src/gallium/drivers/llvmpipe/lp_bld_blend.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2012 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "pipe/p_state.h"
  29 #include "util/u_debug.h"
  30
  31 #include "gallivm/lp_bld_type.h"
  32 #include "gallivm/lp_bld_arit.h"
  33 #include "gallivm/lp_bld_const.h"
  34 #include "gallivm/lp_bld_logic.h"
  35 #include "gallivm/lp_bld_swizzle.h"
  36 #include "gallivm/lp_bld_flow.h"
  37 #include "gallivm/lp_bld_debug.h"
  38 #include "gallivm/lp_bld_pack.h"
  39
  40 #include "lp_bld_blend.h"
  41
  42 /**
  43  * Is (a OP b) == (b OP a)?
  44  */
  45 boolean
  46 lp_build_blend_func_commutative(unsigned func)
  47 {
  48    switch (func) {
  49    case PIPE_BLEND_ADD:
  50    case PIPE_BLEND_MIN:
  51    case PIPE_BLEND_MAX:
  52       return TRUE;
  53    case PIPE_BLEND_SUBTRACT:
  54    case PIPE_BLEND_REVERSE_SUBTRACT:
  55       return FALSE;
  56    default:
  57       assert(0);
  58       return TRUE;
  59    }
  60 }
  61
  62
  63 /**
  64  * Whether the blending functions are the reverse of each other.
  65  */
  66 boolean
  67 lp_build_blend_func_reverse(unsigned rgb_func, unsigned alpha_func)
  68 {
  69    if (rgb_func == alpha_func)
  70       return FALSE;
  71    if (rgb_func == PIPE_BLEND_SUBTRACT && alpha_func == PIPE_BLEND_REVERSE_SUBTRACT)
  72       return TRUE;
  73    if (rgb_func == PIPE_BLEND_REVERSE_SUBTRACT && alpha_func == PIPE_BLEND_SUBTRACT)
  74       return TRUE;
  75    return FALSE;
  76 }
  77
  78
  79 /**
  80  * Whether the blending factors are complementary of each other.
  81  */
  82 static inline boolean
  83 lp_build_blend_factor_complementary(unsigned src_factor, unsigned dst_factor)
  84 {
  85    STATIC_ASSERT((PIPE_BLENDFACTOR_ZERO ^ 0x10) == PIPE_BLENDFACTOR_ONE);
  86    STATIC_ASSERT((PIPE_BLENDFACTOR_CONST_COLOR ^ 0x10) ==
  87                  PIPE_BLENDFACTOR_INV_CONST_COLOR);
  88    return dst_factor == (src_factor ^ 0x10);
  89 }
  90
  91
  92 /**
  93  * Whether this is a inverse blend factor
  94  */
  95 static inline boolean
  96 is_inverse_factor(unsigned factor)
  97 {
  98    STATIC_ASSERT(PIPE_BLENDFACTOR_ZERO == 0x11);
  99    return factor > 0x11;
 100 }
 101
 102
 103 /**
 104  * Calculates the (expanded to wider type) multiplication
 105  * of 2 normalized numbers.
 106  */
 107 static void
 108 lp_build_mul_norm_expand(struct lp_build_context *bld,
 109                          LLVMValueRef a, LLVMValueRef b,
 110                          LLVMValueRef *resl, LLVMValueRef *resh,
 111                          boolean signedness_differs)
 112 {
 113    const struct lp_type type = bld->type;
 114    struct lp_type wide_type = lp_wider_type(type);
 115    struct lp_type wide_type2 = wide_type;
 116    struct lp_type type2 = type;
 117    LLVMValueRef al, ah, bl, bh;
 118
 119    assert(lp_check_value(type, a));
 120    assert(lp_check_value(type, b));
 121    assert(!type.floating && !type.fixed && type.norm);
 122
 123    if (a == bld->zero || b == bld->zero) {
 124       LLVMValueRef zero = LLVMConstNull(lp_build_vec_type(bld->gallivm, wide_type));
 125       *resl = zero;
 126       *resh = zero;
 127       return;
 128    }
 129
 130    if (signedness_differs) {
 131       type2.sign = !type.sign;
 132       wide_type2.sign = !wide_type2.sign;
 133    }
 134
 135    lp_build_unpack2_native(bld->gallivm, type, wide_type, a, &al, &ah);
 136    lp_build_unpack2_native(bld->gallivm, type2, wide_type2, b, &bl, &bh);
 137
 138    *resl = lp_build_mul_norm(bld->gallivm, wide_type, al, bl);
 139    *resh = lp_build_mul_norm(bld->gallivm, wide_type, ah, bh);
 140 }
 141
 142
 143 /**
 144  * @sa http://www.opengl.org/sdk/docs/man/xhtml/glBlendEquationSeparate.xml
 145  */
 146 LLVMValueRef
 147 lp_build_blend_func(struct lp_build_context *bld,
 148                     unsigned func,
 149                     LLVMValueRef term1,
 150                     LLVMValueRef term2)
 151 {
 152    switch (func) {
 153    case PIPE_BLEND_ADD:
 154       return lp_build_add(bld, term1, term2);
 155    case PIPE_BLEND_SUBTRACT:
 156       return lp_build_sub(bld, term1, term2);
 157    case PIPE_BLEND_REVERSE_SUBTRACT:
 158       return lp_build_sub(bld, term2, term1);
 159    case PIPE_BLEND_MIN:
 160       return lp_build_min(bld, term1, term2);
 161    case PIPE_BLEND_MAX:
 162       return lp_build_max(bld, term1, term2);
 163    default:
 164       assert(0);
 165       return bld->zero;
 166    }
 167 }
 168
 169
 170 /**
 171  * Performs optimisations and blending independent of SoA/AoS
 172  *
 173  * @param func                   the blend function
 174  * @param factor_src             PIPE_BLENDFACTOR_xxx
 175  * @param factor_dst             PIPE_BLENDFACTOR_xxx
 176  * @param src                    source rgba
 177  * @param dst                    dest rgba
 178  * @param src_factor             src factor computed value
 179  * @param dst_factor             dst factor computed value
 180  * @param not_alpha_dependent    same factors accross all channels of src/dst
 181  *
 182  * not_alpha_dependent should be:
 183  *  SoA: always true as it is only one channel at a time
 184  *  AoS: rgb_src_factor == alpha_src_factor && rgb_dst_factor == alpha_dst_factor
 185  *
 186  * Note that pretty much every possible optimisation can only be done on non-unorm targets
 187  * due to unorm values not going above 1.0 meaning factorisation can change results.
 188  * e.g. (0.9 * 0.9) + (0.9 * 0.9) != 0.9 * (0.9 + 0.9) as result of + is always <= 1.
 189  */
 190 LLVMValueRef
 191 lp_build_blend(struct lp_build_context *bld,
 192                unsigned func,
 193                unsigned factor_src,
 194                unsigned factor_dst,
 195                LLVMValueRef src,
 196                LLVMValueRef dst,
 197                LLVMValueRef src_factor,
 198                LLVMValueRef dst_factor,
 199                boolean not_alpha_dependent,
 200                boolean optimise_only)
 201 {
 202    LLVMValueRef result, src_term, dst_term;
 203
 204    /* If we are not alpha dependent we can mess with the src/dst factors */
 205    if (not_alpha_dependent) {
 206       if (lp_build_blend_factor_complementary(factor_src, factor_dst)) {
 207          if (func == PIPE_BLEND_ADD) {
 208             if (factor_src < factor_dst) {
 209                return lp_build_lerp(bld, src_factor, dst, src, 0);
 210             } else {
 211                return lp_build_lerp(bld, dst_factor, src, dst, 0);
 212             }
 213          } else if (bld->type.floating && func == PIPE_BLEND_SUBTRACT) {
 214             result = lp_build_add(bld, src, dst);
 215
 216             if (factor_src < factor_dst) {
 217                result = lp_build_mul(bld, result, src_factor);
 218                return lp_build_sub(bld, result, dst);
 219             } else {
 220                result = lp_build_mul(bld, result, dst_factor);
 221                return lp_build_sub(bld, src, result);
 222             }
 223          } else if (bld->type.floating && func == PIPE_BLEND_REVERSE_SUBTRACT) {
 224             result = lp_build_add(bld, src, dst);
 225
 226             if (factor_src < factor_dst) {
 227                result = lp_build_mul(bld, result, src_factor);
 228                return lp_build_sub(bld, dst, result);
 229             } else {
 230                result = lp_build_mul(bld, result, dst_factor);
 231                return lp_build_sub(bld, result, src);
 232             }
 233          }
 234       }
 235
 236       if (bld->type.floating && factor_src == factor_dst) {
 237          if (func == PIPE_BLEND_ADD ||
 238              func == PIPE_BLEND_SUBTRACT ||
 239              func == PIPE_BLEND_REVERSE_SUBTRACT) {
 240             LLVMValueRef result;
 241             result = lp_build_blend_func(bld, func, src, dst);
 242             return lp_build_mul(bld, result, src_factor);
 243          }
 244       }
 245    }
 246
 247    if (optimise_only)
 248       return NULL;
 249
 250    if ((bld->type.norm && bld->type.sign) &&
 251        (is_inverse_factor(factor_src) || is_inverse_factor(factor_dst))) {
 252       /*
 253        * With snorm blending, the inverse blend factors range from [0,2]
 254        * instead of [-1,1], so the ordinary signed normalized arithmetic
 255        * doesn't quite work. Unpack must be unsigned, and the add/sub
 256        * must be done with wider type.
 257        * (Note that it's not quite obvious what the blend equation wrt to
 258        * clamping should actually be based on GL spec in this case, but
 259        * really the incoming src values are clamped to [-1,1] (the dst is
 260        * always clamped already), and then NO further clamping occurs until
 261        * the end.)
 262        */
 263       struct lp_build_context bldw;
 264       struct lp_type wide_type = lp_wider_type(bld->type);
 265       LLVMValueRef src_terml, src_termh, dst_terml, dst_termh;
 266       LLVMValueRef resl, resh;
 267
 268       /*
 269        * We don't need saturate math for the sub/add, since we have
 270        * x+1 bit numbers in x*2 wide type (result is x+2 bits).
 271        * (Doesn't really matter on x86 sse2 though as we use saturated
 272        * intrinsics.)
 273        */
 274       wide_type.norm = 0;
 275       lp_build_context_init(&bldw, bld->gallivm, wide_type);
 276
 277       /*
 278        * XXX This is a bit hackish. Note that -128 really should
 279        * be -1.0, the same as -127. However, we did not actually clamp
 280        * things anywhere (relying on pack intrinsics instead) therefore
 281        * we will get -128, and the inverted factor then 255. But the mul
 282        * can overflow in this case (rather the rounding fixups for the mul,
 283        * -128*255 will be positive).
 284        * So we clamp the src and dst up here but only when necessary (we
 285        * should do this before calculating blend factors but it's enough
 286        * for avoiding overflow).
 287        */
 288       if (is_inverse_factor(factor_src)) {
 289          src = lp_build_max(bld, src,
 290                             lp_build_const_vec(bld->gallivm, bld->type, -1.0));
 291       }
 292       if (is_inverse_factor(factor_dst)) {
 293          dst = lp_build_max(bld, dst,
 294                             lp_build_const_vec(bld->gallivm, bld->type, -1.0));
 295       }
 296
 297       lp_build_mul_norm_expand(bld, src, src_factor, &src_terml, &src_termh,
 298                                is_inverse_factor(factor_src) ? TRUE : FALSE);
 299       lp_build_mul_norm_expand(bld, dst, dst_factor, &dst_terml, &dst_termh,
 300                                is_inverse_factor(factor_dst) ? TRUE : FALSE);
 301       resl = lp_build_blend_func(&bldw, func, src_terml, dst_terml);
 302       resh = lp_build_blend_func(&bldw, func, src_termh, dst_termh);
 303
 304       /*
 305        * XXX pack2_native is not ok because the values have to be in dst
 306        * range. We need native pack though for the correct order on avx2.
 307        * Will break on everything not implementing clamping pack intrinsics
 308        * (i.e. everything but sse2 and altivec).
 309        */
 310       return lp_build_pack2_native(bld->gallivm, wide_type, bld->type, resl, resh);
 311    } else {
 312       src_term = lp_build_mul(bld, src, src_factor);
 313       dst_term = lp_build_mul(bld, dst, dst_factor);
 314       return lp_build_blend_func(bld, func, src_term, dst_term);
 315    }
 316 }
 317
 318 void
 319 lp_build_alpha_to_coverage(struct gallivm_state *gallivm,
 320                            struct lp_type type,
 321                            struct lp_build_mask_context *mask,
 322                            LLVMValueRef alpha,
 323                            boolean do_branch)
 324 {
 325    struct lp_build_context bld;
 326    LLVMValueRef test;
 327    LLVMValueRef alpha_ref_value;
 328
 329    lp_build_context_init(&bld, gallivm, type);
 330
 331    alpha_ref_value = lp_build_const_vec(gallivm, type, 0.5);
 332
 333    test = lp_build_cmp(&bld, PIPE_FUNC_GREATER, alpha, alpha_ref_value);
 334
 335    lp_build_name(test, "alpha_to_coverage");
 336
 337    lp_build_mask_update(mask, test);
 338
 339    if (do_branch)
 340       lp_build_mask_check(mask);
 341 }