/**
* Generate blend code in SOA mode.
* \param rt render target index (to index the blend / colormask state)
* \param src src/fragment color
* \param dst dst/framebuffer color
* \param con constant blend color
* \param res the result/output
*/
void
lp_build_blend_soa(struct gallivm_state *gallivm,
const struct pipe_blend_state *blend,
struct lp_type type,
unsigned rt,
LLVMValueRef src[4],
LLVMValueRef dst[4],
LLVMValueRef con[4],
LLVMValueRef res[4])
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_blend_soa_context bld;
unsigned i, j, k;
assert(rt < PIPE_MAX_COLOR_BUFS);
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, gallivm, type);
for (i = 0; i < 4; ++i) {
bld.src[i] = src[i];
bld.dst[i] = dst[i];
bld.con[i] = con[i];
}
for (i = 0; i < 4; ++i) {
/* only compute blending for the color channels enabled for writing */
if (blend->rt[rt].colormask & (1 << i)) {
if (blend->logicop_enable) {
if(!type.floating) {
res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
}
else
res[i] = dst[i];
}
else if (blend->rt[rt].blend_enable) {
unsigned src_factor = i < 3 ? blend->rt[rt].rgb_src_factor : blend->rt[rt].alpha_src_factor;
unsigned dst_factor = i < 3 ? blend->rt[rt].rgb_dst_factor : blend->rt[rt].alpha_dst_factor;
unsigned func = i < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
boolean func_commutative = lp_build_blend_func_commutative(func);
/*
* Compute src/dst factors.
*/
bld.factor[0][0][i] = src[i];
bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
bld.factor[1][0][i] = dst[i];
bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
/*
* Check if lp_build_blend can perform any optimisations
*/
res[i] = lp_build_blend(&bld.base,
func,
src_factor,
dst_factor,
bld.factor[0][0][i],
bld.factor[1][0][i],
bld.factor[0][1][i],
bld.factor[1][1][i],
true,
true);
if (res[i]) {
continue;
}
/*
* Compute src/dst terms
*/
for(k = 0; k < 2; ++k) {
/* See if this multiplication has been previously computed */
for(j = 0; j < i; ++j) {
if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
bld.factor[k][1][j] == bld.factor[k][1][i]) ||
(bld.factor[k][0][j] == bld.factor[k][1][i] &&
bld.factor[k][1][j] == bld.factor[k][0][i]))
break;
}
if(j < i && bld.term[k][j])
bld.term[k][i] = bld.term[k][j];
else
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
if (src_factor == PIPE_BLENDFACTOR_ZERO &&
(dst_factor == PIPE_BLENDFACTOR_DST_ALPHA ||
dst_factor == PIPE_BLENDFACTOR_INV_DST_ALPHA)) {
/* XXX special case these combos to work around an apparent
* bug in LLVM.
* This hack disables the check for multiplication by zero
* in lp_bld_mul(). When we optimize away the
* multiplication, something goes wrong during code
* generation and we segfault at runtime.
*/
LLVMValueRef zeroSave = bld.base.zero;
bld.base.zero = NULL;
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i],
bld.factor[k][1][i]);
bld.base.zero = zeroSave;
}
}
/*
* Combine terms
*/
/* See if this function has been previously applied */
for(j = 0; j < i; ++j) {
unsigned prev_func = j < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
if((!func_reverse &&
bld.term[0][j] == bld.term[0][i] &&
bld.term[1][j] == bld.term[1][i]) ||
((func_commutative || func_reverse) &&
bld.term[0][j] == bld.term[1][i] &&
bld.term[1][j] == bld.term[0][i]))
break;
}
if(j < i)
res[i] = res[j];
else
res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
}
else {
res[i] = src[i];
}
}
else {
res[i] = dst[i];
}
}
}
/**
* Generate blend code in SOA mode.
* \param src src/fragment color
* \param dst dst/framebuffer color
* \param con constant blend color
* \param res the result/output
*/
void
lp_build_blend_soa(LLVMBuilderRef builder,
const struct pipe_blend_state *blend,
struct lp_type type,
LLVMValueRef src[4],
LLVMValueRef dst[4],
LLVMValueRef con[4],
LLVMValueRef res[4])
{
struct lp_build_blend_soa_context bld;
unsigned i, j, k;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.base, builder, type);
for (i = 0; i < 4; ++i) {
bld.src[i] = src[i];
bld.dst[i] = dst[i];
bld.con[i] = con[i];
}
for (i = 0; i < 4; ++i) {
if (blend->colormask & (1 << i)) {
if (blend->logicop_enable) {
if(!type.floating) {
res[i] = lp_build_logicop(builder, blend->logicop_func, src[i], dst[i]);
}
else
res[i] = dst[i];
}
else if (blend->blend_enable) {
unsigned src_factor = i < 3 ? blend->rgb_src_factor : blend->alpha_src_factor;
unsigned dst_factor = i < 3 ? blend->rgb_dst_factor : blend->alpha_dst_factor;
unsigned func = i < 3 ? blend->rgb_func : blend->alpha_func;
boolean func_commutative = lp_build_blend_func_commutative(func);
/* It makes no sense to blend unless values are normalized */
assert(type.norm);
/*
* Compute src/dst factors.
*/
bld.factor[0][0][i] = src[i];
bld.factor[0][1][i] = lp_build_blend_soa_factor(&bld, src_factor, i);
bld.factor[1][0][i] = dst[i];
bld.factor[1][1][i] = lp_build_blend_soa_factor(&bld, dst_factor, i);
/*
* Compute src/dst terms
*/
for(k = 0; k < 2; ++k) {
/* See if this multiplication has been previously computed */
for(j = 0; j < i; ++j) {
if((bld.factor[k][0][j] == bld.factor[k][0][i] &&
bld.factor[k][1][j] == bld.factor[k][1][i]) ||
(bld.factor[k][0][j] == bld.factor[k][1][i] &&
bld.factor[k][1][j] == bld.factor[k][0][i]))
break;
}
if(j < i)
bld.term[k][i] = bld.term[k][j];
else
bld.term[k][i] = lp_build_mul(&bld.base, bld.factor[k][0][i], bld.factor[k][1][i]);
}
/*
* Combine terms
*/
/* See if this function has been previously applied */
for(j = 0; j < i; ++j) {
unsigned prev_func = j < 3 ? blend->rgb_func : blend->alpha_func;
unsigned func_reverse = lp_build_blend_func_reverse(func, prev_func);
if((!func_reverse &&
bld.term[0][j] == bld.term[0][i] &&
bld.term[1][j] == bld.term[1][i]) ||
((func_commutative || func_reverse) &&
bld.term[0][j] == bld.term[1][i] &&
bld.term[1][j] == bld.term[0][i]))
break;
}
if(j < i)
res[i] = res[j];
else
res[i] = lp_build_blend_func(&bld.base, func, bld.term[0][i], bld.term[1][i]);
}
else {
res[i] = src[i];
}
}
else {
res[i] = dst[i];
}
}
}
/**
* Performs optimisations and blending independent of SoA/AoS
*
* @param func the blend function
* @param factor_src PIPE_BLENDFACTOR_xxx
* @param factor_dst PIPE_BLENDFACTOR_xxx
* @param src source rgba
* @param dst dest rgba
* @param src_factor src factor computed value
* @param dst_factor dst factor computed value
* @param not_alpha_dependent same factors accross all channels of src/dst
*
* not_alpha_dependent should be:
* SoA: always true as it is only one channel at a time
* AoS: rgb_src_factor == alpha_src_factor && rgb_dst_factor == alpha_dst_factor
*
* Note that pretty much every possible optimisation can only be done on non-unorm targets
* due to unorm values not going above 1.0 meaning factorisation can change results.
* e.g. (0.9 * 0.9) + (0.9 * 0.9) != 0.9 * (0.9 + 0.9) as result of + is always <= 1.
*/
LLVMValueRef
lp_build_blend(struct lp_build_context *bld,
unsigned func,
unsigned factor_src,
unsigned factor_dst,
LLVMValueRef src,
LLVMValueRef dst,
LLVMValueRef src_factor,
LLVMValueRef dst_factor,
boolean not_alpha_dependent,
boolean optimise_only)
{
LLVMValueRef result, src_term, dst_term;
/* If we are not alpha dependent we can mess with the src/dst factors */
if (not_alpha_dependent) {
if (lp_build_blend_factor_complementary(factor_src, factor_dst)) {
if (func == PIPE_BLEND_ADD) {
if (factor_src < factor_dst) {
return lp_build_lerp(bld, src_factor, dst, src, 0);
} else {
return lp_build_lerp(bld, dst_factor, src, dst, 0);
}
} else if(bld->type.floating && func == PIPE_BLEND_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, result, dst);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, src, result);
}
} else if(bld->type.floating && func == PIPE_BLEND_REVERSE_SUBTRACT) {
result = lp_build_add(bld, src, dst);
if (factor_src < factor_dst) {
result = lp_build_mul(bld, result, src_factor);
return lp_build_sub(bld, dst, result);
} else {
result = lp_build_mul(bld, result, dst_factor);
return lp_build_sub(bld, result, src);
}
}
}
if (bld->type.floating && factor_src == factor_dst) {
if (func == PIPE_BLEND_ADD ||
func == PIPE_BLEND_SUBTRACT ||
func == PIPE_BLEND_REVERSE_SUBTRACT) {
LLVMValueRef result;
result = lp_build_blend_func(bld, func, src, dst);
return lp_build_mul(bld, result, src_factor);
}
}
}
if (optimise_only)
return NULL;
src_term = lp_build_mul(bld, src, src_factor);
dst_term = lp_build_mul(bld, dst, dst_factor);
return lp_build_blend_func(bld, func, src_term, dst_term);
}