llvmpipe: Unified common code between AoS and SoA blending.

Added a new file lp_bld_blend.c for the common code.
Merged and added some simple optimisations.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
This commit is contained in:
James Benton 2012-07-03 17:01:14 +01:00 committed by José Fonseca
parent 636646a481
commit 8684ffc141
6 changed files with 286 additions and 158 deletions

View File

@ -5,6 +5,7 @@ LIBNAME = llvmpipe
C_SOURCES = \
lp_bld_alpha.c \
lp_bld_blend.c \
lp_bld_blend_aos.c \
lp_bld_blend_logicop.c \
lp_bld_blend_soa.c \

View File

@ -32,6 +32,7 @@ llvmpipe = env.ConvenienceLibrary(
target = 'llvmpipe',
source = [
'lp_bld_alpha.c',
'lp_bld_blend.c',
'lp_bld_blend_aos.c',
'lp_bld_blend_logicop.c',
'lp_bld_blend_soa.c',

View File

@ -0,0 +1,191 @@
/**************************************************************************
*
* Copyright 2012 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sub license, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice (including the
* next paragraph) shall be included in all copies or substantial portions
* of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
* IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
* ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*
**************************************************************************/
#include "pipe/p_state.h"
#include "util/u_debug.h"
#include "gallivm/lp_bld_type.h"
#include "gallivm/lp_bld_arit.h"
#include "lp_bld_blend.h"
/**
* Is (a OP b) == (b OP a)?
*/
boolean
lp_build_blend_func_commutative(unsigned func)
{
switch (func) {
case PIPE_BLEND_ADD:
case PIPE_BLEND_MIN:
case PIPE_BLEND_MAX:
return TRUE;
case PIPE_BLEND_SUBTRACT:
case PIPE_BLEND_REVERSE_SUBTRACT:
return FALSE;
default:
assert(0);
return TRUE;
}
}
/**
* Whether the blending functions are the reverse of each other.
*/
boolean
lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func)
{
if(rgb_func == alpha_func)
return FALSE;
if(rgb_func == PIPE_BLEND_SUBTRACT && alpha_func == PIPE_BLEND_REVERSE_SUBTRACT)
return TRUE;
if(rgb_func == PIPE_BLEND_REVERSE_SUBTRACT && alpha_func == PIPE_BLEND_SUBTRACT)
return TRUE;
return FALSE;
}
/**
* Whether the blending factors are complementary of each other.
*/
static INLINE boolean
lp_build_blend_factor_complementary(unsigned src_factor, unsigned dst_factor)
{
return dst_factor == (src_factor ^ 0x10);
}
/**
* @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendEquationSeparate.xml
*/
LLVMValueRef
lp_build_blend_func(struct lp_build_context *bld,
unsigned func,
LLVMValueRef term1,
LLVMValueRef term2)
{
switch (func) {
case PIPE_BLEND_ADD:
return lp_build_add(bld, term1, term2);
case PIPE_BLEND_SUBTRACT:
return lp_build_sub(bld, term1, term2);
case PIPE_BLEND_REVERSE_SUBTRACT:
return lp_build_sub(bld, term2, term1);
case PIPE_BLEND_MIN:
return lp_build_min(bld, term1, term2);
case PIPE_BLEND_MAX:
return lp_build_max(bld, term1, term2);
default:
assert(0);
return bld->zero;
}
}
/**
* 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);
} else {
return lp_build_lerp(bld, dst_factor, src, dst);
}
} else if(bld->type.floating && func == PIPE_BLEND_SUBTRACT) {
result = lp_build_add(bld, src, dst);
result = lp_build_mul(bld, result, src_factor);
if (factor_src < factor_dst) {
return lp_build_sub(bld, result, dst);
} else {
return lp_build_sub(bld, src, result);
}
} else if(bld->type.floating && func == PIPE_BLEND_REVERSE_SUBTRACT) {
result = lp_build_add(bld, src, dst);
result = lp_build_mul(bld, result, src_factor);
if (factor_src < factor_dst) {
return lp_build_sub(bld, dst, result);
} else {
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);
}

View File

@ -40,25 +40,17 @@ struct lp_type;
struct lp_build_context;
/**
* Whether the blending function is commutative or not.
*/
boolean
lp_build_blend_func_commutative(unsigned func);
/**
* Whether the blending functions are the reverse of each other.
*/
boolean
lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func);
LLVMValueRef
lp_build_blend_func(struct lp_build_context *bld,
unsigned func,
LLVMValueRef term1,
LLVMValueRef term2);
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
@ -98,4 +90,20 @@ lp_build_logicop(LLVMBuilderRef builder,
LLVMValueRef dst);
LLVMValueRef
lp_build_blend_func(struct lp_build_context *bld,
unsigned func,
LLVMValueRef term1,
LLVMValueRef term2);
boolean
lp_build_blend_func_reverse(unsigned rgb_func,
unsigned alpha_func);
boolean
lp_build_blend_func_commutative(unsigned func);
#endif /* !LP_BLD_BLEND_H */

View File

@ -220,84 +220,21 @@ lp_build_blend_swizzle(struct lp_build_blend_aos_context *bld,
*/
static LLVMValueRef
lp_build_blend_factor(struct lp_build_blend_aos_context *bld,
LLVMValueRef factor1,
unsigned rgb_factor,
unsigned alpha_factor,
unsigned alpha_swizzle)
{
LLVMValueRef rgb_factor_;
LLVMValueRef alpha_factor_;
LLVMValueRef factor2;
LLVMValueRef rgb_factor_, alpha_factor_;
enum lp_build_blend_swizzle rgb_swizzle;
rgb_factor_ = lp_build_blend_factor_unswizzled(bld, rgb_factor, FALSE);
alpha_factor_ = lp_build_blend_factor_unswizzled(bld, alpha_factor, TRUE);
rgb_factor_ = lp_build_blend_factor_unswizzled(bld, rgb_factor, FALSE);
rgb_swizzle = lp_build_blend_factor_swizzle(rgb_factor);
factor2 = lp_build_blend_swizzle(bld, rgb_factor_, alpha_factor_, rgb_swizzle, alpha_swizzle);
return lp_build_mul(&bld->base, factor1, factor2);
}
/**
* Is (a OP b) == (b OP a)?
*/
boolean
lp_build_blend_func_commutative(unsigned func)
{
switch (func) {
case PIPE_BLEND_ADD:
case PIPE_BLEND_MIN:
case PIPE_BLEND_MAX:
return TRUE;
case PIPE_BLEND_SUBTRACT:
case PIPE_BLEND_REVERSE_SUBTRACT:
return FALSE;
default:
assert(0);
return TRUE;
}
}
boolean
lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func)
{
if(rgb_func == alpha_func)
return FALSE;
if(rgb_func == PIPE_BLEND_SUBTRACT && alpha_func == PIPE_BLEND_REVERSE_SUBTRACT)
return TRUE;
if(rgb_func == PIPE_BLEND_REVERSE_SUBTRACT && alpha_func == PIPE_BLEND_SUBTRACT)
return TRUE;
return FALSE;
}
/**
* @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendEquationSeparate.xml
*/
LLVMValueRef
lp_build_blend_func(struct lp_build_context *bld,
unsigned func,
LLVMValueRef term1,
LLVMValueRef term2)
{
switch (func) {
case PIPE_BLEND_ADD:
return lp_build_add(bld, term1, term2);
case PIPE_BLEND_SUBTRACT:
return lp_build_sub(bld, term1, term2);
case PIPE_BLEND_REVERSE_SUBTRACT:
return lp_build_sub(bld, term2, term1);
case PIPE_BLEND_MIN:
return lp_build_min(bld, term1, term2);
case PIPE_BLEND_MAX:
return lp_build_max(bld, term1, term2);
default:
assert(0);
return bld->zero;
if (alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE) {
rgb_swizzle = lp_build_blend_factor_swizzle(rgb_factor);
alpha_factor_ = lp_build_blend_factor_unswizzled(bld, alpha_factor, TRUE);
return lp_build_blend_swizzle(bld, rgb_factor_, alpha_factor_, rgb_swizzle, alpha_swizzle);
} else {
return rgb_factor_;
}
}
@ -329,9 +266,9 @@ lp_build_blend_aos(struct gallivm_state *gallivm,
LLVMValueRef const_,
const unsigned char swizzle[4])
{
const struct pipe_rt_blend_state * state = &blend->rt[rt];
struct lp_build_blend_aos_context bld;
LLVMValueRef src_term;
LLVMValueRef dst_term;
LLVMValueRef src_factor, dst_factor;
LLVMValueRef result;
unsigned alpha_swizzle = swizzle[3];
boolean fullcolormask;
@ -343,45 +280,60 @@ lp_build_blend_aos(struct gallivm_state *gallivm,
bld.dst = dst;
bld.const_ = const_;
if (!blend->rt[rt].blend_enable) {
if (swizzle[3] > UTIL_FORMAT_SWIZZLE_W || swizzle[3] == swizzle[0])
alpha_swizzle = UTIL_FORMAT_SWIZZLE_NONE;
if (!state->blend_enable) {
result = src;
} else {
boolean rgb_alpha_same = state->rgb_src_factor == state->rgb_dst_factor && state->alpha_src_factor == state->alpha_dst_factor;
assert(rgb_alpha_same || alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE);
/* TODO: There are still a few optimization opportunities here. For certain
* combinations it is possible to reorder the operations and therefore saving
* some instructions. */
src_factor = lp_build_blend_factor(&bld, state->rgb_src_factor,
state->alpha_src_factor, alpha_swizzle);
dst_factor = lp_build_blend_factor(&bld, state->rgb_dst_factor,
state->alpha_dst_factor, alpha_swizzle);
src_term = lp_build_blend_factor(&bld, src, blend->rt[rt].rgb_src_factor,
blend->rt[rt].alpha_src_factor, alpha_swizzle);
dst_term = lp_build_blend_factor(&bld, dst, blend->rt[rt].rgb_dst_factor,
blend->rt[rt].alpha_dst_factor, alpha_swizzle);
result = lp_build_blend(&bld.base,
state->rgb_func,
state->rgb_src_factor,
state->rgb_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
lp_build_name(src_term, "src_term");
lp_build_name(dst_term, "dst_term");
if(blend->rt[rt].rgb_func == blend->rt[rt].alpha_func) {
result = lp_build_blend_func(&bld.base, blend->rt[rt].rgb_func, src_term, dst_term);
}
else {
/* Seperate RGB / A functions */
LLVMValueRef rgb;
if(state->rgb_func != state->alpha_func && alpha_swizzle != UTIL_FORMAT_SWIZZLE_NONE) {
LLVMValueRef alpha;
rgb = lp_build_blend_func(&bld.base, blend->rt[rt].rgb_func, src_term, dst_term);
alpha = lp_build_blend_func(&bld.base, blend->rt[rt].alpha_func, src_term, dst_term);
alpha = lp_build_blend(&bld.base,
state->alpha_func,
state->alpha_src_factor,
state->alpha_dst_factor,
src,
dst,
src_factor,
dst_factor,
rgb_alpha_same,
false);
result = lp_build_blend_swizzle(&bld, rgb, alpha, LP_BUILD_BLEND_SWIZZLE_RGBA, alpha_swizzle);
result = lp_build_blend_swizzle(&bld,
result,
alpha,
LP_BUILD_BLEND_SWIZZLE_RGBA,
alpha_swizzle);
}
}
/* Check if color mask is necessary */
fullcolormask = util_format_colormask_full(util_format_description(cbuf_format[rt]), blend->rt[rt].colormask);
fullcolormask = util_format_colormask_full(util_format_description(cbuf_format[rt]), state->colormask);
if (!fullcolormask) {
LLVMValueRef color_mask;
color_mask = lp_build_const_mask_aos_swizzled(gallivm, bld.base.type, blend->rt[rt].colormask, swizzle);
color_mask = lp_build_const_mask_aos_swizzled(gallivm, bld.base.type, state->colormask, swizzle);
lp_build_name(color_mask, "color_mask");
/* Combine with input mask if necessary */

View File

@ -196,13 +196,6 @@ lp_build_blend_soa_factor(struct lp_build_blend_soa_context *bld,
}
static boolean
lp_build_blend_factor_complementary(unsigned src_factor, unsigned dst_factor)
{
return dst_factor == (src_factor ^ 0x10);
}
/**
* Generate blend code in SOA mode.
* \param rt render target index (to index the blend / colormask state)
@ -252,42 +245,6 @@ lp_build_blend_soa(struct gallivm_state *gallivm,
unsigned func = i < 3 ? blend->rt[rt].rgb_func : blend->rt[rt].alpha_func;
boolean func_commutative = lp_build_blend_func_commutative(func);
if (func == PIPE_BLEND_ADD &&
lp_build_blend_factor_complementary(src_factor, dst_factor) && 0) {
/*
* Special case linear interpolation, (i.e., complementary factors).
*/
LLVMValueRef weight;
if (src_factor < dst_factor) {
weight = lp_build_blend_soa_factor(&bld, src_factor, i);
res[i] = lp_build_lerp(&bld.base, weight, dst[i], src[i]);
} else {
weight = lp_build_blend_soa_factor(&bld, dst_factor, i);
res[i] = lp_build_lerp(&bld.base, weight, src[i], dst[i]);
}
continue;
}
if ((func == PIPE_BLEND_ADD ||
func == PIPE_BLEND_SUBTRACT ||
func == PIPE_BLEND_REVERSE_SUBTRACT) &&
src_factor == dst_factor &&
type.floating) {
/*
* Special common factor.
*
* XXX: Only for floating points for now, since saturation will
* cause different results.
*/
LLVMValueRef factor;
factor = lp_build_blend_soa_factor(&bld, src_factor, i);
res[i] = lp_build_blend_func(&bld.base, func, src[i], dst[i]);
res[i] = lp_build_mul(&bld.base, res[i], factor);
continue;
}
/*
* Compute src/dst factors.
*/
@ -297,6 +254,24 @@ lp_build_blend_soa(struct gallivm_state *gallivm,
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
*/
@ -311,7 +286,7 @@ lp_build_blend_soa(struct gallivm_state *gallivm,
break;
}
if(j < i)
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]);