Edit
IABSD.fr/xenocara/lib/mesa/src/intel/compiler/brw_ir_performance.cpp
jsg
- lib/mesa/src/intel/compiler/brw_ir_performance.cpp
-
/* * Copyright © 2020 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. */ #include "brw_eu.h" #include "brw_fs.h" #include "brw_cfg.h" using namespace brw; namespace { /** * Enumeration representing the various asynchronous units that can run * computations in parallel on behalf of a shader thread. */ enum intel_eu_unit { /** EU front-end. */ EU_UNIT_FE, /** EU FPU0 (Note that co-issue to FPU1 is currently not modeled here). */ EU_UNIT_FPU, /** Extended Math unit (AKA FPU1 on Gfx8-11, part of the EU on Gfx6+). */ EU_UNIT_EM, /** Sampler shared function. */ EU_UNIT_SAMPLER, /** Pixel Interpolator shared function. */ EU_UNIT_PI, /** Unified Return Buffer shared function. */ EU_UNIT_URB, /** Data Port Data Cache shared function. */ EU_UNIT_DP_DC, /** Data Port Render Cache shared function. */ EU_UNIT_DP_RC, /** Data Port Constant Cache shared function. */ EU_UNIT_DP_CC, /** Message Gateway shared function. */ EU_UNIT_GATEWAY, /** Thread Spawner shared function. */ EU_UNIT_SPAWNER, /* EU_UNIT_VME, */ /* EU_UNIT_CRE, */ /** Number of asynchronous units currently tracked. */ EU_NUM_UNITS, /** Dummy unit for instructions that don't consume runtime from the above. */ EU_UNIT_NULL = EU_NUM_UNITS }; /** * Enumeration representing a computation result another computation can * potentially depend on. */ enum intel_eu_dependency_id { /* Register part of the GRF. */ EU_DEPENDENCY_ID_GRF0 = 0, /* Address register part of the ARF. */ EU_DEPENDENCY_ID_ADDR0 = EU_DEPENDENCY_ID_GRF0 + XE3_MAX_GRF, /* Accumulator register part of the ARF. */ EU_DEPENDENCY_ID_ACCUM0 = EU_DEPENDENCY_ID_ADDR0 + 1, /* Flag register part of the ARF. */ EU_DEPENDENCY_ID_FLAG0 = EU_DEPENDENCY_ID_ACCUM0 + 12, /* SBID token write completion. Only used on Gfx12+. */ EU_DEPENDENCY_ID_SBID_WR0 = EU_DEPENDENCY_ID_FLAG0 + 8, /* SBID token read completion. Only used on Gfx12+. */ EU_DEPENDENCY_ID_SBID_RD0 = EU_DEPENDENCY_ID_SBID_WR0 + 32, /* Number of computation dependencies currently tracked. */ EU_NUM_DEPENDENCY_IDS = EU_DEPENDENCY_ID_SBID_RD0 + 32 }; /** * State of our modeling of the program execution. */ struct state { state() : unit_ready(), dep_ready(), unit_busy(), weight(1.0) {} /** * Time at which a given unit will be ready to execute the next * computation, in clock units. */ unsigned unit_ready[EU_NUM_UNITS]; /** * Time at which an instruction dependent on a given dependency ID will * be ready to execute, in clock units. */ unsigned dep_ready[EU_NUM_DEPENDENCY_IDS]; /** * Aggregated utilization of a given unit excluding idle cycles, * in clock units. */ float unit_busy[EU_NUM_UNITS]; /** * Factor of the overhead of a computation accounted for in the * aggregated utilization calculation. */ float weight; }; /** * Information derived from an IR instruction used to compute performance * estimates. Allows the timing calculation to work on both FS and VEC4 * instructions. */ struct instruction_info { instruction_info(const struct brw_isa_info *isa, const fs_inst *inst) : isa(isa), devinfo(isa->devinfo), op(inst->opcode), td(inst->dst.type), sd(DIV_ROUND_UP(inst->size_written, REG_SIZE)), tx(get_exec_type(inst)), sx(0), ss(0), sc(has_bank_conflict(isa, inst) ? sd : 0), desc(inst->desc), sfid(inst->sfid) { /* We typically want the maximum source size, except for split send * messages which require the total size. */ if (inst->opcode == SHADER_OPCODE_SEND) { ss = DIV_ROUND_UP(inst->size_read(devinfo, 2), REG_SIZE) + DIV_ROUND_UP(inst->size_read(devinfo, 3), REG_SIZE); } else if (inst->opcode == SHADER_OPCODE_SEND_GATHER) { ss = inst->mlen; /* If haven't lowered yet, count the sources. */ if (!ss) { for (int i = 3; i < inst->sources; i++) ss += DIV_ROUND_UP(inst->size_read(devinfo, i), REG_SIZE); } } else { for (unsigned i = 0; i < inst->sources; i++) ss = MAX2(ss, DIV_ROUND_UP(inst->size_read(devinfo, i), REG_SIZE)); } /* Convert the execution size to GRF units. */ sx = DIV_ROUND_UP(inst->exec_size * brw_type_size_bytes(tx), REG_SIZE); /* 32x32 integer multiplication has half the usual ALU throughput. * Treat it as double-precision. */ if ((inst->opcode == BRW_OPCODE_MUL || inst->opcode == BRW_OPCODE_MAD) && !brw_type_is_float(tx) && brw_type_size_bytes(tx) == 4 && brw_type_size_bytes(inst->src[0].type) == brw_type_size_bytes(inst->src[1].type)) tx = brw_int_type(8, tx == BRW_TYPE_D); rcount = inst->opcode == BRW_OPCODE_DPAS ? inst->rcount : 0; } /** ISA encoding information */ const struct brw_isa_info *isa; /** Device information. */ const struct intel_device_info *devinfo; /** Instruction opcode. */ opcode op; /** Destination type. */ brw_reg_type td; /** Destination size in GRF units. */ unsigned sd; /** Execution type. */ brw_reg_type tx; /** Execution size in GRF units. */ unsigned sx; /** Source size. */ unsigned ss; /** Bank conflict penalty size in GRF units (equal to sd if non-zero). */ unsigned sc; /** Send message descriptor. */ uint32_t desc; /** Send message shared function ID. */ uint8_t sfid; /** Repeat count for DPAS instructions. */ uint8_t rcount; }; /** * Timing information of an instruction used to estimate the performance of * the program. */ struct perf_desc { perf_desc(enum intel_eu_unit u, int df, int db, int ls, int ld, int la, int lf) : u(u), df(df), db(db), ls(ls), ld(ld), la(la), lf(lf) {} /** * Back-end unit its runtime shall be accounted to, in addition to the * EU front-end which is always assumed to be involved. */ enum intel_eu_unit u; /** * Overhead cycles from the time that the EU front-end starts executing * the instruction until it's ready to execute the next instruction. */ int df; /** * Overhead cycles from the time that the back-end starts executing the * instruction until it's ready to execute the next instruction. */ int db; /** * Latency cycles from the time that the back-end starts executing the * instruction until its sources have been read from the register file. */ int ls; /** * Latency cycles from the time that the back-end starts executing the * instruction until its regular destination has been written to the * register file. */ int ld; /** * Latency cycles from the time that the back-end starts executing the * instruction until its accumulator destination has been written to the * ARF file. * * Note that this is an approximation of the real behavior of * accumulating instructions in the hardware: Instead of modeling a pair * of back-to-back accumulating instructions as a first computation with * latency equal to ld followed by another computation with a * mid-pipeline stall (e.g. after the "M" part of a MAC instruction), we * model the stall as if it occurred at the top of the pipeline, with * the latency of the accumulator computation offset accordingly. */ int la; /** * Latency cycles from the time that the back-end starts executing the * instruction until its flag destination has been written to the ARF * file. */ int lf; }; /** * Compute the timing information of an instruction based on any relevant * information from the IR and a number of parameters specifying a linear * approximation: Parameter X_Y specifies the derivative of timing X * relative to info field Y, while X_1 specifies the independent term of * the approximation of timing X. */ perf_desc calculate_desc(const instruction_info &info, enum intel_eu_unit u, int df_1, int df_sd, int df_sc, int db_1, int db_sx, int ls_1, int ld_1, int la_1, int lf_1, int l_ss, int l_sd) { return perf_desc(u, df_1 + df_sd * int(info.sd) + df_sc * int(info.sc), db_1 + db_sx * int(info.sx), ls_1 + l_ss * int(info.ss), ld_1 + l_ss * int(info.ss) + l_sd * int(info.sd), la_1, lf_1); } /** * Compute the timing information of an instruction based on any relevant * information from the IR and a number of linear approximation parameters * hard-coded for each IR instruction. * * Most timing parameters are obtained from the multivariate linear * regression of a sample of empirical timings measured using the tm0 * register (as can be done today by using the shader_time debugging * option). The Gfx4-5 math timings are obtained from BSpec Volume 5c.3 * "Shared Functions - Extended Math", Section 3.2 "Performance". * Parameters marked XXX shall be considered low-quality, they're possibly * high variance or completely guessed in cases where experimental data was * unavailable. */ const perf_desc instruction_desc(const instruction_info &info) { const struct intel_device_info *devinfo = info.devinfo; switch (info.op) { case BRW_OPCODE_SYNC: case BRW_OPCODE_SEL: case BRW_OPCODE_NOT: case BRW_OPCODE_AND: case BRW_OPCODE_OR: case BRW_OPCODE_XOR: case BRW_OPCODE_SHR: case BRW_OPCODE_SHL: case BRW_OPCODE_ASR: case BRW_OPCODE_CMPN: case BRW_OPCODE_BFREV: case BRW_OPCODE_BFI1: case BRW_OPCODE_AVG: case BRW_OPCODE_FRC: case BRW_OPCODE_RNDU: case BRW_OPCODE_RNDD: case BRW_OPCODE_RNDE: case BRW_OPCODE_RNDZ: case BRW_OPCODE_MAC: case BRW_OPCODE_MACH: case BRW_OPCODE_LZD: case BRW_OPCODE_FBH: case BRW_OPCODE_FBL: case BRW_OPCODE_CBIT: case BRW_OPCODE_ADDC: case BRW_OPCODE_ROR: case BRW_OPCODE_ROL: case BRW_OPCODE_SUBB: case BRW_OPCODE_LINE: case BRW_OPCODE_NOP: case SHADER_OPCODE_CLUSTER_BROADCAST: case SHADER_OPCODE_SCRATCH_HEADER: case FS_OPCODE_DDX_COARSE: case FS_OPCODE_DDX_FINE: case FS_OPCODE_DDY_COARSE: case FS_OPCODE_PIXEL_X: case FS_OPCODE_PIXEL_Y: if (devinfo->ver >= 11) { return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 10, 6 /* XXX */, 14, 0, 0); } else { if (brw_type_size_bytes(info.tx) > 4) return calculate_desc(info, EU_UNIT_FPU, 0, 4, 0, 0, 4, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 8, 4, 12, 0, 0); } case BRW_OPCODE_MOV: case BRW_OPCODE_CMP: case BRW_OPCODE_ADD: case BRW_OPCODE_ADD3: case BRW_OPCODE_MUL: case SHADER_OPCODE_MOV_RELOC_IMM: if (devinfo->ver >= 11) { return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 10, 6, 14, 0, 0); } else { if (brw_type_size_bytes(info.tx) > 4) return calculate_desc(info, EU_UNIT_FPU, 0, 4, 0, 0, 4, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 8, 4, 12, 0, 0); } case BRW_OPCODE_BFE: case BRW_OPCODE_BFI2: case BRW_OPCODE_CSEL: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, 8, 4 /* XXX */, 12 /* XXX */, 0, 0); case BRW_OPCODE_MAD: if (devinfo->ver >= 11) { return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); } else { if (brw_type_size_bytes(info.tx) > 4) return calculate_desc(info, EU_UNIT_FPU, 0, 4, 1, 0, 4, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, 8, 4 /* XXX */, 12 /* XXX */, 0, 0); } case BRW_OPCODE_DP4: case BRW_OPCODE_DPH: case BRW_OPCODE_DP3: case BRW_OPCODE_DP2: return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); case BRW_OPCODE_DP4A: if (devinfo->ver >= 12) return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); else abort(); case BRW_OPCODE_DPAS: { unsigned ld; switch (info.rcount) { case 1: ld = 21; break; case 2: ld = 22; break; case 8: default: ld = 32; break; } /* DPAS cannot write the accumulator or the flags, so pass UINT_MAX * for la and lf. */ if (devinfo->verx10 >= 125) return calculate_desc(info, EU_UNIT_FPU, 0, 2, 1, 0, 2, 0, ld, UINT_MAX, UINT_MAX, 0, 0); else abort(); } case SHADER_OPCODE_RCP: case SHADER_OPCODE_RSQ: case SHADER_OPCODE_SQRT: case SHADER_OPCODE_EXP2: case SHADER_OPCODE_LOG2: case SHADER_OPCODE_SIN: case SHADER_OPCODE_COS: return calculate_desc(info, EU_UNIT_EM, -2, 4, 0, 0, 4, 0, 16, 0, 0, 0, 0); case SHADER_OPCODE_POW: return calculate_desc(info, EU_UNIT_EM, -2, 4, 0, 0, 8, 0, 24, 0, 0, 0, 0); case SHADER_OPCODE_INT_QUOTIENT: case SHADER_OPCODE_INT_REMAINDER: return calculate_desc(info, EU_UNIT_EM, 2, 0, 0, 26, 0, 0, 28 /* XXX */, 0, 0, 0, 0); case BRW_OPCODE_DO: return calculate_desc(info, EU_UNIT_NULL, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); case BRW_OPCODE_IF: case BRW_OPCODE_ELSE: case BRW_OPCODE_ENDIF: case BRW_OPCODE_WHILE: case BRW_OPCODE_BREAK: case BRW_OPCODE_CONTINUE: case BRW_OPCODE_HALT: return calculate_desc(info, EU_UNIT_NULL, 8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); case BRW_OPCODE_PLN: return calculate_desc(info, EU_UNIT_FPU, 0, 4, 0, 0, 4, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); case BRW_OPCODE_LRP: return calculate_desc(info, EU_UNIT_FPU, 0, 4, 1, 0, 4, 0, 12, 8 /* XXX */, 16 /* XXX */, 0, 0); case FS_OPCODE_PACK_HALF_2x16_SPLIT: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 20, 6, 0, 0, 6, 0, 10 /* XXX */, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 16, 6, 0, 0, 6, 0, 8 /* XXX */, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_READ_ARCH_REG: if (devinfo->ver >= 12) { return calculate_desc(info, EU_UNIT_FPU, 20, 6, 0, 0, 6, 0, 10, 6 /* XXX */, 14, 0, 0); } else { return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 8, 4, 12, 0, 0); } case SHADER_OPCODE_MOV_INDIRECT: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 34, 0, 0, 34, 0, 0, 10 /* XXX */, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 34, 0, 0, 34, 0, 0, 8 /* XXX */, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_BROADCAST: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 20 /* XXX */, 0, 0, 4, 0, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 18, 0, 0, 4, 0, 0, 8, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_FIND_LIVE_CHANNEL: case SHADER_OPCODE_FIND_LAST_LIVE_CHANNEL: case SHADER_OPCODE_LOAD_LIVE_CHANNELS: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 2, 0, 0, 2, 0, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 2, 0, 0, 2, 0, 0, 8, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_RND_MODE: case SHADER_OPCODE_FLOAT_CONTROL_MODE: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 24 /* XXX */, 0, 0, 4 /* XXX */, 0, 0, 0, 0, 0, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 20 /* XXX */, 0, 0, 4 /* XXX */, 0, 0, 0, 0, 0, 0, 0); case SHADER_OPCODE_SHUFFLE: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 44 /* XXX */, 0, 0, 44 /* XXX */, 0, 0, 10 /* XXX */, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 42 /* XXX */, 0, 0, 42 /* XXX */, 0, 0, 8 /* XXX */, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_SEL_EXEC: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 10 /* XXX */, 4 /* XXX */, 0, 0, 4 /* XXX */, 0, 10 /* XXX */, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 8 /* XXX */, 4 /* XXX */, 0, 0, 4 /* XXX */, 0, 8 /* XXX */, 4 /* XXX */, 12 /* XXX */, 0, 0); case SHADER_OPCODE_QUAD_SWIZZLE: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 0 /* XXX */, 8 /* XXX */, 0, 0, 8 /* XXX */, 0, 10 /* XXX */, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0 /* XXX */, 8 /* XXX */, 0, 0, 8 /* XXX */, 0, 8 /* XXX */, 4 /* XXX */, 12 /* XXX */, 0, 0); case FS_OPCODE_DDY_FINE: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 0, 14, 0, 0, 4, 0, 10, 6 /* XXX */, 14 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2, 0, 0, 2, 0, 8, 4 /* XXX */, 12 /* XXX */, 0, 0); case FS_OPCODE_LOAD_LIVE_CHANNELS: if (devinfo->ver >= 11) return calculate_desc(info, EU_UNIT_FPU, 2 /* XXX */, 0, 0, 2 /* XXX */, 0, 0, 0, 0, 10 /* XXX */, 0, 0); else return calculate_desc(info, EU_UNIT_FPU, 0, 2 /* XXX */, 0, 0, 2 /* XXX */, 0, 0, 0, 8 /* XXX */, 0, 0); case SHADER_OPCODE_GET_BUFFER_SIZE: return calculate_desc(info, EU_UNIT_SAMPLER, 2, 0, 0, 0, 16 /* XXX */, 8 /* XXX */, 750 /* XXX */, 0, 0, 2 /* XXX */, 0); case SHADER_OPCODE_MEMORY_FENCE: case SHADER_OPCODE_INTERLOCK: switch (info.sfid) { case GFX6_SFID_DATAPORT_RENDER_CACHE: return calculate_desc(info, EU_UNIT_DP_RC, 2, 0, 0, 30 /* XXX */, 0, 10 /* XXX */, 300 /* XXX */, 0, 0, 0, 0); case BRW_SFID_URB: case GFX7_SFID_DATAPORT_DATA_CACHE: case GFX12_SFID_SLM: case GFX12_SFID_TGM: case GFX12_SFID_UGM: case HSW_SFID_DATAPORT_DATA_CACHE_1: return calculate_desc(info, EU_UNIT_DP_DC, 2, 0, 0, 30 /* XXX */, 0, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 0); default: abort(); } case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD: return calculate_desc(info, EU_UNIT_DP_CC, 2, 0, 0, 0, 16 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 0); case FS_OPCODE_INTERPOLATE_AT_SAMPLE: case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET: case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET: return calculate_desc(info, EU_UNIT_PI, 2, 0, 0, 14 /* XXX */, 0, 0, 90 /* XXX */, 0, 0, 0, 0); case SHADER_OPCODE_BARRIER: return calculate_desc(info, EU_UNIT_GATEWAY, 90 /* XXX */, 0, 0, 0 /* XXX */, 0, 0, 0, 0, 0, 0, 0); case SHADER_OPCODE_SEND: case SHADER_OPCODE_SEND_GATHER: switch (info.sfid) { case GFX6_SFID_DATAPORT_CONSTANT_CACHE: /* See FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD */ return calculate_desc(info, EU_UNIT_DP_CC, 2, 0, 0, 0, 16 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 0); case GFX6_SFID_DATAPORT_RENDER_CACHE: switch (brw_dp_desc_msg_type(devinfo, info.desc)) { case GFX7_DATAPORT_RC_TYPED_ATOMIC_OP: return calculate_desc(info, EU_UNIT_DP_RC, 2, 0, 0, 30 /* XXX */, 450 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 400 /* XXX */); default: return calculate_desc(info, EU_UNIT_DP_RC, 2, 0, 0, 0, 450 /* XXX */, 10 /* XXX */, 300 /* XXX */, 0, 0, 0, 0); } case BRW_SFID_SAMPLER: { return calculate_desc(info, EU_UNIT_SAMPLER, 2, 0, 0, 0, 16, 8, 750, 0, 0, 2, 0); } case GFX7_SFID_DATAPORT_DATA_CACHE: case HSW_SFID_DATAPORT_DATA_CACHE_1: switch (brw_dp_desc_msg_type(devinfo, info.desc)) { 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: return calculate_desc(info, EU_UNIT_DP_DC, 2, 0, 0, 30 /* XXX */, 400 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 400 /* XXX */); default: return calculate_desc(info, EU_UNIT_DP_DC, 2, 0, 0, 0, 20 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 0); } case GFX7_SFID_PIXEL_INTERPOLATOR: return calculate_desc(info, EU_UNIT_PI, 2, 0, 0, 14 /* XXX */, 0, 0, 90 /* XXX */, 0, 0, 0, 0); case GFX12_SFID_UGM: case GFX12_SFID_TGM: case GFX12_SFID_SLM: switch (lsc_msg_desc_opcode(devinfo, info.desc)) { case LSC_OP_LOAD: case LSC_OP_STORE: case LSC_OP_LOAD_CMASK: case LSC_OP_STORE_CMASK: return calculate_desc(info, EU_UNIT_DP_DC, 2, 0, 0, 0, 20 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 0); case LSC_OP_FENCE: case LSC_OP_ATOMIC_INC: case LSC_OP_ATOMIC_DEC: case LSC_OP_ATOMIC_LOAD: case LSC_OP_ATOMIC_STORE: case LSC_OP_ATOMIC_ADD: case LSC_OP_ATOMIC_SUB: case LSC_OP_ATOMIC_MIN: case LSC_OP_ATOMIC_MAX: case LSC_OP_ATOMIC_UMIN: case LSC_OP_ATOMIC_UMAX: case LSC_OP_ATOMIC_CMPXCHG: case LSC_OP_ATOMIC_FADD: case LSC_OP_ATOMIC_FSUB: case LSC_OP_ATOMIC_FMIN: case LSC_OP_ATOMIC_FMAX: case LSC_OP_ATOMIC_FCMPXCHG: case LSC_OP_ATOMIC_AND: case LSC_OP_ATOMIC_OR: case LSC_OP_ATOMIC_XOR: return calculate_desc(info, EU_UNIT_DP_DC, 2, 0, 0, 30 /* XXX */, 400 /* XXX */, 10 /* XXX */, 100 /* XXX */, 0, 0, 0, 400 /* XXX */); default: abort(); } case BRW_SFID_MESSAGE_GATEWAY: case GEN_RT_SFID_BINDLESS_THREAD_DISPATCH: /* or THREAD_SPAWNER */ case GEN_RT_SFID_RAY_TRACE_ACCELERATOR: return calculate_desc(info, EU_UNIT_SPAWNER, 2, 0, 0, 0 /* XXX */, 0, 10 /* XXX */, 0, 0, 0, 0, 0); case BRW_SFID_URB: return calculate_desc(info, EU_UNIT_URB, 2, 0, 0, 0, 6 /* XXX */, 32 /* XXX */, 200 /* XXX */, 0, 0, 0, 0); default: abort(); } case SHADER_OPCODE_UNDEF: case SHADER_OPCODE_HALT_TARGET: case FS_OPCODE_SCHEDULING_FENCE: return calculate_desc(info, EU_UNIT_NULL, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0); default: abort(); } } /** * Model the performance behavior of a stall on the specified dependency * ID. */ void stall_on_dependency(state &st, enum intel_eu_dependency_id id) { if (id < ARRAY_SIZE(st.dep_ready)) st.unit_ready[EU_UNIT_FE] = MAX2(st.unit_ready[EU_UNIT_FE], st.dep_ready[id]); } /** * Model the performance behavior of the front-end and back-end while * executing an instruction with the specified timing information, assuming * all dependencies are already clear. */ void execute_instruction(state &st, const perf_desc &perf) { /* Compute the time at which the front-end will be ready to execute the * next instruction. */ st.unit_ready[EU_UNIT_FE] += perf.df; if (perf.u < EU_NUM_UNITS) { /* Wait for the back-end to be ready to execute this instruction. */ st.unit_ready[EU_UNIT_FE] = MAX2(st.unit_ready[EU_UNIT_FE], st.unit_ready[perf.u]); /* Compute the time at which the back-end will be ready to execute * the next instruction, and update the back-end utilization. */ st.unit_ready[perf.u] = st.unit_ready[EU_UNIT_FE] + perf.db; st.unit_busy[perf.u] += perf.db * st.weight; } } /** * Model the performance behavior of a read dependency provided by an * instruction. */ void mark_read_dependency(state &st, const perf_desc &perf, enum intel_eu_dependency_id id) { if (id < ARRAY_SIZE(st.dep_ready)) st.dep_ready[id] = st.unit_ready[EU_UNIT_FE] + perf.ls; } /** * Model the performance behavior of a write dependency provided by an * instruction. */ void mark_write_dependency(state &st, const perf_desc &perf, enum intel_eu_dependency_id id) { if (id >= EU_DEPENDENCY_ID_ACCUM0 && id < EU_DEPENDENCY_ID_FLAG0) st.dep_ready[id] = st.unit_ready[EU_UNIT_FE] + perf.la; else if (id >= EU_DEPENDENCY_ID_FLAG0 && id < EU_DEPENDENCY_ID_SBID_WR0) st.dep_ready[id] = st.unit_ready[EU_UNIT_FE] + perf.lf; else if (id < ARRAY_SIZE(st.dep_ready)) st.dep_ready[id] = st.unit_ready[EU_UNIT_FE] + perf.ld; } /** * Return the dependency ID of a backend_reg, offset by \p delta GRFs. */ enum intel_eu_dependency_id reg_dependency_id(const intel_device_info *devinfo, const brw_reg &r, const int delta) { if (r.file == VGRF) { const unsigned i = r.nr + r.offset / REG_SIZE + delta; assert(i < EU_DEPENDENCY_ID_ADDR0 - EU_DEPENDENCY_ID_GRF0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_GRF0 + i); } else if (r.file == FIXED_GRF) { const unsigned i = r.nr + delta; assert(i < EU_DEPENDENCY_ID_ADDR0 - EU_DEPENDENCY_ID_GRF0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_GRF0 + i); } else if (r.file == ARF && r.nr >= BRW_ARF_ADDRESS && r.nr < BRW_ARF_ACCUMULATOR) { assert(delta == 0); return EU_DEPENDENCY_ID_ADDR0; } else if (r.file == ARF && r.nr >= BRW_ARF_ACCUMULATOR && r.nr < BRW_ARF_FLAG) { const unsigned i = r.nr - BRW_ARF_ACCUMULATOR + delta; assert(i < EU_DEPENDENCY_ID_FLAG0 - EU_DEPENDENCY_ID_ACCUM0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_ACCUM0 + i); } else { return EU_NUM_DEPENDENCY_IDS; } } /** * Return the dependency ID of flag register starting at offset \p i. */ enum intel_eu_dependency_id flag_dependency_id(unsigned i) { assert(i < EU_DEPENDENCY_ID_SBID_WR0 - EU_DEPENDENCY_ID_FLAG0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_FLAG0 + i); } /** * Return the dependency ID corresponding to the SBID read completion * condition of a Gfx12+ SWSB. */ enum intel_eu_dependency_id tgl_swsb_rd_dependency_id(tgl_swsb swsb) { if (swsb.mode) { assert(swsb.sbid < EU_NUM_DEPENDENCY_IDS - EU_DEPENDENCY_ID_SBID_RD0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_SBID_RD0 + swsb.sbid); } else { return EU_NUM_DEPENDENCY_IDS; } } /** * Return the dependency ID corresponding to the SBID write completion * condition of a Gfx12+ SWSB. */ enum intel_eu_dependency_id tgl_swsb_wr_dependency_id(tgl_swsb swsb) { if (swsb.mode) { assert(swsb.sbid < EU_DEPENDENCY_ID_SBID_RD0 - EU_DEPENDENCY_ID_SBID_WR0); return intel_eu_dependency_id(EU_DEPENDENCY_ID_SBID_WR0 + swsb.sbid); } else { return EU_NUM_DEPENDENCY_IDS; } } /** * Return the implicit accumulator register accessed by channel \p i of the * instruction. */ unsigned accum_reg_of_channel(const intel_device_info *devinfo, const fs_inst *inst, brw_reg_type tx, unsigned i) { assert(inst->reads_accumulator_implicitly() || inst->writes_accumulator_implicitly(devinfo)); const unsigned offset = (inst->group + i) * brw_type_size_bytes(tx) * (brw_type_is_float(tx) ? 1 : 2); return offset / (reg_unit(devinfo) * REG_SIZE) % 2; } /** * Model the performance behavior of an FS back-end instruction. */ void issue_inst(state &st, const struct brw_isa_info *isa, const fs_inst *inst) { const struct intel_device_info *devinfo = isa->devinfo; const instruction_info info(isa, inst); const perf_desc perf = instruction_desc(info); /* Stall on any source dependencies. */ for (unsigned i = 0; i < inst->sources; i++) { for (unsigned j = 0; j < regs_read(devinfo, inst, i); j++) stall_on_dependency( st, reg_dependency_id(devinfo, inst->src[i], j)); } if (inst->reads_accumulator_implicitly()) { for (unsigned j = accum_reg_of_channel(devinfo, inst, info.tx, 0); j <= accum_reg_of_channel(devinfo, inst, info.tx, inst->exec_size - 1); j++) stall_on_dependency( st, reg_dependency_id(devinfo, brw_acc_reg(8), j)); } if (const unsigned mask = inst->flags_read(devinfo)) { for (unsigned i = 0; i < sizeof(mask) * CHAR_BIT; i++) { if (mask & (1 << i)) stall_on_dependency(st, flag_dependency_id(i)); } } /* Stall on any write dependencies. */ if (!inst->no_dd_check) { if (inst->dst.file != BAD_FILE && !inst->dst.is_null()) { for (unsigned j = 0; j < regs_written(inst); j++) stall_on_dependency( st, reg_dependency_id(devinfo, inst->dst, j)); } if (inst->writes_accumulator_implicitly(devinfo)) { for (unsigned j = accum_reg_of_channel(devinfo, inst, info.tx, 0); j <= accum_reg_of_channel(devinfo, inst, info.tx, inst->exec_size - 1); j++) stall_on_dependency( st, reg_dependency_id(devinfo, brw_acc_reg(8), j)); } if (const unsigned mask = inst->flags_written(devinfo)) { for (unsigned i = 0; i < sizeof(mask) * CHAR_BIT; i++) { if (mask & (1 << i)) stall_on_dependency(st, flag_dependency_id(i)); } } } /* Stall on any SBID dependencies. */ if (inst->sched.mode & (TGL_SBID_SET | TGL_SBID_DST)) stall_on_dependency(st, tgl_swsb_wr_dependency_id(inst->sched)); else if (inst->sched.mode & TGL_SBID_SRC) stall_on_dependency(st, tgl_swsb_rd_dependency_id(inst->sched)); /* Execute the instruction. */ execute_instruction(st, perf); /* Mark any source dependencies. */ if (inst->is_send_from_grf()) { for (unsigned i = 0; i < inst->sources; i++) { if (inst->is_payload(i)) { for (unsigned j = 0; j < regs_read(devinfo, inst, i); j++) mark_read_dependency( st, perf, reg_dependency_id(devinfo, inst->src[i], j)); } } } /* Mark any destination dependencies. */ if (inst->dst.file != BAD_FILE && !inst->dst.is_null()) { for (unsigned j = 0; j < regs_written(inst); j++) { mark_write_dependency(st, perf, reg_dependency_id(devinfo, inst->dst, j)); } } if (inst->writes_accumulator_implicitly(devinfo)) { for (unsigned j = accum_reg_of_channel(devinfo, inst, info.tx, 0); j <= accum_reg_of_channel(devinfo, inst, info.tx, inst->exec_size - 1); j++) mark_write_dependency(st, perf, reg_dependency_id(devinfo, brw_acc_reg(8), j)); } if (const unsigned mask = inst->flags_written(devinfo)) { for (unsigned i = 0; i < sizeof(mask) * CHAR_BIT; i++) { if (mask & (1 << i)) mark_write_dependency(st, perf, flag_dependency_id(i)); } } /* Mark any SBID dependencies. */ if (inst->sched.mode & TGL_SBID_SET) { mark_read_dependency(st, perf, tgl_swsb_rd_dependency_id(inst->sched)); mark_write_dependency(st, perf, tgl_swsb_wr_dependency_id(inst->sched)); } } /** * Calculate the maximum possible throughput of the program compatible with * the cycle-count utilization estimated for each asynchronous unit, in * threads-per-cycle units. */ float calculate_thread_throughput(const state &st, float busy) { for (unsigned i = 0; i < EU_NUM_UNITS; i++) busy = MAX2(busy, st.unit_busy[i]); return 1.0 / busy; } /** * Estimate the performance of the specified shader. */ void calculate_performance(performance &p, const fs_visitor *s, unsigned dispatch_width) { /* XXX - Note that the previous version of this code used worst-case * scenario estimation of branching divergence for SIMD32 shaders, * but this heuristic was removed to improve performance in common * scenarios. Wider shader variants are less optimal when divergence * is high, e.g. when application renders complex scene on a small * surface. It is assumed that such renders are short, so their * time doesn't matter and when it comes to the overall performance, * they are dominated by more optimal larger renders. * * It's possible that we could do better with divergence analysis * by isolating branches which are 100% uniform. * * Plumbing the trip counts from NIR loop analysis would allow us * to do a better job regarding the loop weights. * * In the meantime use values that roughly match the control flow * weights used elsewhere in the compiler back-end. * * Note that we provide slightly more pessimistic weights on * Gfx12.x for SIMD32, since the effective warp size on that * platform is 2x the SIMD width due to EU fusion, which increases * the likelihood of divergent control flow in comparison to * previous generations, giving narrower SIMD modes a performance * advantage in several test-cases with non-uniform discard jumps. * EU fusion has been removed on Xe2+ so its divergence behavior is * expected to be closer to pre-Gfx12 platforms. */ const float discard_weight = (dispatch_width > 16 || s->devinfo->ver != 12 ? 1.0 : 0.5); const float loop_weight = 10; unsigned halt_count = 0; unsigned elapsed = 0; state st; foreach_block(block, s->cfg) { const unsigned elapsed0 = elapsed; foreach_inst_in_block(fs_inst, inst, block) { const unsigned clock0 = st.unit_ready[EU_UNIT_FE]; issue_inst(st, &s->compiler->isa, inst); if (inst->opcode == SHADER_OPCODE_HALT_TARGET && halt_count) st.weight /= discard_weight; elapsed += (st.unit_ready[EU_UNIT_FE] - clock0) * st.weight; if (inst->opcode == BRW_OPCODE_DO) st.weight *= loop_weight; else if (inst->opcode == BRW_OPCODE_WHILE) st.weight /= loop_weight; else if (inst->opcode == BRW_OPCODE_HALT && !halt_count++) st.weight *= discard_weight; } p.block_latency[block->num] = elapsed - elapsed0; } p.latency = elapsed; p.throughput = dispatch_width * calculate_thread_throughput(st, elapsed); } } brw::performance::performance(const fs_visitor *v) : block_latency(new unsigned[v->cfg->num_blocks]) { calculate_performance(*this, v, v->dispatch_width); } brw::performance::~performance() { delete[] block_latency; }