static INLINE void
xrender_tex(struct ureg_program *ureg,
struct ureg_dst dst,
struct ureg_src coords,
struct ureg_src sampler,
struct ureg_src imm0,
boolean repeat_none, boolean swizzle, boolean set_alpha)
{
if (repeat_none) {
struct ureg_dst tmp0 = ureg_DECL_temporary(ureg);
struct ureg_dst tmp1 = ureg_DECL_temporary(ureg);
ureg_SGT(ureg, tmp1, ureg_swizzle(coords,
TGSI_SWIZZLE_X,
TGSI_SWIZZLE_Y,
TGSI_SWIZZLE_X,
TGSI_SWIZZLE_Y), ureg_scalar(imm0,
TGSI_SWIZZLE_X));
ureg_SLT(ureg, tmp0,
ureg_swizzle(coords, TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y,
TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y), ureg_scalar(imm0,
TGSI_SWIZZLE_W));
ureg_MIN(ureg, tmp0, ureg_src(tmp0), ureg_src(tmp1));
ureg_MIN(ureg, tmp0, ureg_scalar(ureg_src(tmp0), TGSI_SWIZZLE_X),
ureg_scalar(ureg_src(tmp0), TGSI_SWIZZLE_Y));
ureg_TEX(ureg, tmp1, TGSI_TEXTURE_2D, coords, sampler);
if (swizzle)
ureg_MOV(ureg, tmp1, ureg_swizzle(ureg_src(tmp1),
TGSI_SWIZZLE_Z,
TGSI_SWIZZLE_Y, TGSI_SWIZZLE_X,
TGSI_SWIZZLE_W));
if (set_alpha)
ureg_MOV(ureg,
ureg_writemask(tmp1, TGSI_WRITEMASK_W),
ureg_scalar(imm0, TGSI_SWIZZLE_W));
ureg_MUL(ureg, dst, ureg_src(tmp1), ureg_src(tmp0));
ureg_release_temporary(ureg, tmp0);
ureg_release_temporary(ureg, tmp1);
} else {
if (swizzle) {
struct ureg_dst tmp = ureg_DECL_temporary(ureg);
ureg_TEX(ureg, tmp, TGSI_TEXTURE_2D, coords, sampler);
ureg_MOV(ureg, dst, ureg_swizzle(ureg_src(tmp),
TGSI_SWIZZLE_Z,
TGSI_SWIZZLE_Y, TGSI_SWIZZLE_X,
TGSI_SWIZZLE_W));
ureg_release_temporary(ureg, tmp);
} else {
ureg_TEX(ureg, dst, TGSI_TEXTURE_2D, coords, sampler);
}
if (set_alpha)
ureg_MOV(ureg,
ureg_writemask(dst, TGSI_WRITEMASK_W),
ureg_scalar(imm0, TGSI_SWIZZLE_W));
}
}
static void *
create_frag_shader(struct vl_median_filter *filter,
struct vertex2f *offsets,
unsigned num_offsets)
{
struct pipe_screen *screen = filter->pipe->screen;
struct ureg_program *shader;
struct ureg_src i_vtex;
struct ureg_src sampler;
struct ureg_dst *t_array = MALLOC(sizeof(struct ureg_dst) * num_offsets);
struct ureg_dst o_fragment;
const unsigned median = num_offsets >> 1;
unsigned i, j;
assert(num_offsets & 1); /* we need an odd number of offsets */
if (!(num_offsets & 1)) { /* yeah, we REALLY need an odd number of offsets!!! */
FREE(t_array);
return NULL;
}
if (num_offsets > screen->get_shader_param(
screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_MAX_TEMPS)) {
FREE(t_array);
return NULL;
}
shader = ureg_create(PIPE_SHADER_FRAGMENT);
if (!shader) {
FREE(t_array);
return NULL;
}
i_vtex = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
sampler = ureg_DECL_sampler(shader, 0);
ureg_DECL_sampler_view(shader, 0, TGSI_TEXTURE_2D,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT);
for (i = 0; i < num_offsets; ++i)
t_array[i] = ureg_DECL_temporary(shader);
o_fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
/*
* t_array[0..*] = vtex + offset[0..*]
* t_array[0..*] = tex(t_array[0..*], sampler)
* result = partial_bubblesort(t_array)[mid]
*/
for (i = 0; i < num_offsets; ++i) {
if (!is_vec_zero(offsets[i])) {
ureg_ADD(shader, ureg_writemask(t_array[i], TGSI_WRITEMASK_XY),
i_vtex, ureg_imm2f(shader, offsets[i].x, offsets[i].y));
ureg_MOV(shader, ureg_writemask(t_array[i], TGSI_WRITEMASK_ZW),
ureg_imm1f(shader, 0.0f));
}
}
for (i = 0; i < num_offsets; ++i) {
struct ureg_src src = is_vec_zero(offsets[i]) ? i_vtex : ureg_src(t_array[i]);
ureg_TEX(shader, t_array[i], TGSI_TEXTURE_2D, src, sampler);
}
// TODO: Couldn't this be improved even more?
for (i = 0; i <= median; ++i) {
for (j = 1; j < (num_offsets - i - 1); ++j) {
struct ureg_dst tmp = ureg_DECL_temporary(shader);
ureg_MOV(shader, tmp, ureg_src(t_array[j]));
ureg_MAX(shader, t_array[j], ureg_src(t_array[j]), ureg_src(t_array[j - 1]));
ureg_MIN(shader, t_array[j - 1], ureg_src(tmp), ureg_src(t_array[j - 1]));
ureg_release_temporary(shader, tmp);
}
if (i == median)
ureg_MAX(shader, t_array[j], ureg_src(t_array[j]), ureg_src(t_array[j - 1]));
else
ureg_MIN(shader, t_array[j - 1], ureg_src(t_array[j]), ureg_src(t_array[j - 1]));
}
ureg_MOV(shader, o_fragment, ureg_src(t_array[median]));
ureg_END(shader);
FREE(t_array);
return ureg_create_shader_and_destroy(shader, filter->pipe);
}