);
glDeleteProgram(prog);
return ret;
}
static bool
run_test(const struct image_target_info *target,
const struct image_extent size)
{
const struct grid_info grid = {
GL_FRAGMENT_SHADER_BIT,
get_image_format(GL_RGBA32F),
image_optimal_extent(size)
};
const struct image_info img = {
target, grid.format, size,
image_format_epsilon(grid.format)
};
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(img, ""),
hunk("readonly uniform IMAGE_T src_img;\n"
"writeonly uniform IMAGE_T dst_img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" imageStore(dst_img, IMAGE_ADDR(idx),"
" imageLoad(src_img, IMAGE_ADDR(idx)));\n"
" return x;\n"
"}\n"), NULL));
bool ret = prog && init_fb(grid) &&
init_image(img, 0) &&
init_image(img, 1) &&
set_uniform_int(prog, "src_img", 0) &&
set_uniform_int(prog, "dst_img", 1) &&
draw_grid(grid, prog) &&
check(img);
glDeleteProgram(prog);
return ret;
}
/**
* Bind all image uniforms present in the program to the available
* image units, re-using the same unit several times if necessary in
* cyclical order.
*/
static bool
bind_images(const struct grid_info grid, GLuint prog)
{
const unsigned m = max_image_units();
const struct image_stage_info *stage;
for (stage = image_stages(); stage->name; ++stage) {
if (grid.stages & stage->bit) {
const unsigned first =
num_images_for_stages(grid, stage->bit - 1);
const unsigned n = num_images_for_stage(grid, stage);
const unsigned stage_idx = stage - image_stages();
int i;
for (i = 0; i < n; ++i) {
char *name = NULL;
asprintf(&name, "imgs_%d[%d]", stage_idx, i);
if (!set_uniform_int(prog, name,
(first + i) % m))
return false;
free(name);
}
}
}
return true;
}
static bool
run_test(const struct image_op_info *op,
unsigned w, unsigned h,
bool (*check)(const struct grid_info grid,
const struct image_info img,
unsigned w, unsigned h),
const char *body)
{
const struct grid_info grid =
grid_info(GL_FRAGMENT_SHADER, GL_R32UI, W, H);
const struct image_info img = image_info_for_grid(grid);
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(img, ""),
hunk("uniform IMAGE_T img;\n"),
hunk(op->hunk),
hunk(body), NULL));
bool ret = prog &&
init_fb(grid) &&
init_image(img) &&
set_uniform_int(prog, "img", 0) &&
draw_grid(set_grid_size(grid, w, h), prog) &&
check(grid, img, w, h);
glDeleteProgram(prog);
return ret;
}
static bool
run_test(const struct image_qualifier_info *qual)
{
const struct grid_info grid =
grid_info(GL_FRAGMENT_SHADER, GL_R32UI, W, H);
const struct image_info img =
image_info(GL_TEXTURE_1D, GL_R32UI, W, H);
GLuint prog = generate_program(
grid,
/**
* Write to consecutive locations of an image using a
* the value read from a fixed location of a different
* image uniform which aliases the first image. If
* the implementation incorrectly coalesces repeated
* loads from the fixed location the results of the
* test will be altered.
*/
GL_FRAGMENT_SHADER,
concat(qualifier_hunk(qual),
image_hunk(img, ""),
hunk("IMAGE_Q IMAGE_UNIFORM_T src_img;\n"
"IMAGE_Q IMAGE_UNIFORM_T dst_img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" int i;\n"
"\n"
" for (i = 0; i < N / 2; ++i) {\n"
" imageStore(dst_img, 2 * i,"
" imageLoad(src_img, W) + 1u);\n"
" imageStore(dst_img, 2 * i + 1,"
" imageLoad(src_img, W) - 1u);\n"
" }\n"
"\n"
" return x;\n"
"}\n"), NULL));
bool ret = prog &&
init_fb(grid) &&
init_image(img) &&
set_uniform_int(prog, "src_img", 0) &&
set_uniform_int(prog, "dst_img", 0) &&
draw_grid(set_grid_size(grid, 1, 1), prog) &&
(check(img) || qual->control_test);
glDeleteProgram(prog);
return ret;
}
static bool
run_test(const struct image_qualifier_info *qual,
const struct image_stage_info *stage_w,
const struct image_stage_info *stage_r,
unsigned l)
{
const struct grid_info grid = {
stage_w->bit | stage_r->bit,
get_image_format(GL_RGBA32UI),
{ l, l, 1, 1 }
};
const struct image_info img = image_info_for_grid(grid);
GLuint prog = generate_program(
grid,
/*
* Write (11, 22, 33, 44) to some location on the
* image from the write stage.
*/
stage_w->stage,
concat(qualifier_hunk(qual),
image_hunk(img, ""),
hunk("IMAGE_Q uniform IMAGE_T img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" imageStore(img, idx, DATA_T(11, 22, 33, 44));"
" return x;"
"}\n"), NULL),
/*
* The same location will read back the expected value
* if image access is coherent, as the shader inputs
* of the read stage are dependent on the outputs of
* the write stage and consequently they are
* guaranteed to be executed sequentially.
*/
stage_r->stage,
concat(qualifier_hunk(qual),
image_hunk(img, ""),
hunk("IMAGE_Q uniform IMAGE_T img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" DATA_T v = imageLoad(img, idx);"
" if (v == DATA_T(11, 22, 33, 44))"
" return GRID_T(33, 33, 33, 33);"
" else"
" return GRID_T(77, 77, 77, 77);"
"}\n"), NULL));
bool ret = prog &&
init_fb(grid) &&
init_image(img) &&
set_uniform_int(prog, "img", 0) &&
draw_grid(grid, prog) &&
(check(grid, img) || qual->control_test);
glDeleteProgram(prog);
return ret;
}
/**
* Test skeleton: Init image to \a init_value, run the provided shader
* \a op and check that the resulting image pixels equal \a
* check_value.
*/
static bool
run_test(uint32_t init_value, uint32_t check_value,
const char *op)
{
const struct grid_info grid =
grid_info(GL_FRAGMENT_SHADER, GL_R32UI, W, H);
const struct image_info img = image_info_for_grid(grid);
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(img, ""),
hunk("uniform IMAGE_T img;\n"),
hunk(op), NULL));
bool ret = prog &&
init_fb(grid) &&
init_image(img, init_value) &&
set_uniform_int(prog, "img", 0) &&
draw_grid(grid, prog) &&
check(img, check_value);
glDeleteProgram(prog);
return ret;
}
static bool
run_test_fragment(void)
{
const char *fs_source = "#version 140\n"
"#extension GL_ARB_shader_atomic_counters : enable\n"
"\n"
"out ivec4 fcolor;\n"
"uniform int index;\n"
"layout(binding = 0, offset = 4) uniform atomic_uint x[3];\n"
"\n"
"void main() {\n"
" fcolor.x = int(atomicCounterIncrement(x[1 + index]));\n"
" fcolor.y = int(atomicCounterIncrement(x[0 + index]));\n"
" fcolor.z = int(atomicCounterIncrement(x[1 + index]));\n"
" fcolor.w = int(atomicCounterIncrement(x[0 + index]));\n"
"}\n";
const char *vs_source = "#version 140\n"
"#extension GL_ARB_shader_atomic_counters : enable\n"
"\n"
"in vec4 position;\n"
"\n"
"void main() {\n"
" gl_Position = position;\n"
"}\n";
const uint32_t start[] = { 1, 2, 4, 8 };
const unsigned int expected[] = { 8, 4, 9, 5 };
GLuint prog = glCreateProgram();
bool ret =
atomic_counters_compile(prog, GL_FRAGMENT_SHADER, fs_source) &&
atomic_counters_compile(prog, GL_VERTEX_SHADER, vs_source) &&
set_uniform_int(prog, "index", 1) &&
atomic_counters_draw_point(prog, sizeof(start), start) &&
piglit_probe_rect_rgba_uint(0, 0, 1, 1, expected);
glDeleteProgram(prog);
return ret;
}
/**
* If \a layered is false, bind an individual layer of a texture to an
* image unit, read its contents and write back a different value to
* the same location. If \a layered is true or the texture has a
* single layer, the whole texture will be read and written back.
*
* For textures with a single layer, the arguments \a layered and \a
* layer which are passed to the same arguments of
* glBindImageTexture() should have no effect as required by the spec.
*/
static bool
run_test(const struct image_target_info *target,
bool layered, unsigned layer)
{
const struct image_info real_img = image_info(
target->target, GL_RGBA32F, W, H);
const unsigned slices = (layered ? 1 : image_num_layers(real_img));
/*
* "Slice" of the image that will be bound to the pipeline.
*/
const struct image_info slice_img = image_info(
(layered ? target->target : image_layer_target(target)),
GL_RGBA32F, W, H / slices);
/*
* Grid with as many elements as the slice.
*/
const struct grid_info grid = grid_info(
GL_FRAGMENT_SHADER, GL_RGBA32F, W, H / slices);
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(slice_img, ""),
hunk("IMAGE_UNIFORM_T img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" GRID_T v = imageLoad(img, IMAGE_ADDR(idx));\n"
" imageStore(img, IMAGE_ADDR(idx), DATA_T(33));\n"
" return v;\n"
"}\n"), NULL));
bool ret = prog && init_fb(grid) &&
init_image(real_img, layered, layer) &&
set_uniform_int(prog, "img", 0) &&
draw_grid(grid, prog) &&
check(grid, real_img, (slices == 1 ? 0 : layer));
glDeleteProgram(prog);
return ret;
}
bool
download_image_levels(const struct image_info img, unsigned num_levels,
unsigned unit, uint32_t *r_pixels)
{
const unsigned m = image_num_components(img.format);
int i, l;
glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT |
GL_BUFFER_UPDATE_BARRIER_BIT |
GL_PIXEL_BUFFER_BARRIER_BIT |
GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
glBindTexture(img.target->target, textures[unit]);
switch (img.target->target) {
case GL_TEXTURE_1D:
case GL_TEXTURE_2D:
case GL_TEXTURE_3D:
case GL_TEXTURE_RECTANGLE:
case GL_TEXTURE_1D_ARRAY:
case GL_TEXTURE_2D_ARRAY:
case GL_TEXTURE_CUBE_MAP_ARRAY:
assert(img.target->target != GL_TEXTURE_RECTANGLE ||
num_levels == 1);
for (l = 0; l < num_levels; ++l)
glGetTexImage(img.target->target, l,
img.format->pixel_format,
image_base_type(img.format),
&r_pixels[m * image_level_offset(img, l)]);
break;
case GL_TEXTURE_CUBE_MAP:
for (l = 0; l < num_levels; ++l) {
const unsigned offset = m * image_level_offset(img, l);
const unsigned face_sz =
m * product(image_level_size(img, l)) / 6;
for (i = 0; i < 6; ++i)
glGetTexImage(GL_TEXTURE_CUBE_MAP_POSITIVE_X + i,
l, img.format->pixel_format,
image_base_type(img.format),
&r_pixels[offset + face_sz * i]);
}
break;
case GL_TEXTURE_BUFFER: {
/*
* glGetTexImage() isn't supposed to work with buffer
* textures. We copy the packed pixels to a texture
* with the same internal format as the image to let
* the GL unpack it for us.
*/
const struct image_extent grid = image_optimal_extent(img.size);
GLuint packed_tex;
assert(num_levels == 1);
glGenTextures(1, &packed_tex);
glBindTexture(GL_TEXTURE_2D, packed_tex);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, buffers[unit]);
glTexImage2D(GL_TEXTURE_2D, 0, img.format->format,
grid.x, grid.y, 0, img.format->pixel_format,
img.format->pixel_type, NULL);
glGetTexImage(GL_TEXTURE_2D, 0, img.format->pixel_format,
image_base_type(img.format), r_pixels);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
glDeleteTextures(1, &packed_tex);
break;
}
case GL_TEXTURE_2D_MULTISAMPLE:
case GL_TEXTURE_2D_MULTISAMPLE_ARRAY: {
/*
* GL doesn't seem to provide any direct way to read
* back a multisample texture, so we use imageLoad()
* to copy its contents to a larger single-sample 2D
* texture from the fragment shader.
*/
const struct grid_info grid = {
get_image_stage(GL_FRAGMENT_SHADER)->bit,
img.format,
image_optimal_extent(img.size)
};
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(img, "SRC_"),
image_hunk(image_info_for_grid(grid), "DST_"),
hunk("readonly SRC_IMAGE_UNIFORM_T src_img;\n"
"writeonly DST_IMAGE_UNIFORM_T dst_img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" imageStore(dst_img, DST_IMAGE_ADDR(idx),\n"
" imageLoad(src_img, SRC_IMAGE_ADDR(idx)));\n"
" return x;\n"
"}\n"), NULL));
bool ret = prog && generate_fb(grid, 1);
GLuint tmp_tex;
assert(num_levels == 1);
glGenTextures(1, &tmp_tex);
glBindTexture(GL_TEXTURE_2D, tmp_tex);
glTexImage2D(GL_TEXTURE_2D, 0, img.format->format,
grid.size.x, grid.size.y, 0,
img.format->pixel_format, image_base_type(img.format),
NULL);
glBindImageTexture(unit, textures[unit], 0, GL_TRUE, 0,
GL_READ_ONLY, img.format->format);
glBindImageTexture(6, tmp_tex, 0, GL_TRUE, 0,
GL_WRITE_ONLY, img.format->format);
ret &= set_uniform_int(prog, "src_img", unit) &&
set_uniform_int(prog, "dst_img", 6) &&
draw_grid(grid, prog);
glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT);
glGetTexImage(GL_TEXTURE_2D, 0, img.format->pixel_format,
image_base_type(img.format), r_pixels);
glDeleteProgram(prog);
glDeleteTextures(1, &tmp_tex);
glBindFramebuffer(GL_FRAMEBUFFER, fb[0]);
glViewportIndexedfv(0, vp[0]);
if (!ret)
return false;
break;
}
default:
abort();
}
return piglit_check_gl_error(GL_NO_ERROR);
}
bool
upload_image_levels(const struct image_info img, unsigned num_levels,
unsigned level, unsigned unit, const uint32_t *pixels)
{
const unsigned m = image_num_components(img.format);
int i, l;
if (get_texture(unit)) {
glDeleteTextures(1, &textures[unit]);
textures[unit] = 0;
}
if (get_buffer(unit)) {
glDeleteBuffers(1, &buffers[unit]);
buffers[unit] = 0;
}
glGenTextures(1, &textures[unit]);
glBindTexture(img.target->target, textures[unit]);
switch (img.target->target) {
case GL_TEXTURE_1D:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage1D(GL_TEXTURE_1D, l, img.format->format,
size.x, 0, img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_2D:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage2D(GL_TEXTURE_2D, l, img.format->format,
size.x, size.y, 0,
img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_3D:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage3D(GL_TEXTURE_3D, l, img.format->format,
size.x, size.y, size.z, 0,
img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_RECTANGLE:
assert(num_levels == 1);
glTexImage2D(GL_TEXTURE_RECTANGLE, 0, img.format->format,
img.size.x, img.size.y, 0, img.format->pixel_format,
image_base_type(img.format), pixels);
break;
case GL_TEXTURE_CUBE_MAP:
for (l = 0; l < num_levels; ++l) {
const unsigned offset = m * image_level_offset(img, l);
const struct image_extent size = image_level_size(img, l);
const unsigned face_sz = m * product(size) / 6;
for (i = 0; i < 6; ++i)
glTexImage2D(GL_TEXTURE_CUBE_MAP_POSITIVE_X + i, l,
img.format->format, size.x, size.y, 0,
img.format->pixel_format,
image_base_type(img.format),
&pixels[offset + face_sz * i]);
}
break;
case GL_TEXTURE_BUFFER: {
/*
* glTexImage*() isn't supposed to work with buffer
* textures. We copy the unpacked pixels to a texture
* with the desired internal format to let the GL pack
* them for us.
*/
const struct image_extent grid = image_optimal_extent(img.size);
GLuint packed_tex;
assert(num_levels == 1);
glGenBuffers(1, &buffers[unit]);
glBindBuffer(GL_PIXEL_PACK_BUFFER, buffers[unit]);
glBufferData(GL_PIXEL_PACK_BUFFER,
img.size.x * image_pixel_size(img.format) / 8,
NULL, GL_STATIC_DRAW);
glGenTextures(1, &packed_tex);
glBindTexture(GL_TEXTURE_2D, packed_tex);
glTexImage2D(GL_TEXTURE_2D, 0, img.format->format,
grid.x, grid.y, 0, img.format->pixel_format,
image_base_type(img.format), pixels);
glGetTexImage(GL_TEXTURE_2D, 0, img.format->pixel_format,
img.format->pixel_type, NULL);
glDeleteTextures(1, &packed_tex);
glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
glTexBuffer(GL_TEXTURE_BUFFER, image_compat_format(img.format),
buffers[unit]);
break;
}
case GL_TEXTURE_1D_ARRAY:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage2D(GL_TEXTURE_1D_ARRAY, l, img.format->format,
size.x, size.y, 0, img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_2D_ARRAY:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage3D(GL_TEXTURE_2D_ARRAY, l, img.format->format,
size.x, size.y, size.z, 0,
img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_CUBE_MAP_ARRAY:
for (l = 0; l < num_levels; ++l) {
const struct image_extent size = image_level_size(img, l);
glTexImage3D(GL_TEXTURE_CUBE_MAP_ARRAY, l, img.format->format,
size.x, size.y, size.z, 0,
img.format->pixel_format,
image_base_type(img.format),
&pixels[m * image_level_offset(img, l)]);
}
break;
case GL_TEXTURE_2D_MULTISAMPLE:
case GL_TEXTURE_2D_MULTISAMPLE_ARRAY: {
/*
* GL doesn't seem to provide any direct way to
* initialize a multisample texture, so we use
* imageStore() to render to it from the fragment
* shader copying the contents of a larger
* single-sample 2D texture.
*/
const struct grid_info grid = {
get_image_stage(GL_FRAGMENT_SHADER)->bit,
img.format,
image_optimal_extent(img.size)
};
GLuint prog = generate_program(
grid, GL_FRAGMENT_SHADER,
concat(image_hunk(image_info_for_grid(grid), "SRC_"),
image_hunk(img, "DST_"),
hunk("readonly SRC_IMAGE_UNIFORM_T src_img;\n"
"writeonly DST_IMAGE_UNIFORM_T dst_img;\n"
"\n"
"GRID_T op(ivec2 idx, GRID_T x) {\n"
" imageStore(dst_img, DST_IMAGE_ADDR(idx),\n"
" imageLoad(src_img, SRC_IMAGE_ADDR(idx)));\n"
" return x;\n"
"}\n"), NULL));
bool ret = prog && generate_fb(grid, 1);
GLuint tmp_tex;
assert(num_levels == 1);
glGenTextures(1, &tmp_tex);
glBindTexture(GL_TEXTURE_2D, tmp_tex);
if (img.target->target == GL_TEXTURE_2D_MULTISAMPLE_ARRAY) {
glTexImage3DMultisample(GL_TEXTURE_2D_MULTISAMPLE_ARRAY,
img.size.x, img.format->format,
img.size.y, img.size.z, img.size.w,
GL_FALSE);
} else {
glTexImage2DMultisample(GL_TEXTURE_2D_MULTISAMPLE,
img.size.x, img.format->format,
img.size.y, img.size.z,
GL_FALSE);
}
glTexImage2D(GL_TEXTURE_2D, 0, img.format->format,
grid.size.x, grid.size.y, 0,
img.format->pixel_format, image_base_type(img.format),
pixels);
glBindImageTexture(unit, textures[unit], 0, GL_TRUE, 0,
GL_WRITE_ONLY, img.format->format);
glBindImageTexture(6, tmp_tex, 0, GL_TRUE, 0,
GL_READ_ONLY, img.format->format);
ret &= set_uniform_int(prog, "src_img", 6) &&
set_uniform_int(prog, "dst_img", unit) &&
draw_grid(grid, prog);
glDeleteProgram(prog);
glDeleteTextures(1, &tmp_tex);
glBindFramebuffer(GL_FRAMEBUFFER, fb[0]);
glViewportIndexedfv(0, vp[0]);
glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
if (!ret)
return false;
break;
}
default:
abort();
}
glBindImageTexture(unit, textures[unit], level, GL_TRUE, 0,
GL_READ_WRITE, img.format->format);
return piglit_check_gl_error(GL_NO_ERROR);
}
bool
draw_grid(const struct grid_info grid, GLuint prog)
{
static GLuint lprog;
if (lprog != prog) {
glUseProgram(prog);
lprog = prog;
}
if (grid.stages & GL_COMPUTE_SHADER_BIT) {
set_uniform_int(prog, "ret_img", 7);
glDispatchCompute(1, grid.size.y, 1);
} else if (grid.stages & (GL_TESS_CONTROL_SHADER_BIT |
GL_TESS_EVALUATION_SHADER_BIT)) {
static struct image_extent size;
static GLuint vao, vbo;
if (size.x != grid.size.x || size.y != grid.size.y) {
size = grid.size;
if (!generate_grid_arrays(
&vao, &vbo,
1.0 / size.x - 1.0, 1.0 / size.y - 1.0,
2.0 / size.x, 2.0 / size.y,
size.x, size.y))
return false;
}
glBindVertexArray(vao);
glPatchParameteri(GL_PATCH_VERTICES, 4);
glDrawArrays(GL_PATCHES, 0, product(size));
} else if (grid.stages & (GL_VERTEX_SHADER_BIT |
GL_GEOMETRY_SHADER_BIT)) {
static struct image_extent size;
static GLuint vao, vbo;
if (size.x != grid.size.x || size.y != grid.size.y) {
size = grid.size;
if (!generate_grid_arrays(
&vao, &vbo,
1.0 / size.x - 1.0, 1.0 / size.y - 1.0,
2.0 / size.x, 2.0 / size.y,
size.x, size.y))
return false;
}
glBindVertexArray(vao);
glDrawArrays(GL_POINTS, 0, product(size));
} else {
static struct image_extent size;
static GLuint vao, vbo;
if (size.x != grid.size.x || size.y != grid.size.y) {
float vp[4];
glGetFloati_v(GL_VIEWPORT, 0, vp);
size = grid.size;
if (!generate_grid_arrays(
&vao, &vbo, -1.0, -1.0,
2.0 * size.x / vp[2], 2.0 * size.y / vp[3],
2, 2))
return false;
}
glBindVertexArray(vao);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
return piglit_check_gl_error(GL_NO_ERROR);
}
void FFTPassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
{
Effect::set_gl_state(glsl_program_num, prefix, sampler_num);
int input_size = (direction == VERTICAL) ? input_height : input_width;
// See the comments on changes_output_size() in the .h file to see
// why this is legal. It is _needed_ because it counteracts the
// precision issues we get because we sample the input texture with
// normalized coordinates (especially when the repeat count along
// the axis is not a power of two); we very rapidly end up in narrowly
// missing a texel center, which causes precision loss to propagate
// throughout the FFT.
assert(*sampler_num == 1);
glActiveTexture(GL_TEXTURE0);
check_error();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
check_error();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
check_error();
// The memory layout follows figure 5.2 on page 25 of
// http://gpuwave.sesse.net/gpuwave.pdf -- it can be a bit confusing
// at first, but is classically explained more or less as follows:
//
// The classic Cooley-Tukey decimation-in-time FFT algorithm works
// by first splitting input data into odd and even elements
// (e.g. bit-wise xxxxx0 and xxxxx1 for a size-32 FFT), then FFTing
// them separately and combining them using twiddle factors.
// So the outer pass (done _last_) looks only at the last bit,
// and does one such merge pass of sub-size N/2 (FFT size N).
//
// FFT of the first part must then necessarily be split into xxxx00 and
// xxxx10, and similarly xxxx01 and xxxx11 for the other part. Since
// these two FFTs are handled identically, it means we split into xxxx0x
// and xxxx1x, so that the second-outer pass (done second-to-last)
// looks only at the second last bit, and so on. We do two such merge
// passes of sub-size N/4 (sub-FFT size N/2).
//
// Thus, the inner, Nth pass (done first) splits at the first bit,
// so 0 is paired with 16, 1 with 17 and so on, doing N/2 such merge
// passes of sub-size 1 (sub-FFT size 2). We say that the stride is 16.
// The second-inner, (N-1)th pass (done second) splits at the second
// bit, so the stride is 8, and so on.
assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
float *tmp = new float[fft_size * 4];
int subfft_size = 1 << pass_number;
double mulfac;
if (inverse) {
mulfac = 2.0 * M_PI;
} else {
mulfac = -2.0 * M_PI;
}
assert((fft_size & (fft_size - 1)) == 0); // Must be power of two.
assert(fft_size % subfft_size == 0);
int stride = fft_size / subfft_size;
for (int i = 0; i < fft_size; ++i) {
int k = i / stride; // Element number within this sub-FFT.
int offset = i % stride; // Sub-FFT number.
double twiddle_real, twiddle_imag;
if (k < subfft_size / 2) {
twiddle_real = cos(mulfac * (k / double(subfft_size)));
twiddle_imag = sin(mulfac * (k / double(subfft_size)));
} else {
// This is mathematically equivalent to the twiddle factor calculations
// in the other branch of the if, but not numerically; the range
// reductions on x87 are not all that precise, and this keeps us within
// [0,pi>.
k -= subfft_size / 2;
twiddle_real = -cos(mulfac * (k / double(subfft_size)));
twiddle_imag = -sin(mulfac * (k / double(subfft_size)));
}
// The support texture contains everything we need for the FFT:
// Obviously, the twiddle factor (in the Z and W components), but also
// which two samples to fetch. These are stored as normalized
// X coordinate offsets (Y coordinate for a vertical FFT); the reason
// for using offsets and not direct coordinates as in GPUwave
// is that we can have multiple FFTs along the same line,
// and want to reuse the support texture by repeating it.
int base = k * stride * 2 + offset;
int support_texture_index;
if (direction == FFTPassEffect::VERTICAL) {
// Compensate for OpenGL's bottom-left convention.
support_texture_index = fft_size - i - 1;
} else {
support_texture_index = i;
}
tmp[support_texture_index * 4 + 0] = (base - support_texture_index) / double(input_size);
tmp[support_texture_index * 4 + 1] = (base + stride - support_texture_index) / double(input_size);
tmp[support_texture_index * 4 + 2] = twiddle_real;
tmp[support_texture_index * 4 + 3] = twiddle_imag;
}
glActiveTexture(GL_TEXTURE0 + *sampler_num);
check_error();
glBindTexture(GL_TEXTURE_1D, tex);
check_error();
glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
check_error();
glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
check_error();
glTexParameteri(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_REPEAT);
check_error();
// Supposedly FFTs are very sensitive to inaccuracies in the twiddle factors,
// at least according to a paper by Schatzman (see gpuwave.pdf reference [30]
// for the full reference), so we keep them at 32-bit. However, for
// small sizes, all components are exact anyway, so we can cheat there
// (although noting that the source coordinates become somewhat less
// accurate then, too).
glTexImage1D(GL_TEXTURE_1D, 0, (subfft_size <= 4) ? GL_RGBA16F : GL_RGBA32F, fft_size, 0, GL_RGBA, GL_FLOAT, tmp);
check_error();
delete[] tmp;
set_uniform_int(glsl_program_num, prefix, "support_tex", *sampler_num);
++*sampler_num;
assert(input_size % fft_size == 0);
set_uniform_float(glsl_program_num, prefix, "num_repeats", input_size / fft_size);
}