void
lp_build_alpha_test(struct gallivm_state *gallivm,
unsigned func,
struct lp_type type,
const struct util_format_description *cbuf_format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef alpha,
LLVMValueRef ref,
boolean do_branch)
{
struct lp_build_context bld;
LLVMValueRef test;
lp_build_context_init(&bld, gallivm, type);
/*
* Alpha testing needs to be done in the color buffer precision.
*
* TODO: Ideally, instead of duplicating the color conversion code, we would do
* alpha testing after converting the output colors, but that's not very
* convenient, because it needs to be done before depth testing. Hopefully
* LLVM will detect and remove the duplicate expression.
*
* FIXME: This should be generalized to formats other than rgba8 variants.
*/
if (type.floating &&
util_format_is_rgba8_variant(cbuf_format_desc)) {
const unsigned dst_width = 8;
alpha = lp_build_clamp(&bld, alpha, bld.zero, bld.one);
ref = lp_build_clamp(&bld, ref, bld.zero, bld.one);
alpha = lp_build_clamped_float_to_unsigned_norm(gallivm, type, dst_width, alpha);
ref = lp_build_clamped_float_to_unsigned_norm(gallivm, type, dst_width, ref);
type.floating = 0;
lp_build_context_init(&bld, gallivm, type);
}
test = lp_build_cmp(&bld, func, alpha, ref);
lp_build_name(test, "alpha_mask");
lp_build_mask_update(mask, test);
if (do_branch)
lp_build_mask_check(mask);
}
/**
* For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
* mipmap level index.
* Note: this is all scalar code.
* \param lod scalar float texture level of detail
* \param level_out returns integer
*/
void
lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
unsigned unit,
LLVMValueRef lod_ipart,
LLVMValueRef *level_out)
{
struct lp_build_context *int_bld = &bld->int_bld;
LLVMValueRef first_level, last_level, level;
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
last_level = bld->dynamic_state->last_level(bld->dynamic_state,
bld->gallivm, unit);
/* convert float lod to integer */
level = lp_build_add(int_bld, lod_ipart, first_level);
/* clamp level to legal range of levels */
*level_out = lp_build_clamp(int_bld, level, first_level, last_level);
}
/**
* Convert linear float soa values to packed srgb AoS values.
* This only handles packed formats which are 4x8bit in size
* (rgba and rgbx plus swizzles).
*
* @param src float SoA (vector) values to convert.
*/
LLVMValueRef
lp_build_float_to_srgb_packed(struct gallivm_state *gallivm,
const struct util_format_description *dst_fmt,
struct lp_type src_type,
LLVMValueRef *src)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned chan;
struct lp_build_context f32_bld;
struct lp_type int32_type = lp_int_type(src_type);
LLVMValueRef tmpsrgb[4], alpha, dst;
lp_build_context_init(&f32_bld, gallivm, src_type);
/* rgb is subject to linear->srgb conversion, alpha is not */
for (chan = 0; chan < 3; chan++) {
tmpsrgb[chan] = lp_build_linear_to_srgb(gallivm, src_type, src[chan]);
}
/*
* can't use lp_build_conv since we want to keep values as 32bit
* here so we can interleave with rgb to go from SoA->AoS.
*/
alpha = lp_build_clamp(&f32_bld, src[3], f32_bld.zero, f32_bld.one);
alpha = lp_build_mul(&f32_bld, alpha,
lp_build_const_vec(gallivm, src_type, 255.0f));
tmpsrgb[3] = lp_build_iround(&f32_bld, alpha);
dst = lp_build_zero(gallivm, int32_type);
for (chan = 0; chan < dst_fmt->nr_channels; chan++) {
if (dst_fmt->swizzle[chan] <= UTIL_FORMAT_SWIZZLE_W) {
unsigned ls;
LLVMValueRef shifted, shift_val;
ls = dst_fmt->channel[dst_fmt->swizzle[chan]].shift;
shift_val = lp_build_const_int_vec(gallivm, int32_type, ls);
shifted = LLVMBuildShl(builder, tmpsrgb[chan], shift_val, "");
dst = LLVMBuildOr(builder, dst, shifted, "");
}
}
return dst;
}
static INLINE void
yuv_to_rgb_soa(struct gallivm_state *gallivm,
unsigned n,
LLVMValueRef y, LLVMValueRef u, LLVMValueRef v,
LLVMValueRef *r, LLVMValueRef *g, LLVMValueRef *b)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type;
struct lp_build_context bld;
LLVMValueRef c0;
LLVMValueRef c8;
LLVMValueRef c16;
LLVMValueRef c128;
LLVMValueRef c255;
LLVMValueRef cy;
LLVMValueRef cug;
LLVMValueRef cub;
LLVMValueRef cvr;
LLVMValueRef cvg;
memset(&type, 0, sizeof type);
type.sign = TRUE;
type.width = 32;
type.length = n;
lp_build_context_init(&bld, gallivm, type);
assert(lp_check_value(type, y));
assert(lp_check_value(type, u));
assert(lp_check_value(type, v));
/*
* Constants
*/
c0 = lp_build_const_int_vec(gallivm, type, 0);
c8 = lp_build_const_int_vec(gallivm, type, 8);
c16 = lp_build_const_int_vec(gallivm, type, 16);
c128 = lp_build_const_int_vec(gallivm, type, 128);
c255 = lp_build_const_int_vec(gallivm, type, 255);
cy = lp_build_const_int_vec(gallivm, type, 298);
cug = lp_build_const_int_vec(gallivm, type, -100);
cub = lp_build_const_int_vec(gallivm, type, 516);
cvr = lp_build_const_int_vec(gallivm, type, 409);
cvg = lp_build_const_int_vec(gallivm, type, -208);
/*
* y -= 16;
* u -= 128;
* v -= 128;
*/
y = LLVMBuildSub(builder, y, c16, "");
u = LLVMBuildSub(builder, u, c128, "");
v = LLVMBuildSub(builder, v, c128, "");
/*
* r = 298 * _y + 409 * _v + 128;
* g = 298 * _y - 100 * _u - 208 * _v + 128;
* b = 298 * _y + 516 * _u + 128;
*/
y = LLVMBuildMul(builder, y, cy, "");
y = LLVMBuildAdd(builder, y, c128, "");
*r = LLVMBuildMul(builder, v, cvr, "");
*g = LLVMBuildAdd(builder,
LLVMBuildMul(builder, u, cug, ""),
LLVMBuildMul(builder, v, cvg, ""),
"");
*b = LLVMBuildMul(builder, u, cub, "");
*r = LLVMBuildAdd(builder, *r, y, "");
*g = LLVMBuildAdd(builder, *g, y, "");
*b = LLVMBuildAdd(builder, *b, y, "");
/*
* r >>= 8;
* g >>= 8;
* b >>= 8;
*/
*r = LLVMBuildAShr(builder, *r, c8, "r");
*g = LLVMBuildAShr(builder, *g, c8, "g");
*b = LLVMBuildAShr(builder, *b, c8, "b");
/*
* Clamp
*/
*r = lp_build_clamp(&bld, *r, c0, c255);
*g = lp_build_clamp(&bld, *g, c0, c255);
*b = lp_build_clamp(&bld, *b, c0, c255);
}
/**
* 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);
}
