Edit
IABSD.fr/xenocara/lib/mesa/src/intel/compiler/elk/elk_schedule_instructions.cpp
jsg
- lib/mesa/src/intel/compiler/elk/elk_schedule_instructions.cpp
-
/* * Copyright © 2010 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. * * Authors: * Eric Anholt <eric@anholt.net> * */ #include "elk_eu.h" #include "elk_fs.h" #include "elk_fs_live_variables.h" #include "elk_vec4.h" #include "elk_cfg.h" #include "elk_shader.h" #include <new> using namespace elk; /** @file elk_fs_schedule_instructions.cpp * * List scheduling of FS instructions. * * The basic model of the list scheduler is to take a basic block, * compute a DAG of the dependencies (RAW ordering with latency, WAW * ordering with latency, WAR ordering), and make a list of the DAG heads. * Heuristically pick a DAG head, then put all the children that are * now DAG heads into the list of things to schedule. * * The heuristic is the important part. We're trying to be cheap, * since actually computing the optimal scheduling is NP complete. * What we do is track a "current clock". When we schedule a node, we * update the earliest-unblocked clock time of its children, and * increment the clock. Then, when trying to schedule, we just pick * the earliest-unblocked instruction to schedule. * * Note that often there will be many things which could execute * immediately, and there are a range of heuristic options to choose * from in picking among those. */ static bool debug = false; class elk_instruction_scheduler; struct elk_schedule_node_child; class elk_schedule_node : public exec_node { public: void set_latency_gfx4(); void set_latency_gfx7(const struct elk_isa_info *isa); elk_backend_instruction *inst; elk_schedule_node_child *children; int children_count; int children_cap; int initial_parent_count; int initial_unblocked_time; int latency; /** * This is the sum of the instruction's latency plus the maximum delay of * its children, or just the issue_time if it's a leaf node. */ int delay; /** * Preferred exit node among the (direct or indirect) successors of this * node. Among the scheduler nodes blocked by this node, this will be the * one that may cause earliest program termination, or NULL if none of the * successors is an exit node. */ elk_schedule_node *exit; /** * How many cycles this instruction takes to issue. * * Instructions in gen hardware are handled one simd4 vector at a time, * with 1 cycle per vector dispatched. Thus SIMD8 pixel shaders take 2 * cycles to dispatch and SIMD16 (compressed) instructions take 4. */ int issue_time; /* Temporary data used during the scheduling process. */ struct { int parent_count; int unblocked_time; /** * Which iteration of pushing groups of children onto the candidates list * this node was a part of. */ unsigned cand_generation; } tmp; }; struct elk_schedule_node_child { elk_schedule_node *n; int effective_latency; }; static inline void reset_node_tmp(elk_schedule_node *n) { n->tmp.parent_count = n->initial_parent_count; n->tmp.unblocked_time = n->initial_unblocked_time; n->tmp.cand_generation = 0; } /** * Lower bound of the scheduling time after which one of the instructions * blocked by this node may lead to program termination. * * exit_unblocked_time() determines a strict partial ordering relation '«' on * the set of scheduler nodes as follows: * * n « m <-> exit_unblocked_time(n) < exit_unblocked_time(m) * * which can be used to heuristically order nodes according to how early they * can unblock an exit node and lead to program termination. */ static inline int exit_tmp_unblocked_time(const elk_schedule_node *n) { return n->exit ? n->exit->tmp.unblocked_time : INT_MAX; } static inline int exit_initial_unblocked_time(const elk_schedule_node *n) { return n->exit ? n->exit->initial_unblocked_time : INT_MAX; } void elk_schedule_node::set_latency_gfx4() { int chans = 8; int math_latency = 22; switch (inst->opcode) { case ELK_SHADER_OPCODE_RCP: this->latency = 1 * chans * math_latency; break; case ELK_SHADER_OPCODE_RSQ: this->latency = 2 * chans * math_latency; break; case ELK_SHADER_OPCODE_INT_QUOTIENT: case ELK_SHADER_OPCODE_SQRT: case ELK_SHADER_OPCODE_LOG2: /* full precision log. partial is 2. */ this->latency = 3 * chans * math_latency; break; case ELK_SHADER_OPCODE_INT_REMAINDER: case ELK_SHADER_OPCODE_EXP2: /* full precision. partial is 3, same throughput. */ this->latency = 4 * chans * math_latency; break; case ELK_SHADER_OPCODE_POW: this->latency = 8 * chans * math_latency; break; case ELK_SHADER_OPCODE_SIN: case ELK_SHADER_OPCODE_COS: /* minimum latency, max is 12 rounds. */ this->latency = 5 * chans * math_latency; break; default: this->latency = 2; break; } } void elk_schedule_node::set_latency_gfx7(const struct elk_isa_info *isa) { const bool is_haswell = isa->devinfo->verx10 == 75; switch (inst->opcode) { case ELK_OPCODE_MAD: /* 2 cycles * (since the last two src operands are in different register banks): * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; * * 3 cycles on IVB, 4 on HSW * (since the last two src operands are in the same register bank): * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; * * 18 cycles on IVB, 16 on HSW * (since the last two src operands are in different register banks): * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; * mov(8) null g4<4,5,1>F { align16 WE_normal 1Q }; * * 20 cycles on IVB, 18 on HSW * (since the last two src operands are in the same register bank): * mad(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; */ /* Our register allocator doesn't know about register banks, so use the * higher latency. */ latency = is_haswell ? 16 : 18; break; case ELK_OPCODE_LRP: /* 2 cycles * (since the last two src operands are in different register banks): * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; * * 3 cycles on IVB, 4 on HSW * (since the last two src operands are in the same register bank): * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; * * 16 cycles on IVB, 14 on HSW * (since the last two src operands are in different register banks): * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g3.1<4,4,1>F.x { align16 WE_normal 1Q }; * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; * * 16 cycles * (since the last two src operands are in the same register bank): * lrp(8) g4<1>F g2.2<4,4,1>F.x g2<4,4,1>F.x g2.1<4,4,1>F.x { align16 WE_normal 1Q }; * mov(8) null g4<4,4,1>F { align16 WE_normal 1Q }; */ /* Our register allocator doesn't know about register banks, so use the * higher latency. */ latency = 14; break; case ELK_SHADER_OPCODE_RCP: case ELK_SHADER_OPCODE_RSQ: case ELK_SHADER_OPCODE_SQRT: case ELK_SHADER_OPCODE_LOG2: case ELK_SHADER_OPCODE_EXP2: case ELK_SHADER_OPCODE_SIN: case ELK_SHADER_OPCODE_COS: /* 2 cycles: * math inv(8) g4<1>F g2<0,1,0>F null { align1 WE_normal 1Q }; * * 18 cycles: * math inv(8) g4<1>F g2<0,1,0>F null { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * Same for exp2, log2, rsq, sqrt, sin, cos. */ latency = is_haswell ? 14 : 16; break; case ELK_SHADER_OPCODE_POW: /* 2 cycles: * math pow(8) g4<1>F g2<0,1,0>F g2.1<0,1,0>F { align1 WE_normal 1Q }; * * 26 cycles: * math pow(8) g4<1>F g2<0,1,0>F g2.1<0,1,0>F { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; */ latency = is_haswell ? 22 : 24; break; case ELK_SHADER_OPCODE_TEX: case ELK_SHADER_OPCODE_TXD: case ELK_SHADER_OPCODE_TXF: case ELK_SHADER_OPCODE_TXF_LZ: case ELK_SHADER_OPCODE_TXL: case ELK_SHADER_OPCODE_TXL_LZ: /* 18 cycles: * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; * send(8) g4<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * * 697 +/-49 cycles (min 610, n=26): * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; * send(8) g4<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * So the latency on our first texture load of the batchbuffer takes * ~700 cycles, since the caches are cold at that point. * * 840 +/- 92 cycles (min 720, n=25): * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; * send(8) g4<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * send(8) g4<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * On the second load, it takes just an extra ~140 cycles, and after * accounting for the 14 cycles of the MOV's latency, that makes ~130. * * 683 +/- 49 cycles (min = 602, n=47): * mov(8) g115<1>F 0F { align1 WE_normal 1Q }; * mov(8) g114<1>F 0F { align1 WE_normal 1Q }; * send(8) g4<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * send(8) g50<1>UW g114<8,8,1>F * sampler (10, 0, 0, 1) mlen 2 rlen 4 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * The unit appears to be pipelined, since this matches up with the * cache-cold case, despite there being two loads here. If you replace * the g4 in the MOV to null with g50, it's still 693 +/- 52 (n=39). * * So, take some number between the cache-hot 140 cycles and the * cache-cold 700 cycles. No particular tuning was done on this. * * I haven't done significant testing of the non-TEX opcodes. TXL at * least looked about the same as TEX. */ latency = 200; break; case ELK_SHADER_OPCODE_TXS: /* Testing textureSize(sampler2D, 0), one load was 420 +/- 41 * cycles (n=15): * mov(8) g114<1>UD 0D { align1 WE_normal 1Q }; * send(8) g6<1>UW g114<8,8,1>F * sampler (10, 0, 10, 1) mlen 1 rlen 4 { align1 WE_normal 1Q }; * mov(16) g6<1>F g6<8,8,1>D { align1 WE_normal 1Q }; * * * Two loads was 535 +/- 30 cycles (n=19): * mov(16) g114<1>UD 0D { align1 WE_normal 1H }; * send(16) g6<1>UW g114<8,8,1>F * sampler (10, 0, 10, 2) mlen 2 rlen 8 { align1 WE_normal 1H }; * mov(16) g114<1>UD 0D { align1 WE_normal 1H }; * mov(16) g6<1>F g6<8,8,1>D { align1 WE_normal 1H }; * send(16) g8<1>UW g114<8,8,1>F * sampler (10, 0, 10, 2) mlen 2 rlen 8 { align1 WE_normal 1H }; * mov(16) g8<1>F g8<8,8,1>D { align1 WE_normal 1H }; * add(16) g6<1>F g6<8,8,1>F g8<8,8,1>F { align1 WE_normal 1H }; * * Since the only caches that should matter are just the * instruction/state cache containing the surface state, assume that we * always have hot caches. */ latency = 100; break; case ELK_FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GFX4: case ELK_FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD: case ELK_VS_OPCODE_PULL_CONSTANT_LOAD: /* testing using varying-index pull constants: * * 16 cycles: * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; * send(8) g4<1>F g4<8,8,1>D * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; * * ~480 cycles: * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; * send(8) g4<1>F g4<8,8,1>D * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * ~620 cycles: * mov(8) g4<1>D g2.1<0,1,0>F { align1 WE_normal 1Q }; * send(8) g4<1>F g4<8,8,1>D * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * send(8) g4<1>F g4<8,8,1>D * data (9, 2, 3) mlen 1 rlen 1 { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; * * So, if it's cache-hot, it's about 140. If it's cache cold, it's * about 460. We expect to mostly be cache hot, so pick something more * in that direction. */ latency = 200; break; case ELK_SHADER_OPCODE_GFX7_SCRATCH_READ: /* Testing a load from offset 0, that had been previously written: * * send(8) g114<1>UW g0<8,8,1>F data (0, 0, 0) mlen 1 rlen 1 { align1 WE_normal 1Q }; * mov(8) null g114<8,8,1>F { align1 WE_normal 1Q }; * * The cycles spent seemed to be grouped around 40-50 (as low as 38), * then around 140. Presumably this is cache hit vs miss. */ latency = 50; break; case ELK_VEC4_OPCODE_UNTYPED_ATOMIC: /* See GFX7_DATAPORT_DC_UNTYPED_ATOMIC_OP */ latency = 14000; break; case ELK_VEC4_OPCODE_UNTYPED_SURFACE_READ: case ELK_VEC4_OPCODE_UNTYPED_SURFACE_WRITE: /* See also GFX7_DATAPORT_DC_UNTYPED_SURFACE_READ */ latency = is_haswell ? 300 : 600; break; case ELK_SHADER_OPCODE_SEND: switch (inst->sfid) { case ELK_SFID_SAMPLER: { unsigned msg_type = (inst->desc >> 12) & 0x1f; switch (msg_type) { case GFX5_SAMPLER_MESSAGE_SAMPLE_RESINFO: case GFX6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO: /* See also ELK_SHADER_OPCODE_TXS */ latency = 100; break; default: /* See also ELK_SHADER_OPCODE_TEX */ latency = 200; break; } break; } case GFX6_SFID_DATAPORT_CONSTANT_CACHE: /* See ELK_FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD */ latency = 200; break; case GFX6_SFID_DATAPORT_RENDER_CACHE: switch (elk_fb_desc_msg_type(isa->devinfo, inst->desc)) { case GFX7_DATAPORT_RC_TYPED_SURFACE_WRITE: case GFX7_DATAPORT_RC_TYPED_SURFACE_READ: /* See also ELK_SHADER_OPCODE_TYPED_SURFACE_READ */ assert(!is_haswell); latency = 600; break; case GFX7_DATAPORT_RC_TYPED_ATOMIC_OP: /* See also ELK_SHADER_OPCODE_TYPED_ATOMIC */ assert(!is_haswell); latency = 14000; break; case GFX6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE: /* completely fabricated number */ latency = 600; break; default: unreachable("Unknown render cache message"); } break; case GFX7_SFID_DATAPORT_DATA_CACHE: switch ((inst->desc >> 14) & 0x1f) { case ELK_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ: case GFX7_DATAPORT_DC_UNALIGNED_OWORD_BLOCK_READ: case GFX6_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE: /* We have no data for this but assume it's a little faster than * untyped surface read/write. */ latency = 200; break; case GFX7_DATAPORT_DC_DWORD_SCATTERED_READ: case GFX6_DATAPORT_WRITE_MESSAGE_DWORD_SCATTERED_WRITE: case HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_READ: case HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_WRITE: /* We have no data for this but assume it's roughly the same as * untyped surface read/write. */ latency = 300; break; case GFX7_DATAPORT_DC_UNTYPED_SURFACE_READ: case GFX7_DATAPORT_DC_UNTYPED_SURFACE_WRITE: /* Test code: * mov(8) g112<1>UD 0x00000000UD { align1 WE_all 1Q }; * mov(1) g112.7<1>UD g1.7<0,1,0>UD { align1 WE_all }; * mov(8) g113<1>UD 0x00000000UD { align1 WE_normal 1Q }; * send(8) g4<1>UD g112<8,8,1>UD * data (38, 6, 5) mlen 2 rlen 1 { align1 WE_normal 1Q }; * . * . [repeats 8 times] * . * mov(8) g112<1>UD 0x00000000UD { align1 WE_all 1Q }; * mov(1) g112.7<1>UD g1.7<0,1,0>UD { align1 WE_all }; * mov(8) g113<1>UD 0x00000000UD { align1 WE_normal 1Q }; * send(8) g4<1>UD g112<8,8,1>UD * data (38, 6, 5) mlen 2 rlen 1 { align1 WE_normal 1Q }; * * Running it 100 times as fragment shader on a 128x128 quad * gives an average latency of 583 cycles per surface read, * standard deviation 0.9%. */ assert(!is_haswell); latency = 600; break; case GFX7_DATAPORT_DC_UNTYPED_ATOMIC_OP: /* Test code: * mov(8) g112<1>ud 0x00000000ud { align1 WE_all 1Q }; * mov(1) g112.7<1>ud g1.7<0,1,0>ud { align1 WE_all }; * mov(8) g113<1>ud 0x00000000ud { align1 WE_normal 1Q }; * send(8) g4<1>ud g112<8,8,1>ud * data (38, 5, 6) mlen 2 rlen 1 { align1 WE_normal 1Q }; * * Running it 100 times as fragment shader on a 128x128 quad * gives an average latency of 13867 cycles per atomic op, * standard deviation 3%. Note that this is a rather * pessimistic estimate, the actual latency in cases with few * collisions between threads and favorable pipelining has been * seen to be reduced by a factor of 100. */ assert(!is_haswell); latency = 14000; break; default: unreachable("Unknown data cache message"); } break; case HSW_SFID_DATAPORT_DATA_CACHE_1: switch (elk_dp_desc_msg_type(isa->devinfo, inst->desc)) { case HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ: case HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE: case HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ: case HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE: case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_WRITE: case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_READ: case GFX8_DATAPORT_DC_PORT1_A64_SCATTERED_WRITE: case GFX9_DATAPORT_DC_PORT1_A64_SCATTERED_READ: case GFX8_DATAPORT_DC_PORT1_A64_OWORD_BLOCK_READ: case GFX8_DATAPORT_DC_PORT1_A64_OWORD_BLOCK_WRITE: /* See also GFX7_DATAPORT_DC_UNTYPED_SURFACE_READ */ latency = 300; break; case HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP: case HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2: case HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2: case HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP: case GFX8_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_OP: /* See also GFX7_DATAPORT_DC_UNTYPED_ATOMIC_OP */ latency = 14000; break; default: unreachable("Unknown data cache message"); } break; case GFX7_SFID_PIXEL_INTERPOLATOR: latency = 50; /* TODO */ break; case ELK_SFID_URB: latency = 200; break; default: unreachable("Unknown SFID"); } break; default: /* 2 cycles: * mul(8) g4<1>F g2<0,1,0>F 0.5F { align1 WE_normal 1Q }; * * 16 cycles: * mul(8) g4<1>F g2<0,1,0>F 0.5F { align1 WE_normal 1Q }; * mov(8) null g4<8,8,1>F { align1 WE_normal 1Q }; */ latency = 14; break; } } class elk_instruction_scheduler { public: elk_instruction_scheduler(void *mem_ctx, const elk_backend_shader *s, int grf_count, int grf_write_scale, bool post_reg_alloc): bs(s) { this->mem_ctx = mem_ctx; this->lin_ctx = linear_context(this->mem_ctx); this->grf_count = grf_count; this->post_reg_alloc = post_reg_alloc; this->last_grf_write = linear_zalloc_array(lin_ctx, elk_schedule_node *, grf_count * grf_write_scale); this->nodes_len = s->cfg->last_block()->end_ip + 1; this->nodes = linear_zalloc_array(lin_ctx, elk_schedule_node, this->nodes_len); const struct intel_device_info *devinfo = bs->devinfo; const struct elk_isa_info *isa = &bs->compiler->isa; elk_schedule_node *n = nodes; foreach_block_and_inst(block, elk_backend_instruction, inst, s->cfg) { n->inst = inst; /* We can't measure Gfx6 timings directly but expect them to be much * closer to Gfx7 than Gfx4. */ if (!post_reg_alloc) n->latency = 1; else if (devinfo->ver >= 6) n->set_latency_gfx7(isa); else n->set_latency_gfx4(); n++; } assert(n == nodes + nodes_len); current.block = NULL; current.start = NULL; current.end = NULL; current.len = 0; current.time = 0; current.cand_generation = 0; current.available.make_empty(); } void add_barrier_deps(elk_schedule_node *n); void add_cross_lane_deps(elk_schedule_node *n); void add_dep(elk_schedule_node *before, elk_schedule_node *after, int latency); void add_dep(elk_schedule_node *before, elk_schedule_node *after); void set_current_block(elk_bblock_t *block); void compute_delays(); void compute_exits(); void schedule(elk_schedule_node *chosen); void update_children(elk_schedule_node *chosen); void *mem_ctx; linear_ctx *lin_ctx; elk_schedule_node *nodes; int nodes_len; /* Current block being processed. */ struct { elk_bblock_t *block; /* Range of nodes in the block. End will point to first node * address after the block, i.e. the range is [start, end). */ elk_schedule_node *start; elk_schedule_node *end; int len; int scheduled; unsigned cand_generation; int time; exec_list available; } current; bool post_reg_alloc; int grf_count; const elk_backend_shader *bs; /** * Last instruction to have written the grf (or a channel in the grf, for the * scalar backend) */ elk_schedule_node **last_grf_write; }; class elk_fs_instruction_scheduler : public elk_instruction_scheduler { public: elk_fs_instruction_scheduler(void *mem_ctx, const elk_fs_visitor *v, int grf_count, int hw_reg_count, int block_count, bool post_reg_alloc); void calculate_deps(); bool is_compressed(const elk_fs_inst *inst); elk_schedule_node *choose_instruction_to_schedule(); int calculate_issue_time(elk_backend_instruction *inst); void count_reads_remaining(elk_backend_instruction *inst); void setup_liveness(elk_cfg_t *cfg); void update_register_pressure(elk_backend_instruction *inst); int get_register_pressure_benefit(elk_backend_instruction *inst); void clear_last_grf_write(); void schedule_instructions(); void run(instruction_scheduler_mode mode); const elk_fs_visitor *v; unsigned hw_reg_count; int reg_pressure; instruction_scheduler_mode mode; /* * The register pressure at the beginning of each basic block. */ int *reg_pressure_in; /* * The virtual GRF's whose range overlaps the beginning of each basic block. */ BITSET_WORD **livein; /* * The virtual GRF's whose range overlaps the end of each basic block. */ BITSET_WORD **liveout; /* * The hardware GRF's whose range overlaps the end of each basic block. */ BITSET_WORD **hw_liveout; /* * Whether we've scheduled a write for this virtual GRF yet. */ bool *written; /* * How many reads we haven't scheduled for this virtual GRF yet. */ int *reads_remaining; /* * How many reads we haven't scheduled for this hardware GRF yet. */ int *hw_reads_remaining; }; elk_fs_instruction_scheduler::elk_fs_instruction_scheduler(void *mem_ctx, const elk_fs_visitor *v, int grf_count, int hw_reg_count, int block_count, bool post_reg_alloc) : elk_instruction_scheduler(mem_ctx, v, grf_count, /* grf_write_scale */ 16, post_reg_alloc), v(v) { this->hw_reg_count = hw_reg_count; this->mode = SCHEDULE_NONE; this->reg_pressure = 0; if (!post_reg_alloc) { this->reg_pressure_in = linear_zalloc_array(lin_ctx, int, block_count); this->livein = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count); for (int i = 0; i < block_count; i++) this->livein[i] = linear_zalloc_array(lin_ctx, BITSET_WORD, BITSET_WORDS(grf_count)); this->liveout = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count); for (int i = 0; i < block_count; i++) this->liveout[i] = linear_zalloc_array(lin_ctx, BITSET_WORD, BITSET_WORDS(grf_count)); this->hw_liveout = linear_alloc_array(lin_ctx, BITSET_WORD *, block_count); for (int i = 0; i < block_count; i++) this->hw_liveout[i] = linear_zalloc_array(lin_ctx, BITSET_WORD, BITSET_WORDS(hw_reg_count)); setup_liveness(v->cfg); this->written = linear_alloc_array(lin_ctx, bool, grf_count); this->reads_remaining = linear_alloc_array(lin_ctx, int, grf_count); this->hw_reads_remaining = linear_alloc_array(lin_ctx, int, hw_reg_count); } else { this->reg_pressure_in = NULL; this->livein = NULL; this->liveout = NULL; this->hw_liveout = NULL; this->written = NULL; this->reads_remaining = NULL; this->hw_reads_remaining = NULL; } foreach_block(block, v->cfg) { set_current_block(block); for (elk_schedule_node *n = current.start; n < current.end; n++) n->issue_time = calculate_issue_time(n->inst); calculate_deps(); compute_delays(); compute_exits(); } } static bool is_src_duplicate(elk_fs_inst *inst, int src) { for (int i = 0; i < src; i++) if (inst->src[i].equals(inst->src[src])) return true; return false; } void elk_fs_instruction_scheduler::count_reads_remaining(elk_backend_instruction *be) { assert(reads_remaining); elk_fs_inst *inst = (elk_fs_inst *)be; for (int i = 0; i < inst->sources; i++) { if (is_src_duplicate(inst, i)) continue; if (inst->src[i].file == VGRF) { reads_remaining[inst->src[i].nr]++; } else if (inst->src[i].file == FIXED_GRF) { if (inst->src[i].nr >= hw_reg_count) continue; for (unsigned j = 0; j < regs_read(inst, i); j++) hw_reads_remaining[inst->src[i].nr + j]++; } } } void elk_fs_instruction_scheduler::setup_liveness(elk_cfg_t *cfg) { const fs_live_variables &live = v->live_analysis.require(); /* First, compute liveness on a per-GRF level using the in/out sets from * liveness calculation. */ for (int block = 0; block < cfg->num_blocks; block++) { for (int i = 0; i < live.num_vars; i++) { if (BITSET_TEST(live.block_data[block].livein, i)) { int vgrf = live.vgrf_from_var[i]; if (!BITSET_TEST(livein[block], vgrf)) { reg_pressure_in[block] += v->alloc.sizes[vgrf]; BITSET_SET(livein[block], vgrf); } } if (BITSET_TEST(live.block_data[block].liveout, i)) BITSET_SET(liveout[block], live.vgrf_from_var[i]); } } /* Now, extend the live in/live out sets for when a range crosses a block * boundary, which matches what our register allocator/interference code * does to account for force_writemask_all and incompatible exec_mask's. */ for (int block = 0; block < cfg->num_blocks - 1; block++) { for (int i = 0; i < grf_count; i++) { if (live.vgrf_start[i] <= cfg->blocks[block]->end_ip && live.vgrf_end[i] >= cfg->blocks[block + 1]->start_ip) { if (!BITSET_TEST(livein[block + 1], i)) { reg_pressure_in[block + 1] += v->alloc.sizes[i]; BITSET_SET(livein[block + 1], i); } BITSET_SET(liveout[block], i); } } } int *payload_last_use_ip = ralloc_array(NULL, int, hw_reg_count); v->calculate_payload_ranges(hw_reg_count, payload_last_use_ip); for (unsigned i = 0; i < hw_reg_count; i++) { if (payload_last_use_ip[i] == -1) continue; for (int block = 0; block < cfg->num_blocks; block++) { if (cfg->blocks[block]->start_ip <= payload_last_use_ip[i]) reg_pressure_in[block]++; if (cfg->blocks[block]->end_ip <= payload_last_use_ip[i]) BITSET_SET(hw_liveout[block], i); } } ralloc_free(payload_last_use_ip); } void elk_fs_instruction_scheduler::update_register_pressure(elk_backend_instruction *be) { assert(reads_remaining); elk_fs_inst *inst = (elk_fs_inst *)be; if (inst->dst.file == VGRF) { written[inst->dst.nr] = true; } for (int i = 0; i < inst->sources; i++) { if (is_src_duplicate(inst, i)) continue; if (inst->src[i].file == VGRF) { reads_remaining[inst->src[i].nr]--; } else if (inst->src[i].file == FIXED_GRF && inst->src[i].nr < hw_reg_count) { for (unsigned off = 0; off < regs_read(inst, i); off++) hw_reads_remaining[inst->src[i].nr + off]--; } } } int elk_fs_instruction_scheduler::get_register_pressure_benefit(elk_backend_instruction *be) { elk_fs_inst *inst = (elk_fs_inst *)be; int benefit = 0; const int block_idx = current.block->num; if (inst->dst.file == VGRF) { if (!BITSET_TEST(livein[block_idx], inst->dst.nr) && !written[inst->dst.nr]) benefit -= v->alloc.sizes[inst->dst.nr]; } for (int i = 0; i < inst->sources; i++) { if (is_src_duplicate(inst, i)) continue; if (inst->src[i].file == VGRF && !BITSET_TEST(liveout[block_idx], inst->src[i].nr) && reads_remaining[inst->src[i].nr] == 1) benefit += v->alloc.sizes[inst->src[i].nr]; if (inst->src[i].file == FIXED_GRF && inst->src[i].nr < hw_reg_count) { for (unsigned off = 0; off < regs_read(inst, i); off++) { int reg = inst->src[i].nr + off; if (!BITSET_TEST(hw_liveout[block_idx], reg) && hw_reads_remaining[reg] == 1) { benefit++; } } } } return benefit; } class elk_vec4_instruction_scheduler : public elk_instruction_scheduler { public: elk_vec4_instruction_scheduler(void *mem_ctx, const vec4_visitor *v, int grf_count); void calculate_deps(); elk_schedule_node *choose_instruction_to_schedule(); const vec4_visitor *v; void run(); }; elk_vec4_instruction_scheduler::elk_vec4_instruction_scheduler(void *mem_ctx, const vec4_visitor *v, int grf_count) : elk_instruction_scheduler(mem_ctx, v, grf_count, /* grf_write_scale */ 1, /* post_reg_alloc */ true), v(v) { } void elk_instruction_scheduler::set_current_block(elk_bblock_t *block) { current.block = block; current.start = nodes + block->start_ip; current.len = block->end_ip - block->start_ip + 1; current.end = current.start + current.len; current.time = 0; current.scheduled = 0; current.cand_generation = 1; } /** Computation of the delay member of each node. */ void elk_instruction_scheduler::compute_delays() { for (elk_schedule_node *n = current.end - 1; n >= current.start; n--) { if (!n->children_count) { n->delay = n->issue_time; } else { for (int i = 0; i < n->children_count; i++) { assert(n->children[i].n->delay); n->delay = MAX2(n->delay, n->latency + n->children[i].n->delay); } } } } void elk_instruction_scheduler::compute_exits() { /* Calculate a lower bound of the scheduling time of each node in the * graph. This is analogous to the node's critical path but calculated * from the top instead of from the bottom of the block. */ for (elk_schedule_node *n = current.start; n < current.end; n++) { for (int i = 0; i < n->children_count; i++) { elk_schedule_node_child *child = &n->children[i]; child->n->initial_unblocked_time = MAX2(child->n->initial_unblocked_time, n->initial_unblocked_time + n->issue_time + child->effective_latency); } } /* Calculate the exit of each node by induction based on the exit nodes of * its children. The preferred exit of a node is the one among the exit * nodes of its children which can be unblocked first according to the * optimistic unblocked time estimate calculated above. */ for (elk_schedule_node *n = current.end - 1; n >= current.start; n--) { n->exit = (n->inst->opcode == ELK_OPCODE_HALT ? n : NULL); for (int i = 0; i < n->children_count; i++) { if (exit_initial_unblocked_time(n->children[i].n) < exit_initial_unblocked_time(n)) n->exit = n->children[i].n->exit; } } } /** * Add a dependency between two instruction nodes. * * The @after node will be scheduled after @before. We will try to * schedule it @latency cycles after @before, but no guarantees there. */ void elk_instruction_scheduler::add_dep(elk_schedule_node *before, elk_schedule_node *after, int latency) { if (!before || !after) return; assert(before != after); for (int i = 0; i < before->children_count; i++) { elk_schedule_node_child *child = &before->children[i]; if (child->n == after) { child->effective_latency = MAX2(child->effective_latency, latency); return; } } if (before->children_cap <= before->children_count) { if (before->children_cap < 16) before->children_cap = 16; else before->children_cap *= 2; before->children = reralloc(mem_ctx, before->children, elk_schedule_node_child, before->children_cap); } elk_schedule_node_child *child = &before->children[before->children_count]; child->n = after; child->effective_latency = latency; before->children_count++; after->initial_parent_count++; } void elk_instruction_scheduler::add_dep(elk_schedule_node *before, elk_schedule_node *after) { if (!before) return; add_dep(before, after, before->latency); } static bool is_scheduling_barrier(const elk_backend_instruction *inst) { return inst->opcode == ELK_SHADER_OPCODE_HALT_TARGET || inst->is_control_flow() || inst->has_side_effects(); } static bool has_cross_lane_access(const elk_fs_inst *inst) { /* FINISHME: * * This function is likely incomplete in terms of identify cross lane * accesses. */ if (inst->opcode == ELK_SHADER_OPCODE_BROADCAST || inst->opcode == ELK_SHADER_OPCODE_READ_SR_REG || inst->opcode == ELK_SHADER_OPCODE_CLUSTER_BROADCAST || inst->opcode == ELK_SHADER_OPCODE_SHUFFLE || inst->opcode == ELK_FS_OPCODE_LOAD_LIVE_CHANNELS || inst->opcode == ELK_SHADER_OPCODE_FIND_LAST_LIVE_CHANNEL || inst->opcode == ELK_SHADER_OPCODE_FIND_LIVE_CHANNEL) return true; for (unsigned s = 0; s < inst->sources; s++) { if (inst->src[s].file == VGRF) { if (inst->src[s].stride == 0) return true; } } return false; } /** * Sometimes we really want this node to execute after everything that * was before it and before everything that followed it. This adds * the deps to do so. */ void elk_instruction_scheduler::add_barrier_deps(elk_schedule_node *n) { for (elk_schedule_node *prev = n - 1; prev >= current.start; prev--) { add_dep(prev, n, 0); if (is_scheduling_barrier(prev->inst)) break; } for (elk_schedule_node *next = n + 1; next < current.end; next++) { add_dep(n, next, 0); if (is_scheduling_barrier(next->inst)) break; } } /** * Because some instructions like HALT can disable lanes, scheduling prior to * a cross lane access should not be allowed, otherwise we could end up with * later instructions accessing uninitialized data. */ void elk_instruction_scheduler::add_cross_lane_deps(elk_schedule_node *n) { for (elk_schedule_node *prev = n - 1; prev >= current.start; prev--) { if (has_cross_lane_access((elk_fs_inst*)prev->inst)) add_dep(prev, n, 0); } } /* instruction scheduling needs to be aware of when an MRF write * actually writes 2 MRFs. */ bool elk_fs_instruction_scheduler::is_compressed(const elk_fs_inst *inst) { return inst->exec_size == 16; } /* Clears last_grf_write to be ready to start calculating deps for a block * again. * * Since pre-ra grf_count scales with instructions, and instructions scale with * BBs, we don't want to memset all of last_grf_write per block or you'll end up * O(n^2) with number of blocks. For shaders using softfp64, we get a *lot* of * blocks. * * We don't bother being careful for post-ra, since then grf_count doesn't scale * with instructions. */ void elk_fs_instruction_scheduler::clear_last_grf_write() { if (!post_reg_alloc) { for (elk_schedule_node *n = current.start; n < current.end; n++) { elk_fs_inst *inst = (elk_fs_inst *)n->inst; if (inst->dst.file == VGRF) { /* Don't bother being careful with regs_written(), quicker to just clear 2 cachelines. */ memset(&last_grf_write[inst->dst.nr * 16], 0, sizeof(*last_grf_write) * 16); } } } else { memset(last_grf_write, 0, sizeof(*last_grf_write) * grf_count * 16); } } void elk_fs_instruction_scheduler::calculate_deps() { /* Pre-register-allocation, this tracks the last write per VGRF offset. * After register allocation, reg_offsets are gone and we track individual * GRF registers. */ elk_schedule_node *last_mrf_write[ELK_MAX_MRF_ALL]; elk_schedule_node *last_conditional_mod[8] = {}; elk_schedule_node *last_accumulator_write = NULL; /* Fixed HW registers are assumed to be separate from the virtual * GRFs, so they can be tracked separately. We don't really write * to fixed GRFs much, so don't bother tracking them on a more * granular level. */ elk_schedule_node *last_fixed_grf_write = NULL; memset(last_mrf_write, 0, sizeof(last_mrf_write)); /* top-to-bottom dependencies: RAW and WAW. */ for (elk_schedule_node *n = current.start; n < current.end; n++) { elk_fs_inst *inst = (elk_fs_inst *)n->inst; if (is_scheduling_barrier(inst)) add_barrier_deps(n); if (inst->opcode == ELK_OPCODE_HALT || inst->opcode == ELK_SHADER_OPCODE_HALT_TARGET) add_cross_lane_deps(n); /* read-after-write deps. */ for (int i = 0; i < inst->sources; i++) { if (inst->src[i].file == VGRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_read(inst, i); r++) add_dep(last_grf_write[inst->src[i].nr + r], n); } else { for (unsigned r = 0; r < regs_read(inst, i); r++) { add_dep(last_grf_write[inst->src[i].nr * 16 + inst->src[i].offset / REG_SIZE + r], n); } } } else if (inst->src[i].file == FIXED_GRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_read(inst, i); r++) add_dep(last_grf_write[inst->src[i].nr + r], n); } else { add_dep(last_fixed_grf_write, n); } } else if (inst->src[i].is_accumulator()) { add_dep(last_accumulator_write, n); } else if (inst->src[i].file == ARF && !inst->src[i].is_null()) { add_barrier_deps(n); } } if (inst->base_mrf != -1) { for (int i = 0; i < inst->mlen; i++) { /* It looks like the MRF regs are released in the send * instruction once it's sent, not when the result comes * back. */ add_dep(last_mrf_write[inst->base_mrf + i], n); } } if (const unsigned mask = inst->flags_read(v->devinfo)) { assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { if (mask & (1 << i)) add_dep(last_conditional_mod[i], n); } } if (inst->reads_accumulator_implicitly()) { add_dep(last_accumulator_write, n); } /* write-after-write deps. */ if (inst->dst.file == VGRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_written(inst); r++) { add_dep(last_grf_write[inst->dst.nr + r], n); last_grf_write[inst->dst.nr + r] = n; } } else { for (unsigned r = 0; r < regs_written(inst); r++) { add_dep(last_grf_write[inst->dst.nr * 16 + inst->dst.offset / REG_SIZE + r], n); last_grf_write[inst->dst.nr * 16 + inst->dst.offset / REG_SIZE + r] = n; } } } else if (inst->dst.file == MRF) { int reg = inst->dst.nr & ~ELK_MRF_COMPR4; add_dep(last_mrf_write[reg], n); last_mrf_write[reg] = n; if (is_compressed(inst)) { if (inst->dst.nr & ELK_MRF_COMPR4) reg += 4; else reg++; add_dep(last_mrf_write[reg], n); last_mrf_write[reg] = n; } } else if (inst->dst.file == FIXED_GRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_written(inst); r++) { add_dep(last_grf_write[inst->dst.nr + r], n); last_grf_write[inst->dst.nr + r] = n; } } else { add_dep(last_fixed_grf_write, n); last_fixed_grf_write = n; } } else if (inst->dst.is_accumulator()) { add_dep(last_accumulator_write, n); last_accumulator_write = n; } else if (inst->dst.file == ARF && !inst->dst.is_null()) { add_barrier_deps(n); } if (inst->mlen > 0 && inst->base_mrf != -1) { for (unsigned i = 0; i < inst->implied_mrf_writes(); i++) { add_dep(last_mrf_write[inst->base_mrf + i], n); last_mrf_write[inst->base_mrf + i] = n; } } if (const unsigned mask = inst->flags_written(v->devinfo)) { assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { if (mask & (1 << i)) { add_dep(last_conditional_mod[i], n, 0); last_conditional_mod[i] = n; } } } if (inst->writes_accumulator_implicitly(v->devinfo) && !inst->dst.is_accumulator()) { add_dep(last_accumulator_write, n); last_accumulator_write = n; } } clear_last_grf_write(); /* bottom-to-top dependencies: WAR */ memset(last_mrf_write, 0, sizeof(last_mrf_write)); memset(last_conditional_mod, 0, sizeof(last_conditional_mod)); last_accumulator_write = NULL; last_fixed_grf_write = NULL; for (elk_schedule_node *n = current.end - 1; n >= current.start; n--) { elk_fs_inst *inst = (elk_fs_inst *)n->inst; /* write-after-read deps. */ for (int i = 0; i < inst->sources; i++) { if (inst->src[i].file == VGRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_read(inst, i); r++) add_dep(n, last_grf_write[inst->src[i].nr + r], 0); } else { for (unsigned r = 0; r < regs_read(inst, i); r++) { add_dep(n, last_grf_write[inst->src[i].nr * 16 + inst->src[i].offset / REG_SIZE + r], 0); } } } else if (inst->src[i].file == FIXED_GRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_read(inst, i); r++) add_dep(n, last_grf_write[inst->src[i].nr + r], 0); } else { add_dep(n, last_fixed_grf_write, 0); } } else if (inst->src[i].is_accumulator()) { add_dep(n, last_accumulator_write, 0); } else if (inst->src[i].file == ARF && !inst->src[i].is_null()) { add_barrier_deps(n); } } if (inst->base_mrf != -1) { for (int i = 0; i < inst->mlen; i++) { /* It looks like the MRF regs are released in the send * instruction once it's sent, not when the result comes * back. */ add_dep(n, last_mrf_write[inst->base_mrf + i], 2); } } if (const unsigned mask = inst->flags_read(v->devinfo)) { assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { if (mask & (1 << i)) add_dep(n, last_conditional_mod[i]); } } if (inst->reads_accumulator_implicitly()) { add_dep(n, last_accumulator_write); } /* Update the things this instruction wrote, so earlier reads * can mark this as WAR dependency. */ if (inst->dst.file == VGRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_written(inst); r++) last_grf_write[inst->dst.nr + r] = n; } else { for (unsigned r = 0; r < regs_written(inst); r++) { last_grf_write[inst->dst.nr * 16 + inst->dst.offset / REG_SIZE + r] = n; } } } else if (inst->dst.file == MRF) { int reg = inst->dst.nr & ~ELK_MRF_COMPR4; last_mrf_write[reg] = n; if (is_compressed(inst)) { if (inst->dst.nr & ELK_MRF_COMPR4) reg += 4; else reg++; last_mrf_write[reg] = n; } } else if (inst->dst.file == FIXED_GRF) { if (post_reg_alloc) { for (unsigned r = 0; r < regs_written(inst); r++) last_grf_write[inst->dst.nr + r] = n; } else { last_fixed_grf_write = n; } } else if (inst->dst.is_accumulator()) { last_accumulator_write = n; } else if (inst->dst.file == ARF && !inst->dst.is_null()) { add_barrier_deps(n); } if (inst->mlen > 0 && inst->base_mrf != -1) { for (unsigned i = 0; i < inst->implied_mrf_writes(); i++) { last_mrf_write[inst->base_mrf + i] = n; } } if (const unsigned mask = inst->flags_written(v->devinfo)) { assert(mask < (1 << ARRAY_SIZE(last_conditional_mod))); for (unsigned i = 0; i < ARRAY_SIZE(last_conditional_mod); i++) { if (mask & (1 << i)) last_conditional_mod[i] = n; } } if (inst->writes_accumulator_implicitly(v->devinfo)) { last_accumulator_write = n; } } clear_last_grf_write(); } void elk_vec4_instruction_scheduler::calculate_deps() { elk_schedule_node *last_mrf_write[ELK_MAX_MRF_ALL]; elk_schedule_node *last_conditional_mod = NULL; elk_schedule_node *last_accumulator_write = NULL; /* Fixed HW registers are assumed to be separate from the virtual * GRFs, so they can be tracked separately. We don't really write * to fixed GRFs much, so don't bother tracking them on a more * granular level. */ elk_schedule_node *last_fixed_grf_write = NULL; memset(last_grf_write, 0, grf_count * sizeof(*last_grf_write)); memset(last_mrf_write, 0, sizeof(last_mrf_write)); /* top-to-bottom dependencies: RAW and WAW. */ for (elk_schedule_node *n = current.start; n < current.end; n++) { vec4_instruction *inst = (vec4_instruction *)n->inst; if (is_scheduling_barrier(inst)) add_barrier_deps(n); /* read-after-write deps. */ for (int i = 0; i < 3; i++) { if (inst->src[i].file == VGRF) { for (unsigned j = 0; j < regs_read(inst, i); ++j) add_dep(last_grf_write[inst->src[i].nr + j], n); } else if (inst->src[i].file == FIXED_GRF) { add_dep(last_fixed_grf_write, n); } else if (inst->src[i].is_accumulator()) { assert(last_accumulator_write); add_dep(last_accumulator_write, n); } else if (inst->src[i].file == ARF && !inst->src[i].is_null()) { add_barrier_deps(n); } } if (inst->reads_g0_implicitly()) add_dep(last_fixed_grf_write, n); if (!inst->is_send_from_grf()) { for (int i = 0; i < inst->mlen; i++) { /* It looks like the MRF regs are released in the send * instruction once it's sent, not when the result comes * back. */ add_dep(last_mrf_write[inst->base_mrf + i], n); } } if (inst->reads_flag()) { assert(last_conditional_mod); add_dep(last_conditional_mod, n); } if (inst->reads_accumulator_implicitly()) { assert(last_accumulator_write); add_dep(last_accumulator_write, n); } /* write-after-write deps. */ if (inst->dst.file == VGRF) { for (unsigned j = 0; j < regs_written(inst); ++j) { add_dep(last_grf_write[inst->dst.nr + j], n); last_grf_write[inst->dst.nr + j] = n; } } else if (inst->dst.file == MRF) { add_dep(last_mrf_write[inst->dst.nr], n); last_mrf_write[inst->dst.nr] = n; } else if (inst->dst.file == FIXED_GRF) { add_dep(last_fixed_grf_write, n); last_fixed_grf_write = n; } else if (inst->dst.is_accumulator()) { add_dep(last_accumulator_write, n); last_accumulator_write = n; } else if (inst->dst.file == ARF && !inst->dst.is_null()) { add_barrier_deps(n); } if (inst->mlen > 0 && !inst->is_send_from_grf()) { for (unsigned i = 0; i < inst->implied_mrf_writes(); i++) { add_dep(last_mrf_write[inst->base_mrf + i], n); last_mrf_write[inst->base_mrf + i] = n; } } if (inst->writes_flag(v->devinfo)) { add_dep(last_conditional_mod, n, 0); last_conditional_mod = n; } if (inst->writes_accumulator_implicitly(v->devinfo) && !inst->dst.is_accumulator()) { add_dep(last_accumulator_write, n); last_accumulator_write = n; } } /* bottom-to-top dependencies: WAR */ memset(last_grf_write, 0, grf_count * sizeof(*last_grf_write)); memset(last_mrf_write, 0, sizeof(last_mrf_write)); last_conditional_mod = NULL; last_accumulator_write = NULL; last_fixed_grf_write = NULL; for (elk_schedule_node *n = current.end - 1; n >= current.start; n--) { vec4_instruction *inst = (vec4_instruction *)n->inst; /* write-after-read deps. */ for (int i = 0; i < 3; i++) { if (inst->src[i].file == VGRF) { for (unsigned j = 0; j < regs_read(inst, i); ++j) add_dep(n, last_grf_write[inst->src[i].nr + j]); } else if (inst->src[i].file == FIXED_GRF) { add_dep(n, last_fixed_grf_write); } else if (inst->src[i].is_accumulator()) { add_dep(n, last_accumulator_write); } else if (inst->src[i].file == ARF && !inst->src[i].is_null()) { add_barrier_deps(n); } } if (!inst->is_send_from_grf()) { for (int i = 0; i < inst->mlen; i++) { /* It looks like the MRF regs are released in the send * instruction once it's sent, not when the result comes * back. */ add_dep(n, last_mrf_write[inst->base_mrf + i], 2); } } if (inst->reads_flag()) { add_dep(n, last_conditional_mod); } if (inst->reads_accumulator_implicitly()) { add_dep(n, last_accumulator_write); } /* Update the things this instruction wrote, so earlier reads * can mark this as WAR dependency. */ if (inst->dst.file == VGRF) { for (unsigned j = 0; j < regs_written(inst); ++j) last_grf_write[inst->dst.nr + j] = n; } else if (inst->dst.file == MRF) { last_mrf_write[inst->dst.nr] = n; } else if (inst->dst.file == FIXED_GRF) { last_fixed_grf_write = n; } else if (inst->dst.is_accumulator()) { last_accumulator_write = n; } else if (inst->dst.file == ARF && !inst->dst.is_null()) { add_barrier_deps(n); } if (inst->mlen > 0 && !inst->is_send_from_grf()) { for (unsigned i = 0; i < inst->implied_mrf_writes(); i++) { last_mrf_write[inst->base_mrf + i] = n; } } if (inst->writes_flag(v->devinfo)) { last_conditional_mod = n; } if (inst->writes_accumulator_implicitly(v->devinfo)) { last_accumulator_write = n; } } } elk_schedule_node * elk_fs_instruction_scheduler::choose_instruction_to_schedule() { elk_schedule_node *chosen = NULL; if (mode == SCHEDULE_PRE || mode == SCHEDULE_POST) { int chosen_time = 0; /* Of the instructions ready to execute or the closest to being ready, * choose the one most likely to unblock an early program exit, or * otherwise the oldest one. */ foreach_in_list(elk_schedule_node, n, ¤t.available) { if (!chosen || exit_tmp_unblocked_time(n) < exit_tmp_unblocked_time(chosen) || (exit_tmp_unblocked_time(n) == exit_tmp_unblocked_time(chosen) && n->tmp.unblocked_time < chosen_time)) { chosen = n; chosen_time = n->tmp.unblocked_time; } } } else { int chosen_register_pressure_benefit = 0; /* Before register allocation, we don't care about the latencies of * instructions. All we care about is reducing live intervals of * variables so that we can avoid register spilling, or get SIMD16 * shaders which naturally do a better job of hiding instruction * latency. */ foreach_in_list(elk_schedule_node, n, ¤t.available) { elk_fs_inst *inst = (elk_fs_inst *)n->inst; if (!chosen) { chosen = n; chosen_register_pressure_benefit = get_register_pressure_benefit(chosen->inst); continue; } /* Most important: If we can definitely reduce register pressure, do * so immediately. */ int register_pressure_benefit = get_register_pressure_benefit(n->inst); if (register_pressure_benefit > 0 && register_pressure_benefit > chosen_register_pressure_benefit) { chosen = n; chosen_register_pressure_benefit = register_pressure_benefit; continue; } else if (chosen_register_pressure_benefit > 0 && (register_pressure_benefit < chosen_register_pressure_benefit)) { continue; } if (mode == SCHEDULE_PRE_LIFO) { /* Prefer instructions that recently became available for * scheduling. These are the things that are most likely to * (eventually) make a variable dead and reduce register pressure. * Typical register pressure estimates don't work for us because * most of our pressure comes from texturing, where no single * instruction to schedule will make a vec4 value dead. */ if (n->tmp.cand_generation > chosen->tmp.cand_generation) { chosen = n; chosen_register_pressure_benefit = register_pressure_benefit; continue; } else if (n->tmp.cand_generation < chosen->tmp.cand_generation) { continue; } /* On MRF-using chips, prefer non-SEND instructions. If we don't * do this, then because we prefer instructions that just became * candidates, we'll end up in a pattern of scheduling a SEND, * then the MRFs for the next SEND, then the next SEND, then the * MRFs, etc., without ever consuming the results of a send. */ if (v->devinfo->ver < 7) { elk_fs_inst *chosen_inst = (elk_fs_inst *)chosen->inst; /* We use size_written > 4 * exec_size as our test for the kind * of send instruction to avoid -- only sends generate many * regs, and a single-result send is probably actually reducing * register pressure. */ if (inst->size_written <= 4 * inst->exec_size && chosen_inst->size_written > 4 * chosen_inst->exec_size) { chosen = n; chosen_register_pressure_benefit = register_pressure_benefit; continue; } else if (inst->size_written > chosen_inst->size_written) { continue; } } } /* For instructions pushed on the cands list at the same time, prefer * the one with the highest delay to the end of the program. This is * most likely to have its values able to be consumed first (such as * for a large tree of lowered ubo loads, which appear reversed in * the instruction stream with respect to when they can be consumed). */ if (n->delay > chosen->delay) { chosen = n; chosen_register_pressure_benefit = register_pressure_benefit; continue; } else if (n->delay < chosen->delay) { continue; } /* Prefer the node most likely to unblock an early program exit. */ if (exit_tmp_unblocked_time(n) < exit_tmp_unblocked_time(chosen)) { chosen = n; chosen_register_pressure_benefit = register_pressure_benefit; continue; } else if (exit_tmp_unblocked_time(n) > exit_tmp_unblocked_time(chosen)) { continue; } /* If all other metrics are equal, we prefer the first instruction in * the list (program execution). */ } } return chosen; } elk_schedule_node * elk_vec4_instruction_scheduler::choose_instruction_to_schedule() { elk_schedule_node *chosen = NULL; int chosen_time = 0; /* Of the instructions ready to execute or the closest to being ready, * choose the oldest one. */ foreach_in_list(elk_schedule_node, n, ¤t.available) { if (!chosen || n->tmp.unblocked_time < chosen_time) { chosen = n; chosen_time = n->tmp.unblocked_time; } } return chosen; } int elk_fs_instruction_scheduler::calculate_issue_time(elk_backend_instruction *inst0) { const struct elk_isa_info *isa = &v->compiler->isa; const elk_fs_inst *inst = static_cast<elk_fs_inst *>(inst0); const unsigned overhead = v->grf_used && elk_has_bank_conflict(isa, inst) ? DIV_ROUND_UP(inst->dst.component_size(inst->exec_size), REG_SIZE) : 0; if (is_compressed(inst)) return 4 + overhead; else return 2 + overhead; } void elk_instruction_scheduler::schedule(elk_schedule_node *chosen) { assert(current.scheduled < current.len); current.scheduled++; assert(chosen); chosen->remove(); current.block->instructions.push_tail(chosen->inst); /* If we expected a delay for scheduling, then bump the clock to reflect * that. In reality, the hardware will switch to another hyperthread * and may not return to dispatching our thread for a while even after * we're unblocked. After this, we have the time when the chosen * instruction will start executing. */ current.time = MAX2(current.time, chosen->tmp.unblocked_time); /* Update the clock for how soon an instruction could start after the * chosen one. */ current.time += chosen->issue_time; if (debug) { fprintf(stderr, "clock %4d, scheduled: ", current.time); bs->dump_instruction(chosen->inst); } } void elk_instruction_scheduler::update_children(elk_schedule_node *chosen) { /* Now that we've scheduled a new instruction, some of its * children can be promoted to the list of instructions ready to * be scheduled. Update the children's unblocked time for this * DAG edge as we do so. */ for (int i = chosen->children_count - 1; i >= 0; i--) { elk_schedule_node_child *child = &chosen->children[i]; child->n->tmp.unblocked_time = MAX2(child->n->tmp.unblocked_time, current.time + child->effective_latency); if (debug) { fprintf(stderr, "\tchild %d, %d parents: ", i, child->n->tmp.parent_count); bs->dump_instruction(child->n->inst); } child->n->tmp.cand_generation = current.cand_generation; child->n->tmp.parent_count--; if (child->n->tmp.parent_count == 0) { if (debug) { fprintf(stderr, "\t\tnow available\n"); } current.available.push_head(child->n); } } current.cand_generation++; /* Shared resource: the mathbox. There's one mathbox per EU on Gfx6+ * but it's more limited pre-gfx6, so if we send something off to it then * the next math instruction isn't going to make progress until the first * is done. */ if (bs->devinfo->ver < 6 && chosen->inst->is_math()) { foreach_in_list(elk_schedule_node, n, ¤t.available) { if (n->inst->is_math()) n->tmp.unblocked_time = MAX2(n->tmp.unblocked_time, current.time + chosen->latency); } } } void elk_fs_instruction_scheduler::schedule_instructions() { if (!post_reg_alloc) reg_pressure = reg_pressure_in[current.block->num]; assert(current.available.is_empty()); for (elk_schedule_node *n = current.start; n < current.end; n++) { reset_node_tmp(n); /* Add DAG heads to the list of available instructions. */ if (n->tmp.parent_count == 0) current.available.push_tail(n); } current.block->instructions.make_empty(); while (!current.available.is_empty()) { elk_schedule_node *chosen = choose_instruction_to_schedule(); schedule(chosen); if (!post_reg_alloc) { reg_pressure -= get_register_pressure_benefit(chosen->inst); update_register_pressure(chosen->inst); if (debug) fprintf(stderr, "(register pressure %d)\n", reg_pressure); } update_children(chosen); } } void elk_fs_instruction_scheduler::run(instruction_scheduler_mode mode) { this->mode = mode; if (debug && !post_reg_alloc) { fprintf(stderr, "\nInstructions before scheduling (reg_alloc %d)\n", post_reg_alloc); bs->dump_instructions(); } if (!post_reg_alloc) { memset(reads_remaining, 0, grf_count * sizeof(*reads_remaining)); memset(hw_reads_remaining, 0, hw_reg_count * sizeof(*hw_reads_remaining)); memset(written, 0, grf_count * sizeof(*written)); } foreach_block(block, v->cfg) { set_current_block(block); if (!post_reg_alloc) { for (elk_schedule_node *n = current.start; n < current.end; n++) count_reads_remaining(n->inst); } schedule_instructions(); } if (debug && !post_reg_alloc) { fprintf(stderr, "\nInstructions after scheduling (reg_alloc %d)\n", post_reg_alloc); bs->dump_instructions(); } } void elk_vec4_instruction_scheduler::run() { foreach_block(block, v->cfg) { set_current_block(block); for (elk_schedule_node *n = current.start; n < current.end; n++) { /* We always execute as two vec4s in parallel. */ n->issue_time = 2; } calculate_deps(); compute_delays(); compute_exits(); assert(current.available.is_empty()); for (elk_schedule_node *n = current.start; n < current.end; n++) { reset_node_tmp(n); /* Add DAG heads to the list of available instructions. */ if (n->tmp.parent_count == 0) current.available.push_tail(n); } current.block->instructions.make_empty(); while (!current.available.is_empty()) { elk_schedule_node *chosen = choose_instruction_to_schedule(); schedule(chosen); update_children(chosen); } } } elk_fs_instruction_scheduler * elk_fs_visitor::prepare_scheduler(void *mem_ctx) { const int grf_count = alloc.count; elk_fs_instruction_scheduler *empty = rzalloc(mem_ctx, elk_fs_instruction_scheduler); return new (empty) elk_fs_instruction_scheduler(mem_ctx, this, grf_count, first_non_payload_grf, cfg->num_blocks, /* post_reg_alloc */ false); } void elk_fs_visitor::schedule_instructions_pre_ra(elk_fs_instruction_scheduler *sched, instruction_scheduler_mode mode) { if (mode == SCHEDULE_NONE) return; sched->run(mode); invalidate_analysis(DEPENDENCY_INSTRUCTIONS); } void elk_fs_visitor::schedule_instructions_post_ra() { const bool post_reg_alloc = true; const int grf_count = reg_unit(devinfo) * grf_used; void *mem_ctx = ralloc_context(NULL); elk_fs_instruction_scheduler sched(mem_ctx, this, grf_count, first_non_payload_grf, cfg->num_blocks, post_reg_alloc); sched.run(SCHEDULE_POST); ralloc_free(mem_ctx); invalidate_analysis(DEPENDENCY_INSTRUCTIONS); } void vec4_visitor::opt_schedule_instructions() { void *mem_ctx = ralloc_context(NULL); elk_vec4_instruction_scheduler sched(mem_ctx, this, prog_data->total_grf); sched.run(); ralloc_free(mem_ctx); invalidate_analysis(DEPENDENCY_INSTRUCTIONS); }