lp_state_fs.c revision 7df4c88088ecf34764c558b4f7fe7ef6c82327bb
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 * - stipple (TBI) 355811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - early depth test 365811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - fragment shader 375811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - alpha test 385811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - depth/stencil test (stencil TBI) 395811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * - blending 405811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 415811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * This file has only the glue to assembly the fragment pipeline. The actual 425811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the 435811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we 445811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * muster the LLVM JIT execution engine to create a function that follows an 455811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * established binary interface and that can be called from C directly. 465811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 475811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * A big source of complexity here is that we often want to run different 485811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * stages with different precisions and data types and precisions. For example, 495811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * the fragment shader needs typically to be done in floats, but the 505811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * depth/stencil test and blending is better done in the type that most closely 515811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * matches the depth/stencil and color buffer respectively. 525811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 535811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Since the width of a SIMD vector register stays the same regardless of the 545811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * element type, different types imply different number of elements, so we must 555811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * code generate more instances of the stages with larger types to be able to 565811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * feed/consume the stages with smaller types. 575811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * 585811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * @author Jose Fonseca <jfonseca@vmware.com> 595811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 605811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 61f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#include <limits.h> 62946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "pipe/p_defines.h" 63946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "util/u_memory.h" 64e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca#include "util/u_format.h" 659ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca#include "util/u_debug_dump.h" 66946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "pipe/internal/p_winsys_screen.h" 67946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "pipe/p_shader_tokens.h" 68946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "draw/draw_context.h" 69946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_dump.h" 70946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_scan.h" 71946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca#include "tgsi/tgsi_parse.h" 7273af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_bld_type.h" 73f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca#include "lp_bld_const.h" 74e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca#include "lp_bld_conv.h" 75c40eddd294abfe8af3619d08ccd7e9c8f1660fcbJosé Fonseca#include "lp_bld_intr.h" 7698971802798354cdba45c421cc340ec938143e03José Fonseca#include "lp_bld_logic.h" 77e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca#include "lp_bld_depth.h" 78f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca#include "lp_bld_interp.h" 7973af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_bld_tgsi.h" 809ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca#include "lp_bld_alpha.h" 8198971802798354cdba45c421cc340ec938143e03José Fonseca#include "lp_bld_blend.h" 827c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca#include "lp_bld_swizzle.h" 833d7a88674f9eb3320eeff511968f041426e25023José Fonseca#include "lp_bld_flow.h" 84818d444e12bb57568fbf3bf5f06ee24c6c73a61aJosé Fonseca#include "lp_bld_debug.h" 8573af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_screen.h" 8673af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_context.h" 87c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca#include "lp_buffer.h" 8885999695829823e459e11822b4846ed1db5c055dJosé Fonseca#include "lp_setup.h" 8973af91e938eb27b001404f11195fb06ff9b08903José Fonseca#include "lp_state.h" 908be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca#include "lp_tex_sample.h" 91dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell#include "lp_debug.h" 9273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 9373af91e938eb27b001404f11195fb06ff9b08903José Fonseca 947c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonsecastatic const unsigned char quad_offset_x[4] = {0, 1, 0, 1}; 957c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonsecastatic const unsigned char quad_offset_y[4] = {0, 0, 1, 1}; 967c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 977c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 98f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca/* 99f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca * Derive from the quad's upper left scalar coordinates the coordinates for 100f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca * all other quad pixels 1015811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 1027c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonsecastatic void 103f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonsecagenerate_pos0(LLVMBuilderRef builder, 104f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef x, 105f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef y, 106f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef *x0, 107f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef *y0) 1087c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca{ 1097c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMTypeRef int_elem_type = LLVMInt32Type(); 1107c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE); 1117c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMTypeRef elem_type = LLVMFloatType(); 1127c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE); 1137c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef x_offsets[QUAD_SIZE]; 1147c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef y_offsets[QUAD_SIZE]; 1157c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca unsigned i; 1167c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 1177c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca x = lp_build_broadcast(builder, int_vec_type, x); 1187c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca y = lp_build_broadcast(builder, int_vec_type, y); 1197c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 1207c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca for(i = 0; i < QUAD_SIZE; ++i) { 1217c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0); 1227c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0); 1237c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca } 1247c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 1257c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), ""); 1267c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), ""); 1277c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 128f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca *x0 = LLVMBuildSIToFP(builder, x, vec_type, ""); 129f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca *y0 = LLVMBuildSIToFP(builder, y, vec_type, ""); 1307c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca} 1317c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 1327c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 1335811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca/** 1345811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the depth test. 1355811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 136e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonsecastatic void 137cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonsecagenerate_depth(LLVMBuilderRef builder, 138cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca const struct lp_fragment_shader_variant_key *key, 139b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type src_type, 14098971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_build_mask_context *mask, 14198971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef src, 14298971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef dst_ptr) 143e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca{ 144e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca const struct util_format_description *format_desc; 145b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type dst_type; 146e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 147cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca if(!key->depth.enabled) 148e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca return; 149e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 150cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca format_desc = util_format_description(key->zsbuf_format); 151e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca assert(format_desc); 152e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 1532282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca /* 1542282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca * Depths are expected to be between 0 and 1, even if they are stored in 1552282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca * floats. Setting these bits here will ensure that the lp_build_conv() call 1562282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca * below won't try to unnecessarily clamp the incoming values. 1572282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca */ 1582282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca if(src_type.floating) { 1592282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca src_type.sign = FALSE; 1602282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca src_type.norm = TRUE; 1612282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca } 1622282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca else { 1632282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca assert(!src_type.sign); 1642282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca assert(src_type.norm); 1652282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca } 1662282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca 1675811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* Pick the depth type. */ 168e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca dst_type = lp_depth_type(format_desc, src_type.width*src_type.length); 169e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 1705811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* FIXME: Cope with a depth test type with a different bit width. */ 171e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca assert(dst_type.width == src_type.width); 172e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca assert(dst_type.length == src_type.length); 173e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 174e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1); 1752282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca 1762282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca dst_ptr = LLVMBuildBitCast(builder, 1772282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca dst_ptr, 1782282fb7710d386bd10ccdd18f030069fae0a5d55José Fonseca LLVMPointerType(lp_build_vec_type(dst_type), 0), ""); 179e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 180e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca lp_build_depth_test(builder, 181cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca &key->depth, 182e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca dst_type, 183e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca format_desc, 184e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca mask, 185e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca src, 186e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca dst_ptr); 187e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca} 188e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 189e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 19098971802798354cdba45c421cc340ec938143e03José Fonseca/** 191ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * Generate the code to do inside/outside triangle testing for the 192ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * four pixels in a 2x2 quad. This will set the four elements of the 193ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * quad mask vector to 0 or ~0. 194ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * \param i which quad of the quad group to test, in [0,3] 195ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul */ 196ab9438193083b7f9a3180cb9cea45e269131048aBrian Paulstatic void 197ab9438193083b7f9a3180cb9cea45e269131048aBrian Paulgenerate_tri_edge_mask(LLVMBuilderRef builder, 198ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul unsigned i, 199ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef *mask, /* ivec4, out */ 200ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c0, /* int32 */ 201ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c1, /* int32 */ 202ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c2, /* int32 */ 203ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step0_ptr, /* ivec4 */ 204ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step1_ptr, /* ivec4 */ 205ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step2_ptr) /* ivec4 */ 206ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul{ 2074061ca02dd837950201a9ada462f944ae25deeb5Brian Paul#define OPTIMIZE_IN_OUT_TEST 0 2084061ca02dd837950201a9ada462f944ae25deeb5Brian Paul#if OPTIMIZE_IN_OUT_TEST 209f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul struct lp_build_if_state ifctx; 2104061ca02dd837950201a9ada462f944ae25deeb5Brian Paul LLVMValueRef not_draw_all; 2114061ca02dd837950201a9ada462f944ae25deeb5Brian Paul#endif 2124061ca02dd837950201a9ada462f944ae25deeb5Brian Paul struct lp_build_flow_context *flow; 213ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul struct lp_type i32_type; 214f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMTypeRef i32vec4_type, mask_type; 215ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c0_vec, c1_vec, c2_vec; 216f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMValueRef in_out_mask; 217ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 218ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul assert(i < 4); 219ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 220ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul /* int32 vector type */ 221ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul memset(&i32_type, 0, sizeof i32_type); 222ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32_type.floating = FALSE; /* values are integers */ 223ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32_type.sign = TRUE; /* values are signed */ 224ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32_type.norm = FALSE; /* values are not normalized */ 225ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32_type.width = 32; /* 32-bit int values */ 226ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32_type.length = 4; /* 4 elements per vector */ 227ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 228ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul i32vec4_type = lp_build_int32_vec4_type(); 229ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 230f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul mask_type = LLVMIntType(32 * 4); 231f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul 232f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* 233f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul * Use a conditional here to do detailed pixel in/out testing. 2349a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul * We only have to do this if c0 != INT_MIN. 235f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul */ 236f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul flow = lp_build_flow_create(builder); 237f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_flow_scope_begin(flow); 238ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 23946b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul { 240f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#if OPTIMIZE_IN_OUT_TEST 2419a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* not_draw_all = (c0 != INT_MIN) */ 242f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul not_draw_all = LLVMBuildICmp(builder, 24346b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul LLVMIntNE, 24446b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul c0, 24546b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul LLVMConstInt(LLVMInt32Type(), INT_MIN, 0), 246f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul ""); 247ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 24846b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul in_out_mask = lp_build_int_const_scalar(i32_type, ~0); 24946b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul 25046b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul 251f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_flow_scope_declare(flow, &in_out_mask); 252ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 2539a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* if (not_draw_all) {... */ 254f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_if(&ifctx, flow, builder, not_draw_all); 255f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#endif 256f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul { 257f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMValueRef step0_vec, step1_vec, step2_vec; 258f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMValueRef m0_vec, m1_vec, m2_vec; 259f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMValueRef index, m; 260f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul 26146b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul /* c0_vec = {c0, c0, c0, c0} 26246b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul * Note that we emit this code four times but LLVM optimizes away 26346b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul * three instances of it. 26446b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul */ 26546b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul c0_vec = lp_build_broadcast(builder, i32vec4_type, c0); 26646b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul c1_vec = lp_build_broadcast(builder, i32vec4_type, c1); 26746b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul c2_vec = lp_build_broadcast(builder, i32vec4_type, c2); 26846b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul lp_build_name(c0_vec, "edgeconst0vec"); 26946b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul lp_build_name(c1_vec, "edgeconst1vec"); 27046b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul lp_build_name(c2_vec, "edgeconst2vec"); 27146b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul 2729a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* load step0vec, step1, step2 vec from memory */ 273f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul index = LLVMConstInt(LLVMInt32Type(), i, 0); 274f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), ""); 275f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), ""); 276f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), ""); 277f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_name(step0_vec, "step0vec"); 278f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_name(step1_vec, "step1vec"); 279f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_name(step2_vec, "step2vec"); 280f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul 2819a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* m0_vec = step0_ptr[i] > c0_vec */ 282f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec); 283f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec); 284f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec); 285f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul 2869a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* in_out_mask = m0_vec & m1_vec & m2_vec */ 287f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul m = LLVMBuildAnd(builder, m0_vec, m1_vec, ""); 288f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul in_out_mask = LLVMBuildAnd(builder, m, m2_vec, ""); 289f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_name(in_out_mask, "inoutmaskvec"); 290f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul } 291f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#if OPTIMIZE_IN_OUT_TEST 292f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_endif(&ifctx); 293f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul#endif 2945771f3d483e882d9f5b6c5f3bdb3c39696623b66Brian Paul 29546b5bd6cadd13f47c10aafe9194c90234db91a2aBrian Paul } 296f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_flow_scope_end(flow); 297f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul lp_build_flow_destroy(flow); 2985771f3d483e882d9f5b6c5f3bdb3c39696623b66Brian Paul 2999a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul /* This is the initial alive/dead pixel mask for a quad of four pixels. 3009a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul * It's an int[4] vector with each word set to 0 or ~0. 3019a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul * Words will get cleared when pixels faile the Z test, etc. 3029a10d14a441ca76c5c9ea8986a2eb4b5923a9b9fBrian Paul */ 303f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul *mask = in_out_mask; 304ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul} 305ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 306ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 307ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul/** 3085811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the fragment shader, depth/stencil test, and alpha tests. 309ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * \param i which quad in the tile, in range [0,3] 31098971802798354cdba45c421cc340ec938143e03José Fonseca */ 31198971802798354cdba45c421cc340ec938143e03José Fonsecastatic void 31298971802798354cdba45c421cc340ec938143e03José Fonsecagenerate_fs(struct llvmpipe_context *lp, 31398971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_fragment_shader *shader, 31498971802798354cdba45c421cc340ec938143e03José Fonseca const struct lp_fragment_shader_variant_key *key, 31598971802798354cdba45c421cc340ec938143e03José Fonseca LLVMBuilderRef builder, 316b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type type, 317635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef context_ptr, 31898971802798354cdba45c421cc340ec938143e03José Fonseca unsigned i, 319f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca const struct lp_build_interp_soa_context *interp, 3208be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca struct lp_build_sampler_soa *sampler, 32198971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef *pmask, 322c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef (*color)[4], 323ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef depth_ptr, 324ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c0, 325ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c1, 326ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c2, 327ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step0_ptr, 328ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step1_ptr, 329ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef step2_ptr) 33073af91e938eb27b001404f11195fb06ff9b08903José Fonseca{ 33173af91e938eb27b001404f11195fb06ff9b08903José Fonseca const struct tgsi_token *tokens = shader->base.tokens; 33273af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMTypeRef elem_type; 33373af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMTypeRef vec_type; 3347821664b15501b173b2304bbada758c33c5ff972José Fonseca LLVMTypeRef int_vec_type; 335635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef consts_ptr; 33698971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS]; 337f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef z = interp->pos[2]; 3388e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca struct lp_build_flow_context *flow; 33998971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_build_mask_context mask; 34098971802798354cdba45c421cc340ec938143e03José Fonseca boolean early_depth_test; 34198971802798354cdba45c421cc340ec938143e03José Fonseca unsigned attrib; 34298971802798354cdba45c421cc340ec938143e03José Fonseca unsigned chan; 343c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell unsigned cbuf; 34498971802798354cdba45c421cc340ec938143e03José Fonseca 345ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul assert(i < 4); 346ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 34798971802798354cdba45c421cc340ec938143e03José Fonseca elem_type = lp_build_elem_type(type); 34898971802798354cdba45c421cc340ec938143e03José Fonseca vec_type = lp_build_vec_type(type); 34998971802798354cdba45c421cc340ec938143e03José Fonseca int_vec_type = lp_build_int_vec_type(type); 35098971802798354cdba45c421cc340ec938143e03José Fonseca 351635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca consts_ptr = lp_jit_context_constants(builder, context_ptr); 352635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 3538e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca flow = lp_build_flow_create(builder); 3548e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 3558e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca memset(outputs, 0, sizeof outputs); 3568e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 3578e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca lp_build_flow_scope_begin(flow); 3588e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 3598e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca /* Declare the color and z variables */ 360c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 361c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) { 362c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color[cbuf][chan] = LLVMGetUndef(vec_type); 363c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_flow_scope_declare(flow, &color[cbuf][chan]); 364c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 3658e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca } 3666a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca lp_build_flow_scope_declare(flow, &z); 3678e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca 368ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul /* do triangle edge testing */ 369ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul generate_tri_edge_mask(builder, i, pmask, 370ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul c0, c1, c2, step0_ptr, step1_ptr, step2_ptr); 371ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 3725771f3d483e882d9f5b6c5f3bdb3c39696623b66Brian Paul /* 'mask' will control execution based on quad's pixel alive/killed state */ 3738e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca lp_build_mask_begin(&mask, flow, type, *pmask); 37498971802798354cdba45c421cc340ec938143e03José Fonseca 375ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul 37698971802798354cdba45c421cc340ec938143e03José Fonseca early_depth_test = 377cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca key->depth.enabled && 378cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca !key->alpha.enabled && 379cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca !shader->info.uses_kill && 380cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca !shader->info.writes_z; 38198971802798354cdba45c421cc340ec938143e03José Fonseca 38298971802798354cdba45c421cc340ec938143e03José Fonseca if(early_depth_test) 383cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca generate_depth(builder, key, 384f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca type, &mask, 385f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca z, depth_ptr); 38698971802798354cdba45c421cc340ec938143e03José Fonseca 38798971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_tgsi_soa(builder, tokens, type, &mask, 388f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca consts_ptr, interp->pos, interp->inputs, 3898be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca outputs, sampler); 39098971802798354cdba45c421cc340ec938143e03José Fonseca 39198971802798354cdba45c421cc340ec938143e03José Fonseca for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) { 39298971802798354cdba45c421cc340ec938143e03José Fonseca for(chan = 0; chan < NUM_CHANNELS; ++chan) { 39398971802798354cdba45c421cc340ec938143e03José Fonseca if(outputs[attrib][chan]) { 39498971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]); 39598971802798354cdba45c421cc340ec938143e03José Fonseca 39698971802798354cdba45c421cc340ec938143e03José Fonseca switch (shader->info.output_semantic_name[attrib]) { 39798971802798354cdba45c421cc340ec938143e03José Fonseca case TGSI_SEMANTIC_COLOR: 39898971802798354cdba45c421cc340ec938143e03José Fonseca { 39998971802798354cdba45c421cc340ec938143e03José Fonseca unsigned cbuf = shader->info.output_semantic_index[attrib]; 40098971802798354cdba45c421cc340ec938143e03José Fonseca 40198971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]); 40298971802798354cdba45c421cc340ec938143e03José Fonseca 40398971802798354cdba45c421cc340ec938143e03José Fonseca /* Alpha test */ 40498971802798354cdba45c421cc340ec938143e03José Fonseca /* XXX: should the alpha reference value be passed separately? */ 405c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* XXX: should only test the final assignment to alpha */ 406635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca if(cbuf == 0 && chan == 3) { 407635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef alpha = outputs[attrib][chan]; 408635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef alpha_ref_value; 409635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr); 410635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value); 41198971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_alpha_test(builder, &key->alpha, type, 412635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca &mask, alpha, alpha_ref_value); 413635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca } 41498971802798354cdba45c421cc340ec938143e03José Fonseca 415c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color[cbuf][chan] = outputs[attrib][chan]; 41698971802798354cdba45c421cc340ec938143e03José Fonseca break; 41798971802798354cdba45c421cc340ec938143e03José Fonseca } 41898971802798354cdba45c421cc340ec938143e03José Fonseca 41998971802798354cdba45c421cc340ec938143e03José Fonseca case TGSI_SEMANTIC_POSITION: 42098971802798354cdba45c421cc340ec938143e03José Fonseca if(chan == 2) 421f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca z = outputs[attrib][chan]; 42298971802798354cdba45c421cc340ec938143e03José Fonseca break; 42398971802798354cdba45c421cc340ec938143e03José Fonseca } 42498971802798354cdba45c421cc340ec938143e03José Fonseca } 42598971802798354cdba45c421cc340ec938143e03José Fonseca } 42698971802798354cdba45c421cc340ec938143e03José Fonseca } 42798971802798354cdba45c421cc340ec938143e03José Fonseca 4286a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca if(!early_depth_test) 4296a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca generate_depth(builder, key, 4306a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca type, &mask, 4316a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca z, depth_ptr); 4326a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 4336a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca lp_build_mask_end(&mask); 4346a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 4356a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca lp_build_flow_scope_end(flow); 4366a405b4a21ac1fa45a93da37ce6b95d98d17f0e2José Fonseca 4378e6b925d2a963a2d5a403e106d7d25e3dcca0775José Fonseca lp_build_flow_destroy(flow); 43898971802798354cdba45c421cc340ec938143e03José Fonseca 43998971802798354cdba45c421cc340ec938143e03José Fonseca *pmask = mask.value; 44098971802798354cdba45c421cc340ec938143e03José Fonseca 44198971802798354cdba45c421cc340ec938143e03José Fonseca} 44298971802798354cdba45c421cc340ec938143e03José Fonseca 44398971802798354cdba45c421cc340ec938143e03José Fonseca 44498971802798354cdba45c421cc340ec938143e03José Fonseca/** 4455811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate color blending and color output. 44698971802798354cdba45c421cc340ec938143e03José Fonseca */ 44798971802798354cdba45c421cc340ec938143e03José Fonsecastatic void 44898971802798354cdba45c421cc340ec938143e03José Fonsecagenerate_blend(const struct pipe_blend_state *blend, 44998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMBuilderRef builder, 450b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type type, 451635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef context_ptr, 45298971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef mask, 45398971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef *src, 45498971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef dst_ptr) 45598971802798354cdba45c421cc340ec938143e03José Fonseca{ 45698971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_build_context bld; 457c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca struct lp_build_flow_context *flow; 458c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca struct lp_build_mask_context mask_ctx; 45998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef vec_type; 46098971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef int_vec_type; 461635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMValueRef const_ptr; 46298971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef con[4]; 46398971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef dst[4]; 46498971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef res[4]; 46598971802798354cdba45c421cc340ec938143e03José Fonseca unsigned chan; 46698971802798354cdba45c421cc340ec938143e03José Fonseca 467c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca lp_build_context_init(&bld, builder, type); 468c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 469c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca flow = lp_build_flow_create(builder); 470f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul 471f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* we'll use this mask context to skip blending if all pixels are dead */ 472c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca lp_build_mask_begin(&mask_ctx, flow, type, mask); 473c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 47498971802798354cdba45c421cc340ec938143e03José Fonseca vec_type = lp_build_vec_type(type); 47598971802798354cdba45c421cc340ec938143e03José Fonseca int_vec_type = lp_build_int_vec_type(type); 47698971802798354cdba45c421cc340ec938143e03José Fonseca 477635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca const_ptr = lp_jit_context_blend_color(builder, context_ptr); 478635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca const_ptr = LLVMBuildBitCast(builder, const_ptr, 479635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca LLVMPointerType(vec_type, 0), ""); 480635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 48198971802798354cdba45c421cc340ec938143e03José Fonseca for(chan = 0; chan < 4; ++chan) { 48298971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); 483635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), ""); 48498971802798354cdba45c421cc340ec938143e03José Fonseca 48598971802798354cdba45c421cc340ec938143e03José Fonseca dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), ""); 48698971802798354cdba45c421cc340ec938143e03José Fonseca 48798971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(con[chan], "con.%c", "rgba"[chan]); 48898971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_name(dst[chan], "dst.%c", "rgba"[chan]); 48998971802798354cdba45c421cc340ec938143e03José Fonseca } 49098971802798354cdba45c421cc340ec938143e03José Fonseca 49198971802798354cdba45c421cc340ec938143e03José Fonseca lp_build_blend_soa(builder, blend, type, src, dst, con, res); 49298971802798354cdba45c421cc340ec938143e03José Fonseca 49398971802798354cdba45c421cc340ec938143e03José Fonseca for(chan = 0; chan < 4; ++chan) { 4944c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca if(blend->colormask & (1 << chan)) { 4954c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); 4964c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca lp_build_name(res[chan], "res.%c", "rgba"[chan]); 4974c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); 4984c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); 4994c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 50098971802798354cdba45c421cc340ec938143e03José Fonseca } 501c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca 502c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca lp_build_mask_end(&mask_ctx); 503c3c80c5c22f9ce7fabba90daa5d5142e5fb1c012José Fonseca lp_build_flow_destroy(flow); 50498971802798354cdba45c421cc340ec938143e03José Fonseca} 50598971802798354cdba45c421cc340ec938143e03José Fonseca 50698971802798354cdba45c421cc340ec938143e03José Fonseca 5075811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca/** 5085811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the runtime callable function for the whole fragment pipeline. 509866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * Note that the function which we generate operates on a block of 16 510866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * pixels at at time. The block contains 2x2 quads. Each quad contains 511866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul * 2x2 pixels. 5125811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 5133b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paulstatic void 51498971802798354cdba45c421cc340ec938143e03José Fonsecagenerate_fragment(struct llvmpipe_context *lp, 51598971802798354cdba45c421cc340ec938143e03José Fonseca struct lp_fragment_shader *shader, 5163b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul struct lp_fragment_shader_variant *variant) 51798971802798354cdba45c421cc340ec938143e03José Fonseca{ 51898971802798354cdba45c421cc340ec938143e03José Fonseca struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen); 5193b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul const struct lp_fragment_shader_variant_key *key = &variant->key; 520b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type fs_type; 521b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca struct lp_type blend_type; 52298971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef fs_elem_type; 52398971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef fs_vec_type; 52498971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef fs_int_vec_type; 52598971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef blend_vec_type; 52698971802798354cdba45c421cc340ec938143e03José Fonseca LLVMTypeRef blend_int_vec_type; 527ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMTypeRef arg_types[14]; 52873af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMTypeRef func_type; 529ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type(); 530c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca LLVMValueRef context_ptr; 5317c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef x; 5327c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca LLVMValueRef y; 53373af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef a0_ptr; 53473af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef dadx_ptr; 53573af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMValueRef dady_ptr; 536c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef color_ptr_ptr; 53739352b34438affc36ce22852bb55359d304a522cJosé Fonseca LLVMValueRef depth_ptr; 538ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr; 53973af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMBasicBlockRef block; 54073af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMBuilderRef builder; 541f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef x0; 542f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca LLVMValueRef y0; 5438be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca struct lp_build_sampler_soa *sampler; 544f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca struct lp_build_interp_soa_context interp; 54598971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH]; 546c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH]; 54798971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef blend_mask; 54898971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef blend_in_color[NUM_CHANNELS]; 54998971802798354cdba45c421cc340ec938143e03José Fonseca unsigned num_fs; 55039352b34438affc36ce22852bb55359d304a522cJosé Fonseca unsigned i; 55139352b34438affc36ce22852bb55359d304a522cJosé Fonseca unsigned chan; 552c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell unsigned cbuf; 55373af91e938eb27b001404f11195fb06ff9b08903José Fonseca 5549ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 5555811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* TODO: actually pick these based on the fs and color buffer 5565811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * characteristics. */ 5575811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 558b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca memset(&fs_type, 0, sizeof fs_type); 55998971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.floating = TRUE; /* floating point values */ 56098971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.sign = TRUE; /* values are signed */ 56198971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ 56298971802798354cdba45c421cc340ec938143e03José Fonseca fs_type.width = 32; /* 32-bit float */ 563866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul fs_type.length = 4; /* 4 elements per vector */ 564866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul num_fs = 4; /* number of quads per block */ 56573af91e938eb27b001404f11195fb06ff9b08903José Fonseca 566b4835ea03d64261da5a892f9590c9977b06920e8José Fonseca memset(&blend_type, 0, sizeof blend_type); 56798971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.floating = FALSE; /* values are integers */ 56898971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.sign = FALSE; /* values are unsigned */ 56998971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */ 57098971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.width = 8; /* 8-bit ubyte values */ 57198971802798354cdba45c421cc340ec938143e03José Fonseca blend_type.length = 16; /* 16 elements per vector */ 57298971802798354cdba45c421cc340ec938143e03José Fonseca 5735811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* 5745811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Generate the function prototype. Any change here must be reflected in 57508dd41fd6899bd6b3289d30dc31f8b2998406889José Fonseca * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa. 5765811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 5775811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 57898971802798354cdba45c421cc340ec938143e03José Fonseca fs_elem_type = lp_build_elem_type(fs_type); 57998971802798354cdba45c421cc340ec938143e03José Fonseca fs_vec_type = lp_build_vec_type(fs_type); 58098971802798354cdba45c421cc340ec938143e03José Fonseca fs_int_vec_type = lp_build_int_vec_type(fs_type); 58198971802798354cdba45c421cc340ec938143e03José Fonseca 58298971802798354cdba45c421cc340ec938143e03José Fonseca blend_vec_type = lp_build_vec_type(blend_type); 58398971802798354cdba45c421cc340ec938143e03José Fonseca blend_int_vec_type = lp_build_int_vec_type(blend_type); 58473af91e938eb27b001404f11195fb06ff9b08903José Fonseca 585c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[0] = screen->context_ptr_type; /* context */ 586c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[1] = LLVMInt32Type(); /* x */ 587c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[2] = LLVMInt32Type(); /* y */ 588c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */ 589c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */ 590c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */ 591c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */ 592ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */ 593ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[8] = LLVMInt32Type(); /* c0 */ 594ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[9] = LLVMInt32Type(); /* c1 */ 595ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[10] = LLVMInt32Type(); /* c2 */ 596ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul /* Note: the step arrays are built as int32[16] but we interpret 597ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul * them here as int32_vec4[4]. 598ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul */ 599ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */ 600ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */ 601ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */ 60273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 60373af91e938eb27b001404f11195fb06ff9b08903José Fonseca func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0); 60473af91e938eb27b001404f11195fb06ff9b08903José Fonseca 6059ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant->function = LLVMAddFunction(screen->module, "shader", func_type); 6069ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca LLVMSetFunctionCallConv(variant->function, LLVMCCallConv); 607c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 608c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* XXX: need to propagate noalias down into color param now we are 609c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * passing a pointer-to-pointer? 610c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 611a7f9b915aeade96d8ca58d70a1f1d42cd6a16e3dJosé Fonseca for(i = 0; i < Elements(arg_types); ++i) 6127c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) 6139ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute); 6149ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 615c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca context_ptr = LLVMGetParam(variant->function, 0); 616c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca x = LLVMGetParam(variant->function, 1); 617c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca y = LLVMGetParam(variant->function, 2); 618c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca a0_ptr = LLVMGetParam(variant->function, 3); 619c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca dadx_ptr = LLVMGetParam(variant->function, 4); 620c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca dady_ptr = LLVMGetParam(variant->function, 5); 621c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color_ptr_ptr = LLVMGetParam(variant->function, 6); 622ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul depth_ptr = LLVMGetParam(variant->function, 7); 623ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul c0 = LLVMGetParam(variant->function, 8); 624ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul c1 = LLVMGetParam(variant->function, 9); 625ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul c2 = LLVMGetParam(variant->function, 10); 626ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul step0_ptr = LLVMGetParam(variant->function, 11); 627ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul step1_ptr = LLVMGetParam(variant->function, 12); 628ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul step2_ptr = LLVMGetParam(variant->function, 13); 6297c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca 630c022e15d1e56ba3a9c6b74eef6556d6063e2e322José Fonseca lp_build_name(context_ptr, "context"); 6317c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca lp_build_name(x, "x"); 6327c2dc3faa8e8ba99e29fcc09e7db39fc49d73200José Fonseca lp_build_name(y, "y"); 6335999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(a0_ptr, "a0"); 6345999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(dadx_ptr, "dadx"); 6355999ebfb69a47fa12d1f534871ea01a3f889f62fJosé Fonseca lp_build_name(dady_ptr, "dady"); 636c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_name(color_ptr_ptr, "color_ptr"); 63739352b34438affc36ce22852bb55359d304a522cJosé Fonseca lp_build_name(depth_ptr, "depth"); 638ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(c0, "c0"); 639ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(c1, "c1"); 640ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(c2, "c2"); 641ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(step0_ptr, "step0"); 642ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(step1_ptr, "step1"); 643ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul lp_build_name(step2_ptr, "step2"); 64473af91e938eb27b001404f11195fb06ff9b08903José Fonseca 6455811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca /* 6465811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca * Function body 6475811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 6485811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 6499ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca block = LLVMAppendBasicBlock(variant->function, "entry"); 65073af91e938eb27b001404f11195fb06ff9b08903José Fonseca builder = LLVMCreateBuilder(); 65173af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMPositionBuilderAtEnd(builder, block); 65273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 653f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca generate_pos0(builder, x, y, &x0, &y0); 654f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca 65586f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell lp_build_interp_soa_init(&interp, 65686f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell shader->base.tokens, 65786f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell key->flatshade, 65886f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell builder, fs_type, 659f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca a0_ptr, dadx_ptr, dady_ptr, 660866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul x0, y0); 661f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca 662e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca /* code generated texture sampling */ 663e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr); 664c40eddd294abfe8af3619d08ccd7e9c8f1660fcbJosé Fonseca 665866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul /* loop over quads in the block */ 66698971802798354cdba45c421cc340ec938143e03José Fonseca for(i = 0; i < num_fs; ++i) { 66798971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); 668c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS]; 66998971802798354cdba45c421cc340ec938143e03José Fonseca LLVMValueRef depth_ptr_i; 670c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell int cbuf; 67198971802798354cdba45c421cc340ec938143e03José Fonseca 672f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca if(i != 0) 673866e6856d39efe9b1ec739587f420a640ad8618eBrian Paul lp_build_interp_soa_update(&interp, i); 67498971802798354cdba45c421cc340ec938143e03José Fonseca 67598971802798354cdba45c421cc340ec938143e03José Fonseca depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, ""); 67698971802798354cdba45c421cc340ec938143e03José Fonseca 677635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca generate_fs(lp, shader, key, 67898971802798354cdba45c421cc340ec938143e03José Fonseca builder, 67998971802798354cdba45c421cc340ec938143e03José Fonseca fs_type, 680635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca context_ptr, 68198971802798354cdba45c421cc340ec938143e03José Fonseca i, 682f85c5f8621382ba1c8baa1582d87b46b388258d2José Fonseca &interp, 6838be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca sampler, 684ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul &fs_mask[i], /* output */ 68598971802798354cdba45c421cc340ec938143e03José Fonseca out_color, 686ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul depth_ptr_i, 687ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul c0, c1, c2, 688ab9438193083b7f9a3180cb9cea45e269131048aBrian Paul step0_ptr, step1_ptr, step2_ptr); 68998971802798354cdba45c421cc340ec938143e03José Fonseca 690c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) 691c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) 692c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell fs_out_color[cbuf][chan][i] = out_color[cbuf][chan]; 69398971802798354cdba45c421cc340ec938143e03José Fonseca } 69439352b34438affc36ce22852bb55359d304a522cJosé Fonseca 6958be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca sampler->destroy(sampler); 6968be72bb7646d430e66cb36e09c13c13bee030d53José Fonseca 697c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* Loop over color outputs / color buffers to do blending. 6985811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca */ 699c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 700c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef color_ptr; 701c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0); 7025811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 703c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* 704c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * Convert the fs's output color and mask to fit to the blending type. 705c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 706c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for(chan = 0; chan < NUM_CHANNELS; ++chan) { 707c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_conv(builder, fs_type, blend_type, 708c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell fs_out_color[cbuf][chan], num_fs, 709c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell &blend_in_color[chan], 1); 710c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]); 711c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 712635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 713c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_conv_mask(builder, fs_type, blend_type, 714c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell fs_mask, num_fs, 715c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell &blend_mask, 1); 716c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 717c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color_ptr = LLVMBuildLoad(builder, 718c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""), 719c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell ""); 720c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell lp_build_name(color_ptr, "color_ptr%d", cbuf); 721c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 722c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* 723c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * Blending. 724c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 725c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell generate_blend(&key->blend, 726c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell builder, 727c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell blend_type, 728c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell context_ptr, 729c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell blend_mask, 730c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell blend_in_color, 731c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell color_ptr); 73273af91e938eb27b001404f11195fb06ff9b08903José Fonseca } 73373af91e938eb27b001404f11195fb06ff9b08903José Fonseca 7345811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca LLVMBuildRetVoid(builder); 73573af91e938eb27b001404f11195fb06ff9b08903José Fonseca 73673af91e938eb27b001404f11195fb06ff9b08903José Fonseca LLVMDisposeBuilder(builder); 737946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 7385811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonseca 739f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Verify the LLVM IR. If invalid, dump and abort */ 7404ae3e88dc9856f2f32c37dd04a3321765ed61e07José Fonseca#ifdef DEBUG 741a02ecdf8c2fc5783a4bc82e8cd9d36f0dec7ccecJosé Fonseca if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) { 742f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul if (1) 743f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMDumpValue(variant->function); 744f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul abort(); 745a02ecdf8c2fc5783a4bc82e8cd9d36f0dec7ccecJosé Fonseca } 7464ae3e88dc9856f2f32c37dd04a3321765ed61e07José Fonseca#endif 747a02ecdf8c2fc5783a4bc82e8cd9d36f0dec7ccecJosé Fonseca 748f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Apply optimizations to LLVM IR */ 749f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul if (1) 750f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul LLVMRunFunctionPassManager(screen->pass, variant->function); 75173af91e938eb27b001404f11195fb06ff9b08903José Fonseca 752dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell if (LP_DEBUG & DEBUG_JIT) { 753f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* Print the LLVM IR to stderr */ 754dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell LLVMDumpValue(variant->function); 755dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell debug_printf("\n"); 756dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell } 75773af91e938eb27b001404f11195fb06ff9b08903José Fonseca 758f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul /* 759f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul * Translate the LLVM IR into machine code. 760f4321fbd961a0a891c7f40b16efc61aa791e03a9Brian Paul */ 76108dd41fd6899bd6b3289d30dc31f8b2998406889José Fonseca variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function); 76273af91e938eb27b001404f11195fb06ff9b08903José Fonseca 763dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell if (LP_DEBUG & DEBUG_ASM) 764dec35d04aeb398eef159aaf8cde5e0d04622b811Keith Whitwell lp_disassemble(variant->jit_function); 765818d444e12bb57568fbf3bf5f06ee24c6c73a61aJosé Fonseca 7669ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant->next = shader->variants; 7679ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader->variants = variant; 7683b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul} 7693b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 7703b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 7713b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paulstatic struct lp_fragment_shader_variant * 7723b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paulgenerate_variant(struct llvmpipe_context *lp, 7733b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul struct lp_fragment_shader *shader, 7743b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul const struct lp_fragment_shader_variant_key *key) 7753b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul{ 7763b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul struct lp_fragment_shader_variant *variant; 7773b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 7783b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if (LP_DEBUG & DEBUG_JIT) { 7793b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul unsigned i; 7803b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 7813b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul tgsi_dump(shader->base.tokens, 0); 7823b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(key->depth.enabled) { 7833b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format)); 7843b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE)); 7853b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("depth.writemask = %u\n", key->depth.writemask); 7863b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 7873b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(key->alpha.enabled) { 7883b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE)); 7893b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value); 7903b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 7913b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(key->blend.logicop_enable) { 7923b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func); 7933b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 7943b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul else if(key->blend.blend_enable) { 7953b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE)); 7963b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE)); 7973b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE)); 7983b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE)); 7993b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE)); 8003b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE)); 8013b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 8023b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("blend.colormask = 0x%x\n", key->blend.colormask); 8033b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) { 8043b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(key->sampler[i].format) { 8053b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf("sampler[%u] = \n", i); 8063b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .format = %s\n", 8073b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul pf_name(key->sampler[i].format)); 8083b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .target = %s\n", 8093b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_target(key->sampler[i].target, TRUE)); 8103b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .pot = %u %u %u\n", 8113b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul key->sampler[i].pot_width, 8123b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul key->sampler[i].pot_height, 8133b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul key->sampler[i].pot_depth); 8143b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .wrap = %s %s %s\n", 8153b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE), 8163b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE), 8173b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE)); 8183b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .min_img_filter = %s\n", 8193b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE)); 8203b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .min_mip_filter = %s\n", 8213b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE)); 8223b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .mag_img_filter = %s\n", 8233b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE)); 8243b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE) 8253b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE)); 8263b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords); 8273b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter); 8283b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 8293b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 8303b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul } 8313b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 8323b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul variant = CALLOC_STRUCT(lp_fragment_shader_variant); 8333b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul if(!variant) 8343b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul return NULL; 8353b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 8363b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul variant->shader = shader; 8373b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul memcpy(&variant->key, key, sizeof *key); 8383b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul 8393b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul generate_fragment(lp, shader, variant); 8409ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8419ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca return variant; 8429ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca} 8439ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8449ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8459ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonsecavoid * 8469ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonsecallvmpipe_create_fs_state(struct pipe_context *pipe, 8479ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca const struct pipe_shader_state *templ) 8489ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca{ 8499ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader *shader; 8509ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8519ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader = CALLOC_STRUCT(lp_fragment_shader); 8529ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca if (!shader) 8539ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca return NULL; 8549ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8559ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca /* get/save the summary info for this shader */ 8569ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca tgsi_scan_shader(templ->tokens, &shader->info); 8579ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8589ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca /* we need to keep a local copy of the tokens */ 8599ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader->base.tokens = tgsi_dup_tokens(templ->tokens); 8609ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 86173af91e938eb27b001404f11195fb06ff9b08903José Fonseca return shader; 862946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 863946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 864946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 865946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecavoid 866946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs) 867946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 868946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 869946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 870080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca if (llvmpipe->fs == fs) 871080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca return; 872080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca 873080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca draw_flush(llvmpipe->draw); 874080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca 875080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca llvmpipe->fs = fs; 876946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 877c9a5930fe45a0a0299769bd2b672ca516d1bf39eJosé Fonseca llvmpipe->dirty |= LP_NEW_FS; 878946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 879946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 880946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 881946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecavoid 882946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs) 883946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 884e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 885e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen); 88673af91e938eb27b001404f11195fb06ff9b08903José Fonseca struct lp_fragment_shader *shader = fs; 8879ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader_variant *variant; 888946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 889e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca assert(fs != llvmpipe->fs); 89031d1822473bf9d4105bb82b67572cfeea53aaf94Vinson Lee (void) llvmpipe; 8919ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8929ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant = shader->variants; 8939ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca while(variant) { 8949ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader_variant *next = variant->next; 8959ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 8969ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca if(variant->function) { 8979ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca if(variant->jit_function) 8989ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca LLVMFreeMachineCodeForFunction(screen->engine, variant->function); 8999ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca LLVMDeleteFunction(variant->function); 9009ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca } 9019ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9029ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca FREE(variant); 9039ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9049ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant = next; 90573af91e938eb27b001404f11195fb06ff9b08903José Fonseca } 90673af91e938eb27b001404f11195fb06ff9b08903José Fonseca 90773af91e938eb27b001404f11195fb06ff9b08903José Fonseca FREE((void *) shader->base.tokens); 90873af91e938eb27b001404f11195fb06ff9b08903José Fonseca FREE(shader); 909946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 910946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 911946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 912946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 913946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecavoid 914946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonsecallvmpipe_set_constant_buffer(struct pipe_context *pipe, 915946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca uint shader, uint index, 916c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca const struct pipe_constant_buffer *constants) 917946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca{ 918946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 919c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca struct pipe_buffer *buffer = constants ? constants->buffer : NULL; 920c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca unsigned size = buffer ? buffer->size : 0; 921c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL; 922946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 923946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca assert(shader < PIPE_SHADER_TYPES); 924946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca assert(index == 0); 925946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 926d904ed88c1d957f662497343de7dc3e9fa743e47José Fonseca if(llvmpipe->constants[shader].buffer == buffer) 927d904ed88c1d957f662497343de7dc3e9fa743e47José Fonseca return; 928d904ed88c1d957f662497343de7dc3e9fa743e47José Fonseca 929080703e398f737b71336312fd3dc8d6f38f61e51José Fonseca draw_flush(llvmpipe->draw); 930c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca 931946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca /* note: reference counting */ 932c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer); 933c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca 934c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca if(shader == PIPE_SHADER_VERTEX) { 93589d8577fb3036547ef0b47498cc8dc5c77f886e0Zack Rusin draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, 93689d8577fb3036547ef0b47498cc8dc5c77f886e0Zack Rusin data, size); 937c4d54b62f5491dbec9930563209639f8fb7dcf2eJosé Fonseca } 938946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca 939c9a5930fe45a0a0299769bd2b672ca516d1bf39eJosé Fonseca llvmpipe->dirty |= LP_NEW_CONSTANTS; 940946f432a08112148d743eb9faf6b27bb8cc7fa76José Fonseca} 9419ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 9429ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 943635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca/** 944635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * We need to generate several variants of the fragment pipeline to match 945635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * all the combinations of the contributing state atoms. 946635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * 947635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * TODO: there is actually no reason to tie this to context state -- the 948635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca * generated code could be cached globally in the screen. 949635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca */ 950635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonsecastatic void 951635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonsecamake_variant_key(struct llvmpipe_context *lp, 952e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca struct lp_fragment_shader *shader, 953635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca struct lp_fragment_shader_variant_key *key) 954635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca{ 955e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca unsigned i; 956e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca 957635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca memset(key, 0, sizeof *key); 958635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 959cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca if(lp->framebuffer.zsbuf && 960cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca lp->depth_stencil->depth.enabled) { 961cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca key->zsbuf_format = lp->framebuffer.zsbuf->format; 962cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); 963cdbbcdf3bdb114d79cf7b9474436c3d26b135592José Fonseca } 964635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 965635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca key->alpha.enabled = lp->depth_stencil->alpha.enabled; 966635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca if(key->alpha.enabled) 967635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca key->alpha.func = lp->depth_stencil->alpha.func; 968635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca /* alpha.ref_value is passed in jit_context */ 969635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 97086f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell key->flatshade = lp->rasterizer->flatshade; 97186f450060debebd66dd5fb72f83800d7634efeaaKeith Whitwell 972c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell if (lp->framebuffer.nr_cbufs) { 973c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell memcpy(&key->blend, lp->blend, sizeof key->blend); 974c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell } 975c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 976c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell key->nr_cbufs = lp->framebuffer.nr_cbufs; 977c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell for (i = 0; i < lp->framebuffer.nr_cbufs; i++) { 9784c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca const struct util_format_description *format_desc; 9794c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca unsigned chan; 9804c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 981c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell format_desc = util_format_description(lp->framebuffer.cbufs[i]->format); 9824c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB || 9834c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB); 9844c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca 985c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell /* mask out color channels not present in the color buffer. 986c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell * Should be simple to incorporate per-cbuf writemasks: 987c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell */ 9884c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca for(chan = 0; chan < 4; ++chan) { 9894c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca enum util_format_swizzle swizzle = format_desc->swizzle[chan]; 990c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell 991c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell if(swizzle <= UTIL_FORMAT_SWIZZLE_W) 992c1a04416023e24621e4992caf593e8dfe8d7a2fcKeith Whitwell key->cbuf_blend[i].colormask |= (1 << chan); 9934c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 9944c3a48ad0cb36e6d8601535b91f83caed0d07570José Fonseca } 995e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca 996e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) 997e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) 998e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]); 999635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca} 1000635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 1001635c37e118bbf51fc8c93a2c999281ee817a93e1José Fonseca 10025811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonsecavoid 10035811ed87d732101ab8cfbd087bc99d8c6c963f30José Fonsecallvmpipe_update_fs(struct llvmpipe_context *lp) 10049ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca{ 10059ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader *shader = lp->fs; 1006e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca struct lp_fragment_shader_variant_key key; 10079ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca struct lp_fragment_shader_variant *variant; 1008a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca boolean opaque; 10099ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 1010e4c76c02f77ed6e86537b546f4200f8f8132d114José Fonseca make_variant_key(lp, shader, &key); 1011e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca 10129ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant = shader->variants; 10139ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca while(variant) { 1014e3b38e5ec1ba93e3f1a1b3d5039c70ff7aa3ebe6José Fonseca if(memcmp(&variant->key, &key, sizeof key) == 0) 10159ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca break; 10169ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 10179ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca variant = variant->next; 10189ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca } 10199ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 10209ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca if(!variant) 10213b5d84926847cf2008da4e2dc146090d0c1b5402Brian Paul variant = generate_variant(lp, shader, &key); 10229ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca 10239ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca shader->current = variant; 102485999695829823e459e11822b4846ed1db5c055dJosé Fonseca 1025a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca /* TODO: put this in the variant */ 10267df4c88088ecf34764c558b4f7fe7ef6c82327bbJosé Fonseca /* TODO: most of these can be relaxed, in particular the colormask */ 1027a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca opaque = !key.blend.logicop_enable && 1028a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca !key.blend.blend_enable && 10297df4c88088ecf34764c558b4f7fe7ef6c82327bbJosé Fonseca key.blend.colormask == 0xf && 1030a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca !key.alpha.enabled && 1031a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca !key.depth.enabled && 1032a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca !shader->info.uses_kill 1033a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca ? TRUE : FALSE; 1034a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca 10355ce0380a0f585b9e1fb616b749f7fd18a8afada1Keith Whitwell lp_setup_set_fs_function(lp->setup, 1036a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca shader->current->jit_function, 1037a1acbff299c444913418e65da473745cd901a2dbJosé Fonseca opaque); 10389ae47069b4a2b67e381b02d805f1ca74f31ea7b8José Fonseca} 1039