lp_state_fs.c revision efc82aef35a2aac5d2ed9774f6d28f2626796416
1946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca/************************************************************************** 2946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * 373af91e938eb27b001404f11195fb06ff9b08903José Fonseca * Copyright 2009 VMware, Inc. 4946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. 5946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * All Rights Reserved. 6946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * 7946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * Permission is hereby granted, free of charge, to any person obtaining a 8946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * copy of this software and associated documentation files (the 9946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * "Software"), to deal in the Software without restriction, including 10946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * without limitation the rights to use, copy, modify, merge, publish, 11946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * distribute, sub license, and/or sell copies of the Software, and to 12946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * permit persons to whom the Software is furnished to do so, subject to 13946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * the following conditions: 14946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * 15946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * The above copyright notice and this permission notice (including the 16946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * next paragraph) shall be included in all copies or substantial portions 17946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * of the Software. 18946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * 19946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 20946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 21946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 22946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR 23946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 24946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 25946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 26946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca * 27946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca **************************************************************************/ 28946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 295811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca/** 305811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * @file 315811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Code generate the whole fragment pipeline. 325811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 335811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * The fragment pipeline consists of the following stages: 345811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - early depth test 355811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - fragment shader 365811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - alpha test 37d84b8cb1a8e3ce846de9affe2f72457c6e698872Brian Paul * - depth/stencil test 385811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - blending 395811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 40272f399434ad6b33a8444c287c5126987a222864Brian Paul * This file has only the glue to assemble the fragment pipeline. The actual 415811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the 425811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we 435811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * muster the LLVM JIT execution engine to create a function that follows an 445811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * established binary interface and that can be called from C directly. 455811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 465811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * A big source of complexity here is that we often want to run different 475811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * stages with different precisions and data types and precisions. For example, 485811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * the fragment shader needs typically to be done in floats, but the 495811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * depth/stencil test and blending is better done in the type that most closely 505811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * matches the depth/stencil and color buffer respectively. 515811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 525811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Since the width of a SIMD vector register stays the same regardless of the 535811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * element type, different types imply different number of elements, so we must 545811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * code generate more instances of the stages with larger types to be able to 555811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * feed/consume the stages with smaller types. 565811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 575811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * @author Jose Fonseca <jfonseca@vmware.com> 585811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 595811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 60f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#include <limits.h> 61946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "pipe/p_defines.h" 6228486880ca3ec39419ccee0cb1a3bedc9ef7117cJosé Fonseca#include "util/u_inlines.h" 63946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "util/u_memory.h" 64f503b3dd9d6522abdabab1e25d0652c9d3079421José Fonseca#include "util/u_pointer.h" 65e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca#include "util/u_format.h" 66abdcdb3db73316c5ebca88bda04b90b6659782e8José Fonseca#include "util/u_dump.h" 6781862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca#include "util/u_string.h" 6817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger#include "util/u_simple_list.h" 69a8270c7c29b141025f606c811812304a72e42d2eBrian Paul#include "os/os_time.h" 70946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "pipe/p_shader_tokens.h" 71946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "draw/draw_context.h" 72946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_dump.h" 73946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_scan.h" 74946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_parse.h" 75c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_type.h" 76c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_const.h" 77c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_conv.h" 78078eff659a7ef90691966d983f35ed9e4ce63901Vinson Lee#include "gallivm/lp_bld_init.h" 79c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_intr.h" 80c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_logic.h" 81c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_tgsi.h" 82c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_swizzle.h" 83c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_flow.h" 84c61bf363937f40624a5632745630d4f2b9907082Zack Rusin#include "gallivm/lp_bld_debug.h" 8539a12a73269b3eff46ba66fce03eff73dd5d54fbBrian Paul 8639a12a73269b3eff46ba66fce03eff73dd5d54fbBrian Paul#include "lp_bld_alpha.h" 8739a12a73269b3eff46ba66fce03eff73dd5d54fbBrian Paul#include "lp_bld_blend.h" 8839a12a73269b3eff46ba66fce03eff73dd5d54fbBrian Paul#include "lp_bld_depth.h" 8939a12a73269b3eff46ba66fce03eff73dd5d54fbBrian Paul#include "lp_bld_interp.h" 90e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul#include "lp_context.h" 91524ec6464d5c9d66b58c78c5e68c3bcb6a845f69José Fonseca#include "lp_debug.h" 92e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul#include "lp_perf.h" 93e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul#include "lp_screen.h" 9485999695829823e459e11822b4846ed1db5c055dJosé Fonseca#include "lp_setup.h" 9573af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_state.h" 968be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca#include "lp_tex_sample.h" 9717c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger#include "lp_flush.h" 98d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell#include "lp_state_fs.h" 9973af91e938eb27b001404f11195fb06ff9b08903José Fonseca 10073af91e938eb27b001404f11195fb06ff9b08903José Fonseca 101489af2a3ba467e4341cb8504a0e59cf5828864d4Brian Paul#include <llvm-c/Analysis.h> 1023d7479d70568c84354338d0da0b7bed4d296c169Brian Paul#include <llvm-c/BitWriter.h> 103489af2a3ba467e4341cb8504a0e59cf5828864d4Brian Paul 104489af2a3ba467e4341cb8504a0e59cf5828864d4Brian Paul 105efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul/** Fragment shader number (for debugging) */ 10681862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonsecastatic unsigned fs_no = 0; 10781862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca 10881862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca 10998971802798354cdba45c421cc340ec938143e03José Fonseca/** 110d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * Expand the relevent bits of mask_input to a 4-dword mask for the 111ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * four pixels in a 2x2 quad. This will set the four elements of the 112ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * quad mask vector to 0 or ~0. 113d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * 114d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * \param quad which quad of the quad group to test, in [0,3] 115d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * \param mask_input bitwise mask for the whole 4x4 stamp 116ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul */ 117d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwellstatic LLVMValueRef 118efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulgenerate_quad_mask(struct gallivm_state *gallivm, 119d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell struct lp_type fs_type, 120d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell unsigned quad, 121d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef mask_input) /* int32 */ 122ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul{ 123efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMBuilderRef builder = gallivm->builder; 124d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell struct lp_type mask_type; 125efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); 126d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef bits[4]; 127d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef mask; 12867b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell int shift; 129ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 130f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* 131d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * XXX: We'll need a different path for 16 x u8 132f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul */ 133d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell assert(fs_type.width == 32); 134d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell assert(fs_type.length == 4); 135d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_type = lp_int_type(fs_type); 13646b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul 137d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell /* 138d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * mask_input >>= (quad * 4) 1399a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul */ 14067b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell switch (quad) { 14167b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell case 0: 14267b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell shift = 0; 14367b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell break; 14467b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell case 1: 14567b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell shift = 2; 14667b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell break; 14767b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell case 2: 14867b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell shift = 8; 14967b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell break; 15067b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell case 3: 15167b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell shift = 10; 15267b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell break; 153aeb83928fdc09d9310df0f94aaddf87d9635dc7eVinson Lee default: 154aeb83928fdc09d9310df0f94aaddf87d9635dc7eVinson Lee assert(0); 155aeb83928fdc09d9310df0f94aaddf87d9635dc7eVinson Lee shift = 0; 15667b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell } 157ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 158d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_input = LLVMBuildLShr(builder, 159d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_input, 16067b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell LLVMConstInt(i32t, shift, 0), 161d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell ""); 162ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 163d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell /* 164d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * mask = { mask_input & (1 << i), for i in [0,3] } 165d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell */ 166efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul mask = lp_build_broadcast(gallivm, 167efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_vec_type(gallivm, mask_type), 168efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul mask_input); 1694461442849bfdb817334b38567136f7f9dabdf59Brian Paul 170d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell bits[0] = LLVMConstInt(i32t, 1 << 0, 0); 171d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell bits[1] = LLVMConstInt(i32t, 1 << 1, 0); 17267b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell bits[2] = LLVMConstInt(i32t, 1 << 4, 0); 17367b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell bits[3] = LLVMConstInt(i32t, 1 << 5, 0); 17467b957781d8195b8f8867e994c03b68f8dc5c807Keith Whitwell 175d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), ""); 1764461442849bfdb817334b38567136f7f9dabdf59Brian Paul 177d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell /* 178d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * mask = mask != 0 ? ~0 : 0 179d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell */ 180efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul mask = lp_build_compare(gallivm, 181d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_type, PIPE_FUNC_NOTEQUAL, 182d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask, 183efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_const_int_vec(gallivm, mask_type, 0)); 1844461442849bfdb817334b38567136f7f9dabdf59Brian Paul 185d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell return mask; 1862797f2bf57562c95a601a67edca3089641215cc4Brian Paul} 1872797f2bf57562c95a601a67edca3089641215cc4Brian Paul 1882797f2bf57562c95a601a67edca3089641215cc4Brian Paul 1895b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell#define EARLY_DEPTH_TEST 0x1 1905b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell#define LATE_DEPTH_TEST 0x2 1915b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell#define EARLY_DEPTH_WRITE 0x4 1925b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell#define LATE_DEPTH_WRITE 0x8 1935b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 1945b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwellstatic int 1955b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwellfind_output_by_semantic( const struct tgsi_shader_info *info, 1965b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell unsigned semantic, 1975b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell unsigned index ) 1985b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell{ 1995b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell int i; 2005b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2015b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell for (i = 0; i < info->num_outputs; i++) 2025b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (info->output_semantic_name[i] == semantic && 2035b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell info->output_semantic_index[i] == index) 2045b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell return i; 2055b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2065b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell return -1; 2075b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell} 2085b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2092797f2bf57562c95a601a67edca3089641215cc4Brian Paul 210ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul/** 2115811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the fragment shader, depth/stencil test, and alpha tests. 212ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * \param i which quad in the tile, in range [0,3] 213d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell * \param partial_mask if 1, do mask_input testing 21498971802798354cdba45c421cc340ec938143e03José Fonseca */ 21598971802798354cdba45c421cc340ec938143e03José Fonsecastatic void 216efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulgenerate_fs(struct gallivm_state *gallivm, 217efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul struct lp_fragment_shader *shader, 21898971802798354cdba45c421cc340ec938143e03José Fonseca const struct lp_fragment_shader_variant_key *key, 21998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMBuilderRef builder, 220b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type type, 221635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef context_ptr, 22298971802798354cdba45c421cc340ec938143e03José Fonseca unsigned i, 2238009886b0092df2783472deaac1bcaad4a802c19Keith Whitwell struct lp_build_interp_soa_context *interp, 2248be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca struct lp_build_sampler_soa *sampler, 22598971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef *pmask, 226c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef (*color)[4], 227ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef depth_ptr, 22822e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul LLVMValueRef facing, 229d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell unsigned partial_mask, 230d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef mask_input, 23186afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li LLVMValueRef counter) 23273af91e938eb27b001404f11195fb06ff9b08903José Fonseca{ 2335b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell const struct util_format_description *zs_format_desc = NULL; 23473af91e938eb27b001404f11195fb06ff9b08903José Fonseca const struct tgsi_token *tokens = shader->base.tokens; 23573af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMTypeRef vec_type; 236635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef consts_ptr; 23798971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS]; 2388009886b0092df2783472deaac1bcaad4a802c19Keith Whitwell LLVMValueRef z; 2395b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell LLVMValueRef zs_value = NULL; 240521c61ff017ab15b829abbe9a98b179136a36009Brian Paul LLVMValueRef stencil_refs[2]; 24198971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_build_mask_context mask; 242986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 && 243986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca shader->info.base.num_inputs < 3 && 244986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca shader->info.base.num_instructions < 8); 24598971802798354cdba45c421cc340ec938143e03José Fonseca unsigned attrib; 24698971802798354cdba45c421cc340ec938143e03José Fonseca unsigned chan; 247c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell unsigned cbuf; 2485b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell unsigned depth_mode; 2495b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2505b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (key->depth.enabled || 2515b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell key->stencil[0].enabled || 2525b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell key->stencil[1].enabled) { 2535b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2545b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell zs_format_desc = util_format_description(key->zsbuf_format); 2555b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell assert(zs_format_desc); 2565b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 257986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca if (!shader->info.base.writes_z) { 258986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca if (key->alpha.enabled || shader->info.base.uses_kill) 2595b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* With alpha test and kill, can do the depth test early 2605b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell * and hopefully eliminate some quads. But need to do a 2615b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell * special deferred depth write once the final mask value 2625b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell * is known. 2635b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell */ 2645b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE; 2655b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell else 2665b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE; 2675b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 2685b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell else { 2695b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE; 2705b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 2715b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 2725b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (!(key->depth.enabled && key->depth.writemask) && 2735b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell !(key->stencil[0].enabled && key->stencil[0].writemask)) 2745b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE); 2755b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 2765b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell else { 2775b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_mode = 0; 2785b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 27998971802798354cdba45c421cc340ec938143e03José Fonseca 280ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul assert(i < 4); 28198971802798354cdba45c421cc340ec938143e03José Fonseca 282efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr); 283efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr); 28467b82fc395fc9972fc08233044057ab540c7ab59Brian Paul 285efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul vec_type = lp_build_vec_type(gallivm, type); 28698971802798354cdba45c421cc340ec938143e03José Fonseca 287efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul consts_ptr = lp_jit_context_constants(gallivm, context_ptr); 288635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 2898e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca memset(outputs, 0, sizeof outputs); 2908e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 2918e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca /* Declare the color and z variables */ 292c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 293c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) { 294efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color"); 295c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 2968e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca } 2978e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 298ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul /* do triangle edge testing */ 299d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell if (partial_mask) { 300efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul *pmask = generate_quad_mask(gallivm, type, 301d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell i, mask_input); 3022797f2bf57562c95a601a67edca3089641215cc4Brian Paul } 3032797f2bf57562c95a601a67edca3089641215cc4Brian Paul else { 304efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul *pmask = lp_build_const_int_vec(gallivm, type, ~0); 3052797f2bf57562c95a601a67edca3089641215cc4Brian Paul } 306ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 3075771f3d483e882d9f5b6c5f3bdb3c39696623b66Brian Paul /* 'mask' will control execution based on quad's pixel alive/killed state */ 308efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_mask_begin(&mask, gallivm, type, *pmask); 309aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell 3105b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader) 311aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell lp_build_mask_check(&mask); 312aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell 313efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_interp_soa_update_pos(interp, gallivm, i); 314aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell z = interp->pos[2]; 315aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell 3165b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (depth_mode & EARLY_DEPTH_TEST) { 317efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_depth_stencil_test(gallivm, 3185b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &key->depth, 3195b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell key->stencil, 3205b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell type, 3215b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell zs_format_desc, 3225b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &mask, 3235b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell stencil_refs, 3245b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell z, 3255b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_ptr, facing, 3265b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &zs_value, 3275b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell !simple_shader); 3285b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 32995c18abb03b035c6fa029cd0852f07fb39951279José Fonseca if (depth_mode & EARLY_DEPTH_WRITE) { 33095c18abb03b035c6fa029cd0852f07fb39951279José Fonseca lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value); 33195c18abb03b035c6fa029cd0852f07fb39951279José Fonseca } 3325b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 33398971802798354cdba45c421cc340ec938143e03José Fonseca 334efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_interp_soa_update_inputs(interp, gallivm, i); 3355b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 3365b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Build the actual shader */ 337efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_tgsi_soa(gallivm, tokens, type, &mask, 338f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca consts_ptr, interp->pos, interp->inputs, 339986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca outputs, sampler, &shader->info.base); 34098971802798354cdba45c421cc340ec938143e03José Fonseca 3415b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Alpha test */ 3425b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (key->alpha.enabled) { 343986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca int color0 = find_output_by_semantic(&shader->info.base, 3445b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell TGSI_SEMANTIC_COLOR, 3455b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 0); 3465b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 347ffab84c9a27a229e6fa14c3de63868bb843c0f3eKeith Whitwell if (color0 != -1 && outputs[color0][3]) { 3485b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha"); 3495b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell LLVMValueRef alpha_ref_value; 3505b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 351efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr); 352efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value); 3535b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 354efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_alpha_test(gallivm, key->alpha.func, type, 3555b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &mask, alpha, alpha_ref_value, 3565b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell (depth_mode & LATE_DEPTH_TEST) != 0); 35798971802798354cdba45c421cc340ec938143e03José Fonseca } 35898971802798354cdba45c421cc340ec938143e03José Fonseca } 35998971802798354cdba45c421cc340ec938143e03José Fonseca 3605b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Late Z test */ 3615b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (depth_mode & LATE_DEPTH_TEST) { 362986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca int pos0 = find_output_by_semantic(&shader->info.base, 3635b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell TGSI_SEMANTIC_POSITION, 3645b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 0); 3655b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell 366ffab84c9a27a229e6fa14c3de63868bb843c0f3eKeith Whitwell if (pos0 != -1 && outputs[pos0][2]) { 367a0add0446ca9dce6d4a96014c42ba6cf3a73a44aJosé Fonseca z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z"); 3685b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 3696a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 370efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_depth_stencil_test(gallivm, 3715b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &key->depth, 3725b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell key->stencil, 3735b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell type, 3745b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell zs_format_desc, 3755b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &mask, 3765b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell stencil_refs, 3775b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell z, 3785b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_ptr, facing, 3795b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &zs_value, 3805b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell !simple_shader); 3815b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Late Z write */ 38295c18abb03b035c6fa029cd0852f07fb39951279José Fonseca if (depth_mode & LATE_DEPTH_WRITE) { 38395c18abb03b035c6fa029cd0852f07fb39951279José Fonseca lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value); 38495c18abb03b035c6fa029cd0852f07fb39951279José Fonseca } 3855b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 3865b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell else if ((depth_mode & EARLY_DEPTH_TEST) && 3875b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell (depth_mode & LATE_DEPTH_WRITE)) 3885b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell { 3895b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Need to apply a reduced mask to the depth write. Reload the 3905b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell * depth value, update from zs_value with the new mask value and 3915b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell * write that out. 3925b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell */ 393efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_deferred_depth_write(gallivm, 3945b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell type, 3955b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell zs_format_desc, 3965b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell &mask, 3975b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell depth_ptr, 3985b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell zs_value); 3995b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell } 4006a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 4016a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 4025b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell /* Color write */ 403986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib) 4045b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell { 405ac98519c4eed0daf770a9ba380056978e4420352Keith Whitwell if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR && 406ac98519c4eed0daf770a9ba380056978e4420352Keith Whitwell shader->info.base.output_semantic_index[attrib] < key->nr_cbufs) 4075b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell { 408986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca unsigned cbuf = shader->info.base.output_semantic_index[attrib]; 409f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) { 410f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell if(outputs[attrib][chan]) { 411f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell /* XXX: just initialize outputs to point at colors[] and 412f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell * skip this. 413f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell */ 414f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], ""); 415f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]); 416f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell LLVMBuildStore(builder, out, color[cbuf][chan]); 417f0bd76f28d17da6eabf977a7e619e4ff943a70c7Keith Whitwell } 41898971802798354cdba45c421cc340ec938143e03José Fonseca } 41998971802798354cdba45c421cc340ec938143e03José Fonseca } 42098971802798354cdba45c421cc340ec938143e03José Fonseca } 42198971802798354cdba45c421cc340ec938143e03José Fonseca 4225b7eb868fde98388d80601d8dea39e679828f42fKeith Whitwell if (counter) 423efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_occlusion_count(gallivm, type, 424cc40abad519cc0f765c6d8f6fad4154bed8dd9c2José Fonseca lp_build_mask_value(&mask), counter); 42598971802798354cdba45c421cc340ec938143e03José Fonseca 426cc40abad519cc0f765c6d8f6fad4154bed8dd9c2José Fonseca *pmask = lp_build_mask_end(&mask); 42798971802798354cdba45c421cc340ec938143e03José Fonseca} 42898971802798354cdba45c421cc340ec938143e03José Fonseca 42998971802798354cdba45c421cc340ec938143e03José Fonseca 43098971802798354cdba45c421cc340ec938143e03José Fonseca/** 4315811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate color blending and color output. 432f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param rt the render target index (to index blend, colormask state) 433f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param type the pixel color type 434f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param context_ptr pointer to the runtime JIT context 435f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param mask execution mask (active fragment/pixel mask) 436f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param src colors from the fragment shader 437f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul * \param dst_ptr the destination color buffer pointer 43898971802798354cdba45c421cc340ec938143e03José Fonseca */ 43998971802798354cdba45c421cc340ec938143e03José Fonsecastatic void 440efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulgenerate_blend(struct gallivm_state *gallivm, 441efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul const struct pipe_blend_state *blend, 442f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul unsigned rt, 44398971802798354cdba45c421cc340ec938143e03José Fonseca LLVMBuilderRef builder, 444b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type type, 445635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef context_ptr, 44698971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef mask, 44798971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef *src, 448aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell LLVMValueRef dst_ptr, 449aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell boolean do_branch) 45098971802798354cdba45c421cc340ec938143e03José Fonseca{ 45198971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_build_context bld; 452c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca struct lp_build_mask_context mask_ctx; 45398971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef vec_type; 454635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef const_ptr; 45598971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef con[4]; 45698971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef dst[4]; 45798971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef res[4]; 45898971802798354cdba45c421cc340ec938143e03José Fonseca unsigned chan; 45998971802798354cdba45c421cc340ec938143e03José Fonseca 460efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_context_init(&bld, gallivm, type); 461c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 462efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_mask_begin(&mask_ctx, gallivm, type, mask); 463aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell if (do_branch) 464aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell lp_build_mask_check(&mask_ctx); 465c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 466efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul vec_type = lp_build_vec_type(gallivm, type); 46798971802798354cdba45c421cc340ec938143e03José Fonseca 468efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul const_ptr = lp_jit_context_blend_color(gallivm, context_ptr); 469635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca const_ptr = LLVMBuildBitCast(builder, const_ptr, 470635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMPointerType(vec_type, 0), ""); 471635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 472f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul /* load constant blend color and colors from the dest color buffer */ 47398971802798354cdba45c421cc340ec938143e03José Fonseca for(chan = 0; chan < 4; ++chan) { 474efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMValueRef index = lp_build_const_int32(gallivm, chan); 475635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), ""); 47698971802798354cdba45c421cc340ec938143e03José Fonseca 47798971802798354cdba45c421cc340ec938143e03José Fonseca dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), ""); 47898971802798354cdba45c421cc340ec938143e03José Fonseca 47998971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(con[chan], "con.%c", "rgba"[chan]); 48098971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(dst[chan], "dst.%c", "rgba"[chan]); 48198971802798354cdba45c421cc340ec938143e03José Fonseca } 48298971802798354cdba45c421cc340ec938143e03José Fonseca 483f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul /* do blend */ 484efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res); 48598971802798354cdba45c421cc340ec938143e03José Fonseca 486f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul /* store results to color buffer */ 48798971802798354cdba45c421cc340ec938143e03José Fonseca for(chan = 0; chan < 4; ++chan) { 488f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul if(blend->rt[rt].colormask & (1 << chan)) { 489efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMValueRef index = lp_build_const_int32(gallivm, chan); 4904c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca lp_build_name(res[chan], "res.%c", "rgba"[chan]); 4914c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); 4924c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); 4934c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 49498971802798354cdba45c421cc340ec938143e03José Fonseca } 495c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 496c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca lp_build_mask_end(&mask_ctx); 49798971802798354cdba45c421cc340ec938143e03José Fonseca} 49898971802798354cdba45c421cc340ec938143e03José Fonseca 49998971802798354cdba45c421cc340ec938143e03José Fonseca 5005811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca/** 5015811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the runtime callable function for the whole fragment pipeline. 502866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * Note that the function which we generate operates on a block of 16 503866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * pixels at at time. The block contains 2x2 quads. Each quad contains 504866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * 2x2 pixels. 5055811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 5063b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paulstatic void 507efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulgenerate_fragment(struct llvmpipe_context *lp, 50898971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_fragment_shader *shader, 5092797f2bf57562c95a601a67edca3089641215cc4Brian Paul struct lp_fragment_shader_variant *variant, 510d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell unsigned partial_mask) 51198971802798354cdba45c421cc340ec938143e03José Fonseca{ 512efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul struct gallivm_state *gallivm = lp->gallivm; 5133b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul const struct lp_fragment_shader_variant_key *key = &variant->key; 5145b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS]; 51581862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca char func_name[256]; 516b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type fs_type; 517b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type blend_type; 51898971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef fs_elem_type; 51998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef fs_int_vec_type; 52098971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef blend_vec_type; 521d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMTypeRef arg_types[11]; 52273af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMTypeRef func_type; 523efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context); 524efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context); 525c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca LLVMValueRef context_ptr; 5267c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef x; 5277c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef y; 52873af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef a0_ptr; 52973af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef dadx_ptr; 53073af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef dady_ptr; 531c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef color_ptr_ptr; 53239352b34438affc36ce22852bb55359d304a522cJosé Fonseca LLVMValueRef depth_ptr; 533d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef mask_input; 534d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell LLVMValueRef counter = NULL; 53573af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMBasicBlockRef block; 53673af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMBuilderRef builder; 5378be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca struct lp_build_sampler_soa *sampler; 538f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca struct lp_build_interp_soa_context interp; 53998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH]; 540c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH]; 54198971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef blend_mask; 5423b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul LLVMValueRef function; 54322e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul LLVMValueRef facing; 54495c18abb03b035c6fa029cd0852f07fb39951279José Fonseca const struct util_format_description *zs_format_desc; 54598971802798354cdba45c421cc340ec938143e03José Fonseca unsigned num_fs; 54639352b34438affc36ce22852bb55359d304a522cJosé Fonseca unsigned i; 54739352b34438affc36ce22852bb55359d304a522cJosé Fonseca unsigned chan; 548c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell unsigned cbuf; 54973af91e938eb27b001404f11195fb06ff9b08903José Fonseca 5505b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell /* Adjust color input interpolation according to flatshade state: 5515b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell */ 5520072acd447dc6be652e63752e50215c3105322c8Keith Whitwell memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]); 5530072acd447dc6be652e63752e50215c3105322c8Keith Whitwell for (i = 0; i < shader->info.base.num_inputs; i++) { 5545b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell if (inputs[i].interp == LP_INTERP_COLOR) { 5555b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell if (key->flatshade) 5565b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell inputs[i].interp = LP_INTERP_CONSTANT; 5575b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell else 5585b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell inputs[i].interp = LP_INTERP_LINEAR; 5595b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell } 5605b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell } 5615b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 5629ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 5635811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* TODO: actually pick these based on the fs and color buffer 5645811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * characteristics. */ 5655811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 566b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca memset(&fs_type, 0, sizeof fs_type); 56798971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.floating = TRUE; /* floating point values */ 56898971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.sign = TRUE; /* values are signed */ 56998971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ 57098971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.width = 32; /* 32-bit float */ 571866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul fs_type.length = 4; /* 4 elements per vector */ 572866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul num_fs = 4; /* number of quads per block */ 57373af91e938eb27b001404f11195fb06ff9b08903José Fonseca 574b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca memset(&blend_type, 0, sizeof blend_type); 57598971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.floating = FALSE; /* values are integers */ 57698971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.sign = FALSE; /* values are unsigned */ 57798971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */ 57898971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.width = 8; /* 8-bit ubyte values */ 57998971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.length = 16; /* 16 elements per vector */ 58098971802798354cdba45c421cc340ec938143e03José Fonseca 5815811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* 5825811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the function prototype. Any change here must be reflected in 58308dd41fd6899bd6b3289d30dc31f8b2998406889José Fonseca * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa. 5845811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 5855811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 586efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul fs_elem_type = lp_build_elem_type(gallivm, fs_type); 587efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul fs_int_vec_type = lp_build_int_vec_type(gallivm, fs_type); 58898971802798354cdba45c421cc340ec938143e03José Fonseca 589efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul blend_vec_type = lp_build_vec_type(gallivm, blend_type); 59073af91e938eb27b001404f11195fb06ff9b08903José Fonseca 59181862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s", 592d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell shader->no, variant->no, partial_mask ? "partial" : "whole"); 59381862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca 594efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[0] = lp_jit_get_context_type(lp); /* context */ 595efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[1] = int32_type; /* x */ 596efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[2] = int32_type; /* y */ 597efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[3] = int32_type; /* facing */ 59822e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */ 59922e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */ 60022e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */ 60122e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */ 602efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */ 603efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[9] = int32_type; /* mask_input */ 604efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */ 60573af91e938eb27b001404f11195fb06ff9b08903José Fonseca 606efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context), 607efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul arg_types, Elements(arg_types), 0); 60873af91e938eb27b001404f11195fb06ff9b08903José Fonseca 609efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul function = LLVMAddFunction(gallivm->module, func_name, func_type); 6103b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul LLVMSetFunctionCallConv(function, LLVMCCallConv); 6113b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul 612d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell variant->function[partial_mask] = function; 6133b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul 614c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* XXX: need to propagate noalias down into color param now we are 615c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * passing a pointer-to-pointer? 616c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 617a7f9b915aeade96d8ca58d70a1f1d42cd6a16e3dJosé Fonseca for(i = 0; i < Elements(arg_types); ++i) 6187c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) 6193b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute); 6203b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul 6213b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul context_ptr = LLVMGetParam(function, 0); 6223b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul x = LLVMGetParam(function, 1); 6233b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul y = LLVMGetParam(function, 2); 62422e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul facing = LLVMGetParam(function, 3); 62522e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul a0_ptr = LLVMGetParam(function, 4); 62622e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul dadx_ptr = LLVMGetParam(function, 5); 62722e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul dady_ptr = LLVMGetParam(function, 6); 62822e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul color_ptr_ptr = LLVMGetParam(function, 7); 62922e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul depth_ptr = LLVMGetParam(function, 8); 630d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_input = LLVMGetParam(function, 9); 6317c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 632c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca lp_build_name(context_ptr, "context"); 6337c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca lp_build_name(x, "x"); 6347c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca lp_build_name(y, "y"); 6355999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(a0_ptr, "a0"); 6365999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(dadx_ptr, "dadx"); 6375999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(dady_ptr, "dady"); 638f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul lp_build_name(color_ptr_ptr, "color_ptr_ptr"); 63939352b34438affc36ce22852bb55359d304a522cJosé Fonseca lp_build_name(depth_ptr, "depth"); 640d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell lp_build_name(mask_input, "mask_input"); 64173af91e938eb27b001404f11195fb06ff9b08903José Fonseca 64286afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li if (key->occlusion_count) { 643d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell counter = LLVMGetParam(function, 10); 64486afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li lp_build_name(counter, "counter"); 64586afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li } 64686afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li 6475811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* 6485811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Function body 6495811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 6505811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 651efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry"); 652efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul builder = gallivm->builder; 653efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul assert(builder); 65473af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMPositionBuilderAtEnd(builder, block); 65573af91e938eb27b001404f11195fb06ff9b08903José Fonseca 6566ce68ad3ca242076bbb93fdd99bb448f87a31d15José Fonseca /* 6576ce68ad3ca242076bbb93fdd99bb448f87a31d15José Fonseca * The shader input interpolation info is not explicitely baked in the 6586ce68ad3ca242076bbb93fdd99bb448f87a31d15José Fonseca * shader key, but everything it derives from (TGSI, and flatshade) is 6596ce68ad3ca242076bbb93fdd99bb448f87a31d15José Fonseca * already included in the shader key. 6606ce68ad3ca242076bbb93fdd99bb448f87a31d15José Fonseca */ 66186f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell lp_build_interp_soa_init(&interp, 662efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul gallivm, 6630072acd447dc6be652e63752e50215c3105322c8Keith Whitwell shader->info.base.num_inputs, 6645b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell inputs, 66586f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell builder, fs_type, 666f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca a0_ptr, dadx_ptr, dady_ptr, 667147dc2354c21f098a2a63a085c21ec10252cab24José Fonseca x, y); 668f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca 669e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca /* code generated texture sampling */ 670e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr); 671c40eddd294abfe8af3619d08ccd7e9c8f1660fcbJosé Fonseca 672866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul /* loop over quads in the block */ 67395c18abb03b035c6fa029cd0852f07fb39951279José Fonseca zs_format_desc = util_format_description(key->zsbuf_format); 67495c18abb03b035c6fa029cd0852f07fb39951279José Fonseca 67598971802798354cdba45c421cc340ec938143e03José Fonseca for(i = 0; i < num_fs; ++i) { 676efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMValueRef depth_offset = LLVMConstInt(int32_type, 67795c18abb03b035c6fa029cd0852f07fb39951279José Fonseca i*fs_type.length*zs_format_desc->block.bits/8, 67895c18abb03b035c6fa029cd0852f07fb39951279José Fonseca 0); 679c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS]; 68098971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef depth_ptr_i; 68198971802798354cdba45c421cc340ec938143e03José Fonseca 68295c18abb03b035c6fa029cd0852f07fb39951279José Fonseca depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, ""); 68398971802798354cdba45c421cc340ec938143e03José Fonseca 684efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul generate_fs(gallivm, 685efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul shader, key, 68698971802798354cdba45c421cc340ec938143e03José Fonseca builder, 68798971802798354cdba45c421cc340ec938143e03José Fonseca fs_type, 688635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca context_ptr, 68998971802798354cdba45c421cc340ec938143e03José Fonseca i, 690f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca &interp, 6918be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca sampler, 692ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul &fs_mask[i], /* output */ 69398971802798354cdba45c421cc340ec938143e03José Fonseca out_color, 694ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul depth_ptr_i, 69522e6dc387039e79f6d1435ae8b7422a6514d5d10Brian Paul facing, 696d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell partial_mask, 697d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell mask_input, 698d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell counter); 69998971802798354cdba45c421cc340ec938143e03José Fonseca 700c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) 701c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) 702c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell fs_out_color[cbuf][chan][i] = out_color[cbuf][chan]; 70398971802798354cdba45c421cc340ec938143e03José Fonseca } 70439352b34438affc36ce22852bb55359d304a522cJosé Fonseca 7058be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca sampler->destroy(sampler); 7068be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca 707c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* Loop over color outputs / color buffers to do blending. 7085811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 709c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 710c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef color_ptr; 711efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMValueRef index = lp_build_const_int32(gallivm, cbuf); 712f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul LLVMValueRef blend_in_color[NUM_CHANNELS]; 713f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul unsigned rt; 7145811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 715c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* 716c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * Convert the fs's output color and mask to fit to the blending type. 717c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 718c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) { 71940d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH]; 72040d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell 72140d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell for (i = 0; i < num_fs; i++) { 72240d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell fs_color_vals[i] = 72340d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals"); 72440d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell } 72540d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell 726efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_conv(gallivm, fs_type, blend_type, 72740d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell fs_color_vals, 72840d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell num_fs, 729c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell &blend_in_color[chan], 1); 73040d7be52619fbff2479dcdf56929e3e0c5b12e72Keith Whitwell 731c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]); 732c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 733635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 734d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell if (partial_mask || !variant->opaque) { 735efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp_build_conv_mask(lp->gallivm, fs_type, blend_type, 736d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell fs_mask, num_fs, 737d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell &blend_mask, 1); 738d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell } else { 739efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul blend_mask = lp_build_const_int_vec(lp->gallivm, blend_type, ~0); 740d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell } 741c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 742c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color_ptr = LLVMBuildLoad(builder, 743c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""), 744c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell ""); 745c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_name(color_ptr, "color_ptr%d", cbuf); 746c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 747f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul /* which blend/colormask state to use */ 748f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul rt = key->blend.independent_blend_enable ? cbuf : 0; 749f795735f4251d5f7842ee9e09994641c5c46d25dBrian Paul 750c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* 751c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * Blending. 752c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 753aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell { 754aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell /* Could the 4x4 have been killed? 755aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell */ 756aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) && 757aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell !key->alpha.enabled && 758986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca !shader->info.base.uses_kill); 759aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell 760efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul generate_blend(lp->gallivm, 761efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul &key->blend, 762aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell rt, 763aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell builder, 764aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell blend_type, 765aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell context_ptr, 766aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell blend_mask, 767aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell blend_in_color, 768aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell color_ptr, 769aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell do_branch); 770aa4cb5e2d8d48c7dcc9653c61a9e25494e3e7b2aKeith Whitwell } 77173af91e938eb27b001404f11195fb06ff9b08903José Fonseca } 77273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 7735811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca LLVMBuildRetVoid(builder); 77473af91e938eb27b001404f11195fb06ff9b08903José Fonseca 775f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Verify the LLVM IR. If invalid, dump and abort */ 7764ae3e88dc9856f2f32c37dd04a3321765ed61e07José Fonseca#ifdef DEBUG 7773b1920a34903dfb753bc2a0461fef204d39846c6Brian Paul if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) { 778f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul if (1) 7798ad3e0b55df50beac8ba3c5cafa0be79641a4977José Fonseca lp_debug_dump_value(function); 780f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul abort(); 781a02ecdf8c2fc5783a4bc82e8cd9d36f0dec7ccecJosé Fonseca } 7824ae3e88dc9856f2f32c37dd04a3321765ed61e07José Fonseca#endif 783a02ecdf8c2fc5783a4bc82e8cd9d36f0dec7ccecJosé Fonseca 784f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Apply optimizations to LLVM IR */ 785efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMRunFunctionPassManager(gallivm->passmgr, function); 78673af91e938eb27b001404f11195fb06ff9b08903José Fonseca 787954965366fee3fa2eec8a11b6663d4cf218e1d5dKeith Whitwell if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) { 788f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Print the LLVM IR to stderr */ 7898ad3e0b55df50beac8ba3c5cafa0be79641a4977José Fonseca lp_debug_dump_value(function); 7902584c5bd253e53ba052356360a33b5ec976e9716José Fonseca debug_printf("\n"); 7912584c5bd253e53ba052356360a33b5ec976e9716José Fonseca } 79273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 7933d7479d70568c84354338d0da0b7bed4d296c169Brian Paul /* Dump byte code to a file */ 7943d7479d70568c84354338d0da0b7bed4d296c169Brian Paul if (0) { 795efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc"); 7963d7479d70568c84354338d0da0b7bed4d296c169Brian Paul } 7973d7479d70568c84354338d0da0b7bed4d296c169Brian Paul 798f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* 799f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul * Translate the LLVM IR into machine code. 800f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul */ 8018dc8c3f5b11d5f158b0027d1501555c898e0451eBrian Paul { 802efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul void *f = LLVMGetPointerToGlobal(gallivm->engine, function); 8038dc8c3f5b11d5f158b0027d1501555c898e0451eBrian Paul 804d4b64167b56f780d0dea73193c345622888fbc16Keith Whitwell variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f); 80573af91e938eb27b001404f11195fb06ff9b08903José Fonseca 806954965366fee3fa2eec8a11b6663d4cf218e1d5dKeith Whitwell if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) { 8078dc8c3f5b11d5f158b0027d1501555c898e0451eBrian Paul lp_disassemble(f); 8088a3bcba12c19049eb91cb8ce865212d75be5c859José Fonseca } 809042018a943a403a4d9887b400deb3b3c83ee40c0Zack Rusin lp_func_delete_body(function); 8108dc8c3f5b11d5f158b0027d1501555c898e0451eBrian Paul } 8113b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul} 8123b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 8133b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 814c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paulstatic void 815c2a036d4efeca80263e4d212532e46dfc525b9daBrian Pauldump_fs_variant_key(const struct lp_fragment_shader_variant_key *key) 816c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul{ 817c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul unsigned i; 818c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 819d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("fs variant %p:\n", (void *) key); 820d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul 821446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca if (key->flatshade) { 822446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca debug_printf("flatshade = 1\n"); 823446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca } 82400989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca for (i = 0; i < key->nr_cbufs; ++i) { 82500989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i])); 82600989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca } 827c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul if (key->depth.enabled) { 828c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format)); 829c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE)); 830c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("depth.writemask = %u\n", key->depth.writemask); 831c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 832c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 833c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul for (i = 0; i < 2; ++i) { 834c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul if (key->stencil[i].enabled) { 835c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE)); 836c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE)); 837c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE)); 838c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE)); 839c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask); 840c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask); 841c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 842c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 843c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 844c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul if (key->alpha.enabled) { 845c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE)); 846c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 847c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 848446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca if (key->occlusion_count) { 849446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca debug_printf("occlusion_count = 1\n"); 850446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca } 851446dbb921762710d678486f3f5a6dfdf318fe34cJosé Fonseca 852c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul if (key->blend.logicop_enable) { 85375189ef960d2f7001d2b24e7c1939d0a647c0cf1Brian Paul debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE)); 854c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 855c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul else if (key->blend.rt[0].blend_enable) { 856c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE)); 857d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE)); 858d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE)); 859d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE)); 860d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE)); 861d574ffbb4d3a5e15a0e309686d3888c3112fcf21Brian Paul debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE)); 862c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 863c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask); 864591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell for (i = 0; i < key->nr_samplers; ++i) { 865591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf("sampler[%u] = \n", i); 866591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .format = %s\n", 867591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_format_name(key->sampler[i].format)); 868591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .target = %s\n", 869591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_target(key->sampler[i].target, TRUE)); 870591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .pot = %u %u %u\n", 871591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell key->sampler[i].pot_width, 872591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell key->sampler[i].pot_height, 873591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell key->sampler[i].pot_depth); 874591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .wrap = %s %s %s\n", 875591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE), 876591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE), 877591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE)); 878591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .min_img_filter = %s\n", 879591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE)); 880591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .min_mip_filter = %s\n", 881591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE)); 882591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .mag_img_filter = %s\n", 883591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE)); 884591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE) 885591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE)); 886591e1bc34f5a5dd065614deae41b59682f59ac08Keith Whitwell debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords); 8871c32583581ef5aee59901d9dd8e56cc1a125f0d4José Fonseca debug_printf(" .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal); 8881c32583581ef5aee59901d9dd8e56cc1a125f0d4José Fonseca debug_printf(" .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero); 8891c32583581ef5aee59901d9dd8e56cc1a125f0d4José Fonseca debug_printf(" .apply_min_lod = %u\n", key->sampler[i].apply_min_lod); 8901c32583581ef5aee59901d9dd8e56cc1a125f0d4José Fonseca debug_printf(" .apply_max_lod = %u\n", key->sampler[i].apply_max_lod); 891c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul } 892c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul} 893c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 894c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 89571e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwellvoid 89671e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwelllp_debug_fs_variant(const struct lp_fragment_shader_variant *variant) 89771e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell{ 89871e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n", 89971e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell variant->shader->no, variant->no); 90071e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell tgsi_dump(variant->shader->base.tokens, 0); 90171e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell dump_fs_variant_key(&variant->key); 90271e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell debug_printf("variant->opaque = %u\n", variant->opaque); 90371e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell debug_printf("\n"); 90471e27ef21c4a24aeadadb85f60774ec48a7e3cffKeith Whitwell} 905c2a036d4efeca80263e4d212532e46dfc525b9daBrian Paul 906efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 907efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul/** 908efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * Generate a new fragment shader variant from the shader code and 909efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * other state indicated by the key. 910efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul */ 9113b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paulstatic struct lp_fragment_shader_variant * 912efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulgenerate_variant(struct llvmpipe_context *lp, 9133b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul struct lp_fragment_shader *shader, 9143b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul const struct lp_fragment_shader_variant_key *key) 9153b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul{ 9163b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul struct lp_fragment_shader_variant *variant; 91700989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca boolean fullcolormask; 9183b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 9193b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul variant = CALLOC_STRUCT(lp_fragment_shader_variant); 9203b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(!variant) 9213b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul return NULL; 9223b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 923197636814e4e523b98d3e25d918d589ef61b8acaJosé Fonseca variant->shader = shader; 92417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger variant->list_item_global.base = variant; 92517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger variant->list_item_local.base = variant; 92617c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger variant->no = shader->variants_created++; 92781862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca 9283d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell memcpy(&variant->key, key, shader->variant_key_size); 929818d444e12bb57568fbf3bf5f06ee24c6c73a61aJosé Fonseca 93000989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca /* 93100989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca * Determine whether we are touching all channels in the color buffer. 93200989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca */ 93300989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca fullcolormask = FALSE; 93400989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca if (key->nr_cbufs == 1) { 93500989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca const struct util_format_description *format_desc; 93600989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca format_desc = util_format_description(key->cbuf_format[0]); 93700989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca if ((~key->blend.rt[0].colormask & 93800989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca util_format_colormask(format_desc)) == 0) { 93900989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca fullcolormask = TRUE; 94000989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca } 94100989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca } 94200989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca 94310fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca variant->opaque = 94410fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca !key->blend.logicop_enable && 94510fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca !key->blend.rt[0].blend_enable && 94600989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca fullcolormask && 94710fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca !key->stencil[0].enabled && 94810fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca !key->alpha.enabled && 94910fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca !key->depth.enabled && 950986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca !shader->info.base.uses_kill 95110fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca ? TRUE : FALSE; 95210fdbb9298489e9dfb2ecec0662abe29da5b6239José Fonseca 9539f6e8e1d6b8696a3ee96cba01b2466ba7a1a8ef6Keith Whitwell 95409863554259aeb7b0522b130837b6a1f12ffaebeKeith Whitwell if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) { 9559f6e8e1d6b8696a3ee96cba01b2466ba7a1a8ef6Keith Whitwell lp_debug_fs_variant(variant); 9569f6e8e1d6b8696a3ee96cba01b2466ba7a1a8ef6Keith Whitwell } 9579f6e8e1d6b8696a3ee96cba01b2466ba7a1a8ef6Keith Whitwell 958efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul generate_fragment(lp, shader, variant, RAST_EDGE_TEST); 9599f6e8e1d6b8696a3ee96cba01b2466ba7a1a8ef6Keith Whitwell 960853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca if (variant->opaque) { 961853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca /* Specialized shader, which doesn't need to read the color buffer. */ 962efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul generate_fragment(lp, shader, variant, RAST_WHOLE); 963853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca } else { 964853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST]; 965853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca } 966853953dc3c5bd2ef9cca904c9cabc4a0ef037b16José Fonseca 9679ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca return variant; 9689ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca} 9699ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9709ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9719523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulstatic void * 9729ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonsecallvmpipe_create_fs_state(struct pipe_context *pipe, 9739ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca const struct pipe_shader_state *templ) 9749ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca{ 975955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 9769ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader *shader; 9773d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell int nr_samplers; 9785b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell int i; 9799ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9809ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader = CALLOC_STRUCT(lp_fragment_shader); 9819ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca if (!shader) 9829ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca return NULL; 9839ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 98481862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca shader->no = fs_no++; 98517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger make_empty_list(&shader->variants); 98681862bbbc82341b1271fc4bc7b803d409826aebaJosé Fonseca 9879ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca /* get/save the summary info for this shader */ 988986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca lp_build_tgsi_info(templ->tokens, &shader->info); 9899ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9909ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca /* we need to keep a local copy of the tokens */ 9919ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader->base.tokens = tgsi_dup_tokens(templ->tokens); 9929ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 993955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ); 994955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul if (shader->draw_data == NULL) { 995955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul FREE((void *) shader->base.tokens); 996b556bb7c44236a9fae54f58cc03e1d05eaa2124fJosé Fonseca FREE(shader); 997955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul return NULL; 998955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul } 999955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul 1000986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1; 10013d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell 10023d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key, 10033d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell sampler[nr_samplers]); 10043d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell 10050072acd447dc6be652e63752e50215c3105322c8Keith Whitwell for (i = 0; i < shader->info.base.num_inputs; i++) { 10060072acd447dc6be652e63752e50215c3105322c8Keith Whitwell shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i]; 10075b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 10080072acd447dc6be652e63752e50215c3105322c8Keith Whitwell switch (shader->info.base.input_interpolate[i]) { 10095b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_INTERPOLATE_CONSTANT: 10105b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_CONSTANT; 10115b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10125b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_INTERPOLATE_LINEAR: 10135b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_LINEAR; 10145b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10155b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_INTERPOLATE_PERSPECTIVE: 10165b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_PERSPECTIVE; 10175b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10185b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell default: 10195b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell assert(0); 10205b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10215b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell } 10225b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 10230072acd447dc6be652e63752e50215c3105322c8Keith Whitwell switch (shader->info.base.input_semantic_name[i]) { 10245b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_SEMANTIC_COLOR: 10255b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell /* Colors may be either linearly or constant interpolated in 10265b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell * the fragment shader, but that information isn't available 10275b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell * here. Mark color inputs and fix them up later. 10285b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell */ 10295b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_COLOR; 10305b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10315b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_SEMANTIC_FACE: 10325b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_FACING; 10335b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell break; 10345b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell case TGSI_SEMANTIC_POSITION: 10355b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell /* Position was already emitted above 10365b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell */ 10375b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].interp = LP_INTERP_POSITION; 10385b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].src_index = 0; 10395b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell continue; 10405b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell } 10415b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 10425b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell shader->inputs[i].src_index = i+1; 10435b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell } 10445b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 1045524ec6464d5c9d66b58c78c5e68c3bcb6a845f69José Fonseca if (LP_DEBUG & DEBUG_TGSI) { 1046149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca unsigned attrib; 1047efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul debug_printf("llvmpipe: Create fragment shader #%u %p:\n", 1048efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul shader->no, (void *) shader); 1049ac6725b8f9c331cacd08d95eb22d0c81b0bc9821Brian Paul tgsi_dump(templ->tokens, 0); 1050149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca debug_printf("usage masks:\n"); 1051986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) { 1052986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca unsigned usage_mask = shader->info.base.input_usage_mask[attrib]; 1053149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca debug_printf(" IN[%u].%s%s%s%s\n", 1054149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca attrib, 1055149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca usage_mask & TGSI_WRITEMASK_X ? "x" : "", 1056149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca usage_mask & TGSI_WRITEMASK_Y ? "y" : "", 1057149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca usage_mask & TGSI_WRITEMASK_Z ? "z" : "", 1058149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca usage_mask & TGSI_WRITEMASK_W ? "w" : ""); 1059149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca } 1060149cb7682e37ce719d693f120e68dde60ba73bdfJosé Fonseca debug_printf("\n"); 1061ac6725b8f9c331cacd08d95eb22d0c81b0bc9821Brian Paul } 1062ac6725b8f9c331cacd08d95eb22d0c81b0bc9821Brian Paul 106373af91e938eb27b001404f11195fb06ff9b08903José Fonseca return shader; 1064946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 1065946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1066946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 10679523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulstatic void 1068946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs) 1069946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 1070946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 1071946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1072080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca if (llvmpipe->fs == fs) 1073080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca return; 1074080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca 1075080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca draw_flush(llvmpipe->draw); 1076080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca 10772996ce72b142c774101f8df8bd6050d4755ccdccBrian Paul llvmpipe->fs = (struct lp_fragment_shader *) fs; 10789fbf744389cc13e96324086a1a94afbac933e6eaBrian Paul 1079955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul draw_bind_fragment_shader(llvmpipe->draw, 1080955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL)); 1081955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul 1082c9a5930fe45a0a0299769bd2b672ca516d1bf39eJosé Fonseca llvmpipe->dirty |= LP_NEW_FS; 1083946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 1084946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1085efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1086efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul/** 1087efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * Remove shader variant from two lists: the shader's variant list 1088efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * and the context's variant list. 1089efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul */ 1090efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulvoid 1091efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paulllvmpipe_remove_shader_variant(struct llvmpipe_context *lp, 1092efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul struct lp_fragment_shader_variant *variant) 109317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger{ 109417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger unsigned i; 109517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger 109617c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (gallivm_debug & GALLIVM_DEBUG_IR) { 1097efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached" 1098efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul " #%u v total cached #%u\n", 1099efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul variant->shader->no, 1100efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul variant->no, 1101efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul variant->shader->variants_created, 1102efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul variant->shader->variants_cached, 1103efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul lp->nr_fs_variants); 110417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 1105efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1106efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* free all the variant's JIT'd functions */ 110717c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger for (i = 0; i < Elements(variant->function); i++) { 110817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (variant->function[i]) { 110917c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (variant->jit_function[i]) 1110efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul LLVMFreeMachineCodeForFunction(lp->gallivm->engine, 111117c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger variant->function[i]); 111217c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger LLVMDeleteFunction(variant->function[i]); 111317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 111417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 1115efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1116efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* remove from shader's list */ 111717c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger remove_from_list(&variant->list_item_local); 1118197636814e4e523b98d3e25d918d589ef61b8acaJosé Fonseca variant->shader->variants_cached--; 1119efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1120efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* remove from context's list */ 112117c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger remove_from_list(&variant->list_item_global); 112217c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger lp->nr_fs_variants--; 1123efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 112417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger FREE(variant); 112517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger} 1126946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1127efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 11289523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulstatic void 1129946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs) 1130946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 1131e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 113273af91e938eb27b001404f11195fb06ff9b08903José Fonseca struct lp_fragment_shader *shader = fs; 113317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger struct lp_fs_variant_list_item *li; 1134946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1135e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca assert(fs != llvmpipe->fs); 11369ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 11374d5dd82558fcf1f6fcfdb95b225a4e3f842a9133José Fonseca /* 11384d5dd82558fcf1f6fcfdb95b225a4e3f842a9133José Fonseca * XXX: we need to flush the context until we have some sort of reference 11394d5dd82558fcf1f6fcfdb95b225a4e3f842a9133José Fonseca * counting in fragment shaders as they may still be binned 114017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger * Flushing alone might not sufficient we need to wait on it too. 11414d5dd82558fcf1f6fcfdb95b225a4e3f842a9133José Fonseca */ 1142b6e03eafe3311142445ca42c1574d3f6998eecc3Keith Whitwell llvmpipe_finish(pipe, __FUNCTION__); 11439ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1144efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Delete all the variants */ 114517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger li = first_elem(&shader->variants); 114617c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger while(!at_end(&shader->variants, li)) { 1147d7f4250333f3649d170d71bfdc5f2faba4e8c5f8José Fonseca struct lp_fs_variant_list_item *next = next_elem(li); 1148efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul llvmpipe_remove_shader_variant(llvmpipe, li->base); 1149d7f4250333f3649d170d71bfdc5f2faba4e8c5f8José Fonseca li = next; 115073af91e938eb27b001404f11195fb06ff9b08903José Fonseca } 115173af91e938eb27b001404f11195fb06ff9b08903José Fonseca 1152efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Delete draw module's data */ 1153955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data); 1154955d76c3d2004c058c326d68eddc5a06d1611a41Brian Paul 115517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger assert(shader->variants_cached == 0); 115673af91e938eb27b001404f11195fb06ff9b08903José Fonseca FREE((void *) shader->base.tokens); 115773af91e938eb27b001404f11195fb06ff9b08903José Fonseca FREE(shader); 1158946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 1159946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1160946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1161946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 11629523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulstatic void 1163946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_set_constant_buffer(struct pipe_context *pipe, 1164946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca uint shader, uint index, 1165287c94ea4987033f9c99a2f91c5750c9083504caKeith Whitwell struct pipe_resource *constants) 1166946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 1167946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 1168287c94ea4987033f9c99a2f91c5750c9083504caKeith Whitwell unsigned size = constants ? constants->width0 : 0; 11690639765b2850739af1678f10fc0c5706d5827776Brian Paul const void *data = constants ? llvmpipe_resource_data(constants) : NULL; 1170946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1171946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca assert(shader < PIPE_SHADER_TYPES); 1172a6d9d18faecef9963be3e4b64a21b89889b4670dKeith Whitwell assert(index < PIPE_MAX_CONSTANT_BUFFERS); 1173946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1174a6d9d18faecef9963be3e4b64a21b89889b4670dKeith Whitwell if(llvmpipe->constants[shader][index] == constants) 1175d904ed88c1d957f662497343de7dc3e9fa743e47José Fonseca return; 1176d904ed88c1d957f662497343de7dc3e9fa743e47José Fonseca 1177080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca draw_flush(llvmpipe->draw); 117869588d7ed59a019a5272a9cc391e30c47d006aeeJosé Fonseca 1179946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca /* note: reference counting */ 1180a6d9d18faecef9963be3e4b64a21b89889b4670dKeith Whitwell pipe_resource_reference(&llvmpipe->constants[shader][index], constants); 118169588d7ed59a019a5272a9cc391e30c47d006aeeJosé Fonseca 1182b5e381d9783f17c9a527ac38122444eac6807566Zack Rusin if(shader == PIPE_SHADER_VERTEX || 1183b5e381d9783f17c9a527ac38122444eac6807566Zack Rusin shader == PIPE_SHADER_GEOMETRY) { 1184b5e381d9783f17c9a527ac38122444eac6807566Zack Rusin draw_set_mapped_constant_buffer(llvmpipe->draw, shader, 1185b5e381d9783f17c9a527ac38122444eac6807566Zack Rusin index, data, size); 118669588d7ed59a019a5272a9cc391e30c47d006aeeJosé Fonseca } 1187946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 1188c9a5930fe45a0a0299769bd2b672ca516d1bf39eJosé Fonseca llvmpipe->dirty |= LP_NEW_CONSTANTS; 1189946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 11909ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 11919ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1192635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca/** 119359253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * Return the blend factor equivalent to a destination alpha of one. 119459253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca */ 119559253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonsecastatic INLINE unsigned 119687267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonsecaforce_dst_alpha_one(unsigned factor) 119759253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca{ 119859253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca switch(factor) { 119959253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca case PIPE_BLENDFACTOR_DST_ALPHA: 120059253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca return PIPE_BLENDFACTOR_ONE; 120159253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca case PIPE_BLENDFACTOR_INV_DST_ALPHA: 120259253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca return PIPE_BLENDFACTOR_ZERO; 120359253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: 120459253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca return PIPE_BLENDFACTOR_ZERO; 120559253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca } 120659253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca 120759253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca return factor; 120859253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca} 120959253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca 121059253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca 121159253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca/** 1212635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * We need to generate several variants of the fragment pipeline to match 1213635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * all the combinations of the contributing state atoms. 1214635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * 1215635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * TODO: there is actually no reason to tie this to context state -- the 1216635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * generated code could be cached globally in the screen. 1217635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca */ 1218635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonsecastatic void 1219635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonsecamake_variant_key(struct llvmpipe_context *lp, 1220e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca struct lp_fragment_shader *shader, 1221635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca struct lp_fragment_shader_variant_key *key) 1222635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca{ 1223e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca unsigned i; 1224e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca 12253d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell memset(key, 0, shader->variant_key_size); 1226635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 122767b82fc395fc9972fc08233044057ab540c7ab59Brian Paul if (lp->framebuffer.zsbuf) { 122867b82fc395fc9972fc08233044057ab540c7ab59Brian Paul if (lp->depth_stencil->depth.enabled) { 122967b82fc395fc9972fc08233044057ab540c7ab59Brian Paul key->zsbuf_format = lp->framebuffer.zsbuf->format; 123067b82fc395fc9972fc08233044057ab540c7ab59Brian Paul memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); 123167b82fc395fc9972fc08233044057ab540c7ab59Brian Paul } 123267b82fc395fc9972fc08233044057ab540c7ab59Brian Paul if (lp->depth_stencil->stencil[0].enabled) { 123367b82fc395fc9972fc08233044057ab540c7ab59Brian Paul key->zsbuf_format = lp->framebuffer.zsbuf->format; 123467b82fc395fc9972fc08233044057ab540c7ab59Brian Paul memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil); 123567b82fc395fc9972fc08233044057ab540c7ab59Brian Paul } 1236cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca } 1237635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 1238635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca key->alpha.enabled = lp->depth_stencil->alpha.enabled; 1239635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca if(key->alpha.enabled) 1240635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca key->alpha.func = lp->depth_stencil->alpha.func; 1241635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca /* alpha.ref_value is passed in jit_context */ 1242635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 124386f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell key->flatshade = lp->rasterizer->flatshade; 124486afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li if (lp->active_query_count) { 124586afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li key->occlusion_count = TRUE; 124686afe8250edaa2e6129c937a62a695f616c48d70Qicheng Christopher Li } 12474c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 1248c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell if (lp->framebuffer.nr_cbufs) { 12494c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca memcpy(&key->blend, lp->blend, sizeof key->blend); 1250c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 12514c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 1252c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell key->nr_cbufs = lp->framebuffer.nr_cbufs; 1253c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for (i = 0; i < lp->framebuffer.nr_cbufs; i++) { 125400989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca enum pipe_format format = lp->framebuffer.cbufs[i]->format; 125559253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i]; 12564c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca const struct util_format_description *format_desc; 12574c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 125800989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca key->cbuf_format[i] = format; 125900989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca 126000989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca format_desc = util_format_description(format); 12616ae9975ea08b64d7f4e7a2c6c535c14280bef843José Fonseca assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB || 12626ae9975ea08b64d7f4e7a2c6c535c14280bef843José Fonseca format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB); 12634c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 126459253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca blend_rt->colormask = lp->blend->rt[i].colormask; 1265862d014ba91f0b959465a48005e94f530f0374a5Brian Paul 126600989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca /* 126700989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca * Mask out color channels not present in the color buffer. 1268c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 126900989d5bfc29f632886066d048a366ff4d2d03bcJosé Fonseca blend_rt->colormask &= util_format_colormask(format_desc); 127059253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca 127159253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca /* 127259253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * Our swizzled render tiles always have an alpha channel, but the linear 127359253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * render target format often does not, so force here the dst alpha to be 127459253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * one. 127559253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * 127659253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * This is not a mere optimization. Wrong results will be produced if the 127759253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * dst alpha is used, the dst format does not have alpha, and the previous 127859253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * rendering was not flushed from the swizzled to linear buffer. For 127959253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * example, NonPowTwo DCT. 128059253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * 128159253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * TODO: This should be generalized to all channels for better 128259253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca * performance, but only alpha causes correctness issues. 128387267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca * 128487267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca * Also, force rgb/alpha func/factors match, to make AoS blending easier. 128559253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca */ 128659253a79d823a4cfd1d65466c0a165d6d4c43a92José Fonseca if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) { 128787267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor); 128887267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor); 128987267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca blend_rt->alpha_func = blend_rt->rgb_func; 129087267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca blend_rt->alpha_src_factor = blend_rt->rgb_src_factor; 129187267c71f67d02fcdd59a899fd0eea6d64e523b5José Fonseca blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor; 12924c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 12934c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 1294e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca 12953d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell /* This value will be the same for all the variants of a given shader: 12963d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell */ 1297986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1; 12983d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell 12993d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell for(i = 0; i < key->nr_samplers; ++i) { 1300986cb9d5cf60bc11c7facc19017b5432b17240f7José Fonseca if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) { 13013d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell lp_sampler_static_state(&key->sampler[i], 13023d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell lp->fragment_sampler_views[i], 13033d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell lp->sampler[i]); 13043d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell } 13053d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell } 1306635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca} 1307635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 1308efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1309efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 13104769328fe1ddaa1882dddbaad21239d5fdcddf19Brian Paul/** 1311efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * Update fragment shader state. This is called just prior to drawing 13124769328fe1ddaa1882dddbaad21239d5fdcddf19Brian Paul * something when some fragment-related state has changed. 13134769328fe1ddaa1882dddbaad21239d5fdcddf19Brian Paul */ 13145811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonsecavoid 13155811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonsecallvmpipe_update_fs(struct llvmpipe_context *lp) 13169ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca{ 13179ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader *shader = lp->fs; 1318e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca struct lp_fragment_shader_variant_key key; 131917c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger struct lp_fragment_shader_variant *variant = NULL; 132017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger struct lp_fs_variant_list_item *li; 13219ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1322e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca make_variant_key(lp, shader, &key); 1323e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 1324efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Search the variants for one which matches the key */ 132517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger li = first_elem(&shader->variants); 132617c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger while(!at_end(&shader->variants, li)) { 13273d4b60f1f7be3dc54951c9c414601062e73ca674Keith Whitwell if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) { 132817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger variant = li->base; 13299ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca break; 133017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 133117c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger li = next_elem(li); 13329ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca } 13339ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 133417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (variant) { 1335efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Move this variant to the head of the list to implement LRU 1336efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * deletion of shader's when we have too many. 1337efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul */ 133817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger move_to_head(&lp->fs_variants_list, &variant->list_item_global); 133917c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 134017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger else { 1341efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* variant not found, create it now */ 1342efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul int64_t t0, t1, dt; 134317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger unsigned i; 1344efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1345efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* First, check if we've exceeded the max number of shader variants. 1346efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * If so, free 25% of them (the least recently used ones). 1347efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul */ 134817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) { 134917c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger struct pipe_context *pipe = &lp->pipe; 135017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger 135117c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger /* 1352efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * XXX: we need to flush the context until we have some sort of 1353efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * reference counting in fragment shaders as they may still be binned 135417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger * Flushing alone might not be sufficient we need to wait on it too. 135517c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger */ 1356b6e03eafe3311142445ca42c1574d3f6998eecc3Keith Whitwell llvmpipe_finish(pipe, __FUNCTION__); 135717c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger 135817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) { 1359efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul struct lp_fs_variant_list_item *item; 1360efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul item = last_elem(&lp->fs_variants_list); 1361efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul llvmpipe_remove_shader_variant(lp, item->base); 136217c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 136317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 13649ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1365efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* 1366efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul * Generate the new variant. 1367efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul */ 1368efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul t0 = os_time_get(); 1369efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul variant = generate_variant(lp, shader, &key); 1370a8270c7c29b141025f606c811812304a72e42d2eBrian Paul t1 = os_time_get(); 1371a8270c7c29b141025f606c811812304a72e42d2eBrian Paul dt = t1 - t0; 1372e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul LP_COUNT_ADD(llvm_compile_time, dt); 1373e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */ 137417c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger 1375efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul llvmpipe_variant_count++; 1376efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul 1377efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Put the new variant into the list */ 137817c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger if (variant) { 137917c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger insert_at_head(&shader->variants, &variant->list_item_local); 138017c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger insert_at_head(&lp->fs_variants_list, &variant->list_item_global); 138117c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger lp->nr_fs_variants++; 138217c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger shader->variants_cached++; 138317c9d7eea7b3365c59455a731fcb81e69bb86ce2Roland Scheidegger } 1384e95ad2a2b521514eaec04f9b266ee030ecc639a3Brian Paul } 13859ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1386efc82aef35a2aac5d2ed9774f6d28f2626796416Brian Paul /* Bind this variant */ 138718fb9ff6d8f58a08e559070cf29f26ed0caa567fJosé Fonseca lp_setup_set_fs_variant(lp->setup, variant); 13889ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca} 13899523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul 13909523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul 13919523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul 13925b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 13935b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 13945b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 13955b4c43d98556c5a4806757513bcb196a724518c5Keith Whitwell 13969523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulvoid 13979523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paulllvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe) 13989523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul{ 13999523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state; 14009523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state; 14019523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state; 14029523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul 14039523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer; 14049523d78dde7f2e819275ecb39cdeafe50eb65f13Brian Paul} 1405