/**
* Convert srgb int values to linear float values.
* Several possibilities how to do this, e.g.
* - table
* - doing the pow() with int-to-float and float-to-int tricks
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
* - just using standard polynomial approximation
* (3rd order polynomial is required for crappy but just sufficient accuracy)
*
* @param src integer (vector) value(s) to convert
* (chan_bits bit values unpacked to 32 bit already).
*/
LLVMValueRef
lp_build_srgb_to_linear(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned chan_bits,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, src_type.length * 32);
struct lp_build_context f32_bld;
LLVMValueRef srcf, part_lin, part_pow, is_linear, lin_const, lin_thresh;
double coeffs[4] = {0.0023f,
0.0030f / 255.0f,
0.6935f / (255.0f * 255.0f),
0.3012f / (255.0f * 255.0f * 255.0f)
};
assert(src_type.width == 32);
/* Technically this would work with more bits too but would be inaccurate. */
assert(chan_bits <= 8);
lp_build_context_init(&f32_bld, gallivm, f32_type);
/*
* using polynomial: (src * (src * (src * 0.3012 + 0.6935) + 0.0030) + 0.0023)
* ( poly = 0.3012*x^3 + 0.6935*x^2 + 0.0030*x + 0.0023)
* (found with octave polyfit and some magic as I couldn't get the error
* function right). Using the above mentioned error function, the values stay
* within +-0.35, except for the lowest values - hence tweaking linear segment
* to cover the first 16 instead of the first 11 values (the error stays
* just about acceptable there too).
* Hence: lin = src > 15 ? poly : src / 12.6
* This function really only makes sense for vectors, should use LUT otherwise.
* All in all (including float conversion) 11 instructions (with sse4.1),
* 6 constants (polynomial could be done with 1 instruction less at the cost
* of slightly worse dependency chain, fma should also help).
*/
/* doing the 1/255 mul as part of the approximation */
srcf = lp_build_int_to_float(&f32_bld, src);
if (chan_bits != 8) {
/* could adjust all the constants instead */
LLVMValueRef rescale_const = lp_build_const_vec(gallivm, f32_type,
255.0f / ((1 << chan_bits) - 1));
srcf = lp_build_mul(&f32_bld, srcf, rescale_const);
}
lin_const = lp_build_const_vec(gallivm, f32_type, 1.0f / (12.6f * 255.0f));
part_lin = lp_build_mul(&f32_bld, srcf, lin_const);
part_pow = lp_build_polynomial(&f32_bld, srcf, coeffs, 4);
lin_thresh = lp_build_const_vec(gallivm, f32_type, 15.0f);
is_linear = lp_build_compare(gallivm, f32_type, PIPE_FUNC_LEQUAL, srcf, lin_thresh);
return lp_build_select(&f32_bld, is_linear, part_lin, part_pow);
}
/**
* Generate code to compute coordinate gradient (rho).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
* \param ddy partial derivatives of (s, t, r, q) with respect to Y
*
* XXX: The resulting rho is scalar, so we ignore all but the first element of
* derivatives that are passed by the shader.
*/
static LLVMValueRef
lp_build_rho(struct lp_build_sample_context *bld,
unsigned unit,
const LLVMValueRef ddx[4],
const LLVMValueRef ddy[4])
{
struct lp_build_context *int_size_bld = &bld->int_size_bld;
struct lp_build_context *float_size_bld = &bld->float_size_bld;
struct lp_build_context *float_bld = &bld->float_bld;
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
LLVMValueRef index2 = LLVMConstInt(i32t, 2, 0);
LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
LLVMValueRef rho_x, rho_y;
LLVMValueRef rho_vec;
LLVMValueRef int_size, float_size;
LLVMValueRef rho;
LLVMValueRef first_level, first_level_vec;
dsdx = ddx[0];
dsdy = ddy[0];
if (dims <= 1) {
rho_x = dsdx;
rho_y = dsdy;
}
else {
rho_x = float_size_bld->undef;
rho_y = float_size_bld->undef;
rho_x = LLVMBuildInsertElement(builder, rho_x, dsdx, index0, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dsdy, index0, "");
dtdx = ddx[1];
dtdy = ddy[1];
rho_x = LLVMBuildInsertElement(builder, rho_x, dtdx, index1, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dtdy, index1, "");
if (dims >= 3) {
drdx = ddx[2];
drdy = ddy[2];
rho_x = LLVMBuildInsertElement(builder, rho_x, drdx, index2, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, drdy, index2, "");
}
}
rho_x = lp_build_abs(float_size_bld, rho_x);
rho_y = lp_build_abs(float_size_bld, rho_y);
rho_vec = lp_build_max(float_size_bld, rho_x, rho_y);
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
first_level_vec = lp_build_broadcast_scalar(&bld->int_size_bld, first_level);
int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec);
float_size = lp_build_int_to_float(float_size_bld, int_size);
rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
if (dims <= 1) {
rho = rho_vec;
}
else {
if (dims >= 2) {
LLVMValueRef rho_s, rho_t, rho_r;
rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
rho = lp_build_max(float_bld, rho_s, rho_t);
if (dims >= 3) {
rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
rho = lp_build_max(float_bld, rho, rho_r);
}
}
}
return rho;
}
/**
* Sample a single texture image with (bi-)(tri-)linear sampling.
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_linear(struct lp_build_sample_context *bld,
LLVMValueRef int_size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *colors_lo,
LLVMValueRef *colors_hi)
{
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context i32, h16, u8n;
LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
LLVMValueRef i32_c8, i32_c128, i32_c255;
LLVMValueRef width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
LLVMValueRef t_ipart = NULL, t_fpart = NULL, t_fpart_lo = NULL, t_fpart_hi = NULL;
LLVMValueRef r_ipart = NULL, r_fpart = NULL, r_fpart_lo = NULL, r_fpart_hi = NULL;
LLVMValueRef x_stride, y_stride, z_stride;
LLVMValueRef x_offset0, x_offset1;
LLVMValueRef y_offset0, y_offset1;
LLVMValueRef z_offset0, z_offset1;
LLVMValueRef offset[2][2][2]; /* [z][y][x] */
LLVMValueRef x_subcoord[2], y_subcoord[2], z_subcoord[2];
LLVMValueRef neighbors_lo[2][2][2]; /* [z][y][x] */
LLVMValueRef neighbors_hi[2][2][2]; /* [z][y][x] */
LLVMValueRef packed_lo, packed_hi;
unsigned x, y, z;
unsigned i, j, k;
unsigned numj, numk;
lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32));
lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16));
lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8));
i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type);
h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type);
u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type);
lp_build_extract_image_sizes(bld,
bld->int_size_type,
bld->int_coord_type,
int_size,
&width_vec,
&height_vec,
&depth_vec);
if (bld->static_state->normalized_coords) {
LLVMValueRef scaled_size;
LLVMValueRef flt_size;
/* scale size by 256 (8 fractional bits) */
scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8);
flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size);
lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r);
}
else {
/* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
if (dims >= 2)
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
if (dims >= 3)
r = lp_build_mul_imm(&bld->coord_bld, r, 256);
}
/* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
if (dims >= 2)
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
if (dims >= 3)
r = LLVMBuildFPToSI(builder, r, i32_vec_type, "");
/* subtract 0.5 (add -128) */
i32_c128 = lp_build_const_int_vec(bld->gallivm, i32.type, -128);
s = LLVMBuildAdd(builder, s, i32_c128, "");
if (dims >= 2) {
t = LLVMBuildAdd(builder, t, i32_c128, "");
}
if (dims >= 3) {
r = LLVMBuildAdd(builder, r, i32_c128, "");
}
/* compute floor (shift right 8) */
i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
if (dims >= 2)
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
if (dims >= 3)
r_ipart = LLVMBuildAShr(builder, r, i32_c8, "");
/* compute fractional part (AND with 0xff) */
i32_c255 = lp_build_const_int_vec(bld->gallivm, i32.type, 255);
s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
if (dims >= 2)
t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
if (dims >= 3)
r_fpart = LLVMBuildAnd(builder, r, i32_c255, "");
/* get pixel, row and image strides */
x_stride = lp_build_const_vec(bld->gallivm, bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
y_stride = row_stride_vec;
z_stride = img_stride_vec;
/* do texcoord wrapping and compute texel offsets */
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset0, &x_offset1,
&x_subcoord[0], &x_subcoord[1]);
for (z = 0; z < 2; z++) {
for (y = 0; y < 2; y++) {
offset[z][y][0] = x_offset0;
offset[z][y][1] = x_offset1;
}
}
if (dims >= 2) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, y_stride,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset0, &y_offset1,
&y_subcoord[0], &y_subcoord[1]);
for (z = 0; z < 2; z++) {
for (x = 0; x < 2; x++) {
offset[z][0][x] = lp_build_add(&bld->int_coord_bld,
offset[z][0][x], y_offset0);
offset[z][1][x] = lp_build_add(&bld->int_coord_bld,
offset[z][1][x], y_offset1);
}
}
}
if (dims >= 3) {
lp_build_sample_wrap_linear_int(bld,
bld->format_desc->block.height,
r_ipart, depth_vec, z_stride,
bld->static_state->pot_depth,
bld->static_state->wrap_r,
&z_offset0, &z_offset1,
&z_subcoord[0], &z_subcoord[1]);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset0);
offset[1][y][x] = lp_build_add(&bld->int_coord_bld,
offset[1][y][x], z_offset1);
}
}
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
for (y = 0; y < 2; y++) {
for (x = 0; x < 2; x++) {
/* The r coord is the cube face in [0,5] */
offset[0][y][x] = lp_build_add(&bld->int_coord_bld,
offset[0][y][x], z_offset);
}
}
}
/*
* Transform 4 x i32 in
*
* s_fpart = {s0, s1, s2, s3}
*
* into 8 x i16
*
* s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
*
* into two 8 x i16
*
* s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
* s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
*
* and likewise for t_fpart. There is no risk of loosing precision here
* since the fractional parts only use the lower 8bits.
*/
s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
if (dims >= 2)
t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
if (dims >= 3)
r_fpart = LLVMBuildBitCast(builder, r_fpart, h16_vec_type, "");
{
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
LLVMValueRef shuffle_lo;
LLVMValueRef shuffle_hi;
for (j = 0; j < h16.type.length; j += 4) {
#ifdef PIPE_ARCH_LITTLE_ENDIAN
unsigned subindex = 0;
#else
unsigned subindex = 1;
#endif
LLVMValueRef index;
index = LLVMConstInt(elem_type, j/2 + subindex, 0);
for (i = 0; i < 4; ++i)
shuffles_lo[j + i] = index;
index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
for (i = 0; i < 4; ++i)
shuffles_hi[j + i] = index;
}
shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef,
shuffle_lo, "");
s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef,
shuffle_hi, "");
if (dims >= 2) {
t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_lo, "");
t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef,
shuffle_hi, "");
}
if (dims >= 3) {
r_fpart_lo = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_lo, "");
r_fpart_hi = LLVMBuildShuffleVector(builder, r_fpart, h16.undef,
shuffle_hi, "");
}
}
/*
* Fetch the pixels as 4 x 32bit (rgba order might differ):
*
* rgba0 rgba1 rgba2 rgba3
*
* bit cast them into 16 x u8
*
* r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
*
* unpack them into two 8 x i16:
*
* r0 g0 b0 a0 r1 g1 b1 a1
* r2 g2 b2 a2 r3 g3 b3 a3
*
* The higher 8 bits of the resulting elements will be zero.
*/
numj = 1 + (dims >= 2);
numk = 1 + (dims >= 3);
for (k = 0; k < numk; k++) {
for (j = 0; j < numj; j++) {
for (i = 0; i < 2; i++) {
LLVMValueRef rgba8;
if (util_format_is_rgba8_variant(bld->format_desc)) {
/*
* Given the format is a rgba8, just read the pixels as is,
* without any swizzling. Swizzling will be done later.
*/
rgba8 = lp_build_gather(bld->gallivm,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset[k][j][i]);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset[k][j][i],
x_subcoord[i],
y_subcoord[j]);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
&neighbors_lo[k][j][i], &neighbors_hi[k][j][i]);
}
}
}
/*
* Linear interpolation with 8.8 fixed point.
*/
if (dims == 1) {
/* 1-D lerp */
packed_lo = lp_build_lerp(&h16,
s_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1]);
packed_hi = lp_build_lerp(&h16,
s_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1]);
}
else {
/* 2-D lerp */
packed_lo = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[0][0][0],
neighbors_lo[0][0][1],
neighbors_lo[0][1][0],
neighbors_lo[0][1][1]);
packed_hi = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[0][0][0],
neighbors_hi[0][0][1],
neighbors_hi[0][1][0],
neighbors_hi[0][1][1]);
if (dims >= 3) {
LLVMValueRef packed_lo2, packed_hi2;
/* lerp in the second z slice */
packed_lo2 = lp_build_lerp_2d(&h16,
s_fpart_lo, t_fpart_lo,
neighbors_lo[1][0][0],
neighbors_lo[1][0][1],
neighbors_lo[1][1][0],
neighbors_lo[1][1][1]);
packed_hi2 = lp_build_lerp_2d(&h16,
s_fpart_hi, t_fpart_hi,
neighbors_hi[1][0][0],
neighbors_hi[1][0][1],
neighbors_hi[1][1][0],
neighbors_hi[1][1][1]);
/* interp between two z slices */
packed_lo = lp_build_lerp(&h16, r_fpart_lo,
packed_lo, packed_lo2);
packed_hi = lp_build_lerp(&h16, r_fpart_hi,
packed_hi, packed_hi2);
}
}
*colors_lo = packed_lo;
*colors_hi = packed_hi;
}
/**
* Sample a single texture image with nearest sampling.
* If sampling a cube texture, r = cube face in [0,5].
* Return filtered color as two vectors of 16-bit fixed point values.
*/
static void
lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
LLVMValueRef int_size,
LLVMValueRef row_stride_vec,
LLVMValueRef img_stride_vec,
LLVMValueRef data_ptr,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef *colors_lo,
LLVMValueRef *colors_hi)
{
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
struct lp_build_context i32, h16, u8n;
LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
LLVMValueRef i32_c8;
LLVMValueRef width_vec, height_vec, depth_vec;
LLVMValueRef s_ipart, t_ipart = NULL, r_ipart = NULL;
LLVMValueRef x_stride;
LLVMValueRef x_offset, offset;
LLVMValueRef x_subcoord, y_subcoord, z_subcoord;
lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32));
lp_build_context_init(&h16, bld->gallivm, lp_type_ufixed(16));
lp_build_context_init(&u8n, bld->gallivm, lp_type_unorm(8));
i32_vec_type = lp_build_vec_type(bld->gallivm, i32.type);
h16_vec_type = lp_build_vec_type(bld->gallivm, h16.type);
u8n_vec_type = lp_build_vec_type(bld->gallivm, u8n.type);
lp_build_extract_image_sizes(bld,
bld->int_size_type,
bld->int_coord_type,
int_size,
&width_vec,
&height_vec,
&depth_vec);
if (bld->static_state->normalized_coords) {
LLVMValueRef scaled_size;
LLVMValueRef flt_size;
/* scale size by 256 (8 fractional bits) */
scaled_size = lp_build_shl_imm(&bld->int_size_bld, int_size, 8);
flt_size = lp_build_int_to_float(&bld->float_size_bld, scaled_size);
lp_build_unnormalized_coords(bld, flt_size, &s, &t, &r);
}
else {
/* scale coords by 256 (8 fractional bits) */
s = lp_build_mul_imm(&bld->coord_bld, s, 256);
if (dims >= 2)
t = lp_build_mul_imm(&bld->coord_bld, t, 256);
if (dims >= 3)
r = lp_build_mul_imm(&bld->coord_bld, r, 256);
}
/* convert float to int */
s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
if (dims >= 2)
t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
if (dims >= 3)
r = LLVMBuildFPToSI(builder, r, i32_vec_type, "");
/* compute floor (shift right 8) */
i32_c8 = lp_build_const_int_vec(bld->gallivm, i32.type, 8);
s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
if (dims >= 2)
t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
if (dims >= 3)
r_ipart = LLVMBuildAShr(builder, r, i32_c8, "");
/* get pixel, row, image strides */
x_stride = lp_build_const_vec(bld->gallivm,
bld->int_coord_bld.type,
bld->format_desc->block.bits/8);
/* Do texcoord wrapping, compute texel offset */
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.width,
s_ipart, width_vec, x_stride,
bld->static_state->pot_width,
bld->static_state->wrap_s,
&x_offset, &x_subcoord);
offset = x_offset;
if (dims >= 2) {
LLVMValueRef y_offset;
lp_build_sample_wrap_nearest_int(bld,
bld->format_desc->block.height,
t_ipart, height_vec, row_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_t,
&y_offset, &y_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, y_offset);
if (dims >= 3) {
LLVMValueRef z_offset;
lp_build_sample_wrap_nearest_int(bld,
1, /* block length (depth) */
r_ipart, depth_vec, img_stride_vec,
bld->static_state->pot_height,
bld->static_state->wrap_r,
&z_offset, &z_subcoord);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
LLVMValueRef z_offset;
/* The r coord is the cube face in [0,5] */
z_offset = lp_build_mul(&bld->int_coord_bld, r, img_stride_vec);
offset = lp_build_add(&bld->int_coord_bld, offset, z_offset);
}
}
/*
* Fetch the pixels as 4 x 32bit (rgba order might differ):
*
* rgba0 rgba1 rgba2 rgba3
*
* bit cast them into 16 x u8
*
* r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
*
* unpack them into two 8 x i16:
*
* r0 g0 b0 a0 r1 g1 b1 a1
* r2 g2 b2 a2 r3 g3 b3 a3
*
* The higher 8 bits of the resulting elements will be zero.
*/
{
LLVMValueRef rgba8;
if (util_format_is_rgba8_variant(bld->format_desc)) {
/*
* Given the format is a rgba8, just read the pixels as is,
* without any swizzling. Swizzling will be done later.
*/
rgba8 = lp_build_gather(bld->gallivm,
bld->texel_type.length,
bld->format_desc->block.bits,
bld->texel_type.width,
data_ptr, offset);
rgba8 = LLVMBuildBitCast(builder, rgba8, u8n_vec_type, "");
}
else {
rgba8 = lp_build_fetch_rgba_aos(bld->gallivm,
bld->format_desc,
u8n.type,
data_ptr, offset,
x_subcoord,
y_subcoord);
}
/* Expand one 4*rgba8 to two 2*rgba16 */
lp_build_unpack2(bld->gallivm, u8n.type, h16.type,
rgba8,
colors_lo, colors_hi);
}
}
/**
* Convert linear float values to srgb int values.
* Several possibilities how to do this, e.g.
* - use table (based on exponent/highest order mantissa bits) and do
* linear interpolation (https://gist.github.com/rygorous/2203834)
* - Chebyshev polynomial
* - Approximation using reciprocals
* - using int-to-float and float-to-int tricks for pow()
* (http://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent)
*
* @param src float (vector) value(s) to convert.
*/
static LLVMValueRef
lp_build_linear_to_srgb(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context f32_bld;
LLVMValueRef lin_thresh, lin, lin_const, is_linear, tmp, pow_final;
lp_build_context_init(&f32_bld, gallivm, src_type);
src = lp_build_clamp(&f32_bld, src, f32_bld.zero, f32_bld.one);
if (0) {
/*
* using int-to-float and float-to-int trick for pow().
* This is much more accurate than necessary thanks to the correction,
* but it most certainly makes no sense without rsqrt available.
* Bonus points if you understand how this works...
* All in all (including min/max clamp, conversion) 19 instructions.
*/
float exp_f = 2.0f / 3.0f;
/* some compilers can't do exp2f, so this is exp2f(127.0f/exp_f - 127.0f) */
float exp2f_c = 1.30438178253e+19f;
float coeff_f = 0.62996f;
LLVMValueRef pow_approx, coeff, x2, exponent, pow_1, pow_2;
struct lp_type int_type = lp_int_type(src_type);
/*
* First calculate approx x^8/12
*/
exponent = lp_build_const_vec(gallivm, src_type, exp_f);
coeff = lp_build_const_vec(gallivm, src_type,
exp2f_c * powf(coeff_f, 1.0f / exp_f));
/* premultiply src */
tmp = lp_build_mul(&f32_bld, coeff, src);
/* "log2" */
tmp = LLVMBuildBitCast(builder, tmp, lp_build_vec_type(gallivm, int_type), "");
tmp = lp_build_int_to_float(&f32_bld, tmp);
/* multiply for pow */
tmp = lp_build_mul(&f32_bld, tmp, exponent);
/* "exp2" */
pow_approx = lp_build_itrunc(&f32_bld, tmp);
pow_approx = LLVMBuildBitCast(builder, pow_approx,
lp_build_vec_type(gallivm, src_type), "");
/*
* Since that pow was inaccurate (like 3 bits, though each sqrt step would
* give another bit), compensate the error (which is why we chose another
* exponent in the first place).
*/
/* x * x^(8/12) = x^(20/12) */
pow_1 = lp_build_mul(&f32_bld, pow_approx, src);
/* x * x * x^(-4/12) = x^(20/12) */
/* Should avoid using rsqrt if it's not available, but
* using x * x^(4/12) * x^(4/12) instead will change error weight */
tmp = lp_build_fast_rsqrt(&f32_bld, pow_approx);
x2 = lp_build_mul(&f32_bld, src, src);
pow_2 = lp_build_mul(&f32_bld, x2, tmp);
/* average the values so the errors cancel out, compensate bias,
* we also squeeze the 1.055 mul of the srgb conversion plus the 255.0 mul
* for conversion to int in here */
tmp = lp_build_add(&f32_bld, pow_1, pow_2);
coeff = lp_build_const_vec(gallivm, src_type,
1.0f / (3.0f * coeff_f) * 0.999852f *
powf(1.055f * 255.0f, 4.0f));
pow_final = lp_build_mul(&f32_bld, tmp, coeff);
/* x^(5/12) = rsqrt(rsqrt(x^20/12)) */
if (lp_build_fast_rsqrt_available(src_type)) {
pow_final = lp_build_fast_rsqrt(&f32_bld,
lp_build_fast_rsqrt(&f32_bld, pow_final));
}
else {
pow_final = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, pow_final));
}
pow_final = lp_build_add(&f32_bld, pow_final,
lp_build_const_vec(gallivm, src_type, -0.055f * 255.0f));
}
else {
/*
* using "rational polynomial" approximation here.
* Essentially y = a*x^0.375 + b*x^0.5 + c, with also
* factoring in the 255.0 mul and the scaling mul.
* (a is closer to actual value so has higher weight than b.)
* Note: the constants are magic values. They were found empirically,
* possibly could be improved but good enough (be VERY careful with
* error metric if you'd want to tweak them, they also MUST fit with
* the crappy polynomial above for srgb->linear since it is required
* that each srgb value maps back to the same value).
* This function has an error of max +-0.17 (and we'd only require +-0.6),
* for the approximated srgb->linear values the error is naturally larger
* (+-0.42) but still accurate enough (required +-0.5 essentially).
* All in all (including min/max clamp, conversion) 15 instructions.
* FMA would help (minus 2 instructions).
*/
LLVMValueRef x05, x0375, a_const, b_const, c_const, tmp2;
if (lp_build_fast_rsqrt_available(src_type)) {
tmp = lp_build_fast_rsqrt(&f32_bld, src);
x05 = lp_build_mul(&f32_bld, src, tmp);
}
else {
/*
* I don't really expect this to be practical without rsqrt
* but there's no reason for triple punishment so at least
* save the otherwise resulting division and unnecessary mul...
*/
x05 = lp_build_sqrt(&f32_bld, src);
}
tmp = lp_build_mul(&f32_bld, x05, src);
if (lp_build_fast_rsqrt_available(src_type)) {
x0375 = lp_build_fast_rsqrt(&f32_bld, lp_build_fast_rsqrt(&f32_bld, tmp));
}
else {
x0375 = lp_build_sqrt(&f32_bld, lp_build_sqrt(&f32_bld, tmp));
}
a_const = lp_build_const_vec(gallivm, src_type, 0.675f * 1.0622 * 255.0f);
b_const = lp_build_const_vec(gallivm, src_type, 0.325f * 1.0622 * 255.0f);
c_const = lp_build_const_vec(gallivm, src_type, -0.0620f * 255.0f);
tmp = lp_build_mul(&f32_bld, a_const, x0375);
tmp2 = lp_build_mul(&f32_bld, b_const, x05);
tmp2 = lp_build_add(&f32_bld, tmp2, c_const);
pow_final = lp_build_add(&f32_bld, tmp, tmp2);
}
/* linear part is easy */
lin_const = lp_build_const_vec(gallivm, src_type, 12.92f * 255.0f);
lin = lp_build_mul(&f32_bld, src, lin_const);
lin_thresh = lp_build_const_vec(gallivm, src_type, 0.0031308f);
is_linear = lp_build_compare(gallivm, src_type, PIPE_FUNC_LEQUAL, src, lin_thresh);
tmp = lp_build_select(&f32_bld, is_linear, lin, pow_final);
f32_bld.type.sign = 0;
return lp_build_iround(&f32_bld, tmp);
}
