/* * Copyright (C) 2020 Collabora Ltd. * * 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 "bi_builder.h" #include "compiler.h" /* Not all 8-bit and 16-bit instructions support all swizzles on all sources. * These passes, intended to run after NIR->BIR but before scheduling/RA, lower * away swizzles that cannot be represented. In the future, we should try to * recombine swizzles where we can as an optimization. */ static bool bi_swizzle_replicates_8(enum bi_swizzle swz) { switch (swz) { case BI_SWIZZLE_B0000: case BI_SWIZZLE_B1111: case BI_SWIZZLE_B2222: case BI_SWIZZLE_B3333: return true; default: return false; } } static void lower_swizzle(bi_context *ctx, bi_instr *ins, unsigned src) { /* TODO: Use the opcode table and be a lot more methodical about this... */ switch (ins->op) { /* Some instructions used with 16-bit data never have swizzles */ case BI_OPCODE_CSEL_V2F16: case BI_OPCODE_CSEL_V2I16: case BI_OPCODE_CSEL_V2S16: case BI_OPCODE_CSEL_V2U16: /* Despite ostensibly being 32-bit instructions, CLPER does not * inherently interpret the data, so it can be used for v2f16 * derivatives, which might require swizzle lowering */ case BI_OPCODE_CLPER_I32: case BI_OPCODE_CLPER_OLD_I32: /* Similarly, CSEL.i32 consumes a boolean as a 32-bit argument. If the * boolean is implemented as a 16-bit integer, the swizzle is needed * for correct operation if the instruction producing the 16-bit * boolean does not replicate to both halves of the containing 32-bit * register. As such, we may need to lower a swizzle. * * This is a silly hack. Ideally, code gen would be smart enough to * avoid this case (by replicating). In practice, silly hardware design * decisions force our hand here. */ case BI_OPCODE_MUX_I32: case BI_OPCODE_CSEL_I32: break; case BI_OPCODE_IADD_V2S16: case BI_OPCODE_IADD_V2U16: case BI_OPCODE_ISUB_V2S16: case BI_OPCODE_ISUB_V2U16: if (src == 0 && ins->src[src].swizzle != BI_SWIZZLE_H10) break; else return; case BI_OPCODE_LSHIFT_AND_V2I16: case BI_OPCODE_LSHIFT_OR_V2I16: case BI_OPCODE_LSHIFT_XOR_V2I16: case BI_OPCODE_RSHIFT_AND_V2I16: case BI_OPCODE_RSHIFT_OR_V2I16: case BI_OPCODE_RSHIFT_XOR_V2I16: if (src == 2) return; else break; /* For some reason MUX.v2i16 allows swaps but not replication */ case BI_OPCODE_MUX_V2I16: if (ins->src[src].swizzle == BI_SWIZZLE_H10) return; else break; /* No swizzles supported */ case BI_OPCODE_HADD_V4U8: case BI_OPCODE_HADD_V4S8: case BI_OPCODE_CLZ_V4U8: case BI_OPCODE_IDP_V4I8: case BI_OPCODE_IABS_V4S8: case BI_OPCODE_ICMP_V4I8: case BI_OPCODE_ICMP_V4U8: case BI_OPCODE_MUX_V4I8: case BI_OPCODE_IADD_IMM_V4I8: break; case BI_OPCODE_LSHIFT_AND_V4I8: case BI_OPCODE_LSHIFT_OR_V4I8: case BI_OPCODE_LSHIFT_XOR_V4I8: case BI_OPCODE_RSHIFT_AND_V4I8: case BI_OPCODE_RSHIFT_OR_V4I8: case BI_OPCODE_RSHIFT_XOR_V4I8: /* Last source allows identity or replication */ if (src == 2 && bi_swizzle_replicates_8(ins->src[src].swizzle)) return; /* Others do not allow swizzles */ break; /* We don't want to deal with reswizzling logic in modifier prop. Move * the swizzle outside, it's easier for clamp propagation. */ case BI_OPCODE_FCLAMP_V2F16: { bi_builder b = bi_init_builder(ctx, bi_after_instr(ins)); bi_index dest = ins->dest[0]; bi_index tmp = bi_temp(ctx); bi_index swizzled_src = bi_replace_index(ins->src[0], tmp); ins->src[0].swizzle = BI_SWIZZLE_H01; ins->dest[0] = tmp; bi_swz_v2i16_to(&b, dest, swizzled_src); return; } default: return; } /* First, try to apply a given swizzle to a constant to clear the * runtime swizzle. This is less heavy-handed than ignoring the * swizzle for scalar destinations, since it maintains * replication of the destination. */ if (ins->src[src].type == BI_INDEX_CONSTANT) { ins->src[src].value = bi_apply_swizzle(ins->src[src].value, ins->src[src].swizzle); ins->src[src].swizzle = BI_SWIZZLE_H01; return; } /* Even if the source does not replicate, if the consuming instruction * produces a 16-bit scalar, we can ignore the other component. */ if (ins->dest[0].swizzle == BI_SWIZZLE_H00 && ins->src[src].swizzle == BI_SWIZZLE_H00) { ins->src[src].swizzle = BI_SWIZZLE_H01; return; } /* Lower it away */ bi_builder b = bi_init_builder(ctx, bi_before_instr(ins)); bool is_8 = (bi_opcode_props[ins->op].size == BI_SIZE_8) || (bi_opcode_props[ins->op].size == BI_SIZE_32 && ins->src[src].swizzle >= BI_SWIZZLE_B0000); bi_index orig = ins->src[src]; bi_index stripped = bi_replace_index(bi_null(), orig); stripped.swizzle = ins->src[src].swizzle; bi_index swz = is_8 ? bi_swz_v4i8(&b, stripped) : bi_swz_v2i16(&b, stripped); bi_replace_src(ins, src, swz); ins->src[src].swizzle = BI_SWIZZLE_H01; } static bool bi_swizzle_replicates_16(enum bi_swizzle swz) { switch (swz) { case BI_SWIZZLE_H00: case BI_SWIZZLE_H11: return true; default: /* If a swizzle replicates every 8-bits, it also replicates * every 16-bits, so allow 8-bit replicating swizzles. */ return bi_swizzle_replicates_8(swz); } } static bool bi_instr_replicates(bi_instr *I, BITSET_WORD *replicates_16) { switch (I->op) { /* Instructions that construct vectors have replicated output if their * sources are identical. Check this case first. */ case BI_OPCODE_MKVEC_V2I16: case BI_OPCODE_V2F16_TO_V2S16: case BI_OPCODE_V2F16_TO_V2U16: case BI_OPCODE_V2F32_TO_V2F16: case BI_OPCODE_V2S16_TO_V2F16: case BI_OPCODE_V2S8_TO_V2F16: case BI_OPCODE_V2S8_TO_V2S16: case BI_OPCODE_V2U16_TO_V2F16: case BI_OPCODE_V2U8_TO_V2F16: case BI_OPCODE_V2U8_TO_V2U16: return bi_is_value_equiv(I->src[0], I->src[1]); /* 16-bit transcendentals are defined to output zero in their * upper half, so they do not replicate */ case BI_OPCODE_FRCP_F16: case BI_OPCODE_FRSQ_F16: return false; /* Not sure, be conservative, we don't use these.. */ case BI_OPCODE_VN_ASST1_F16: case BI_OPCODE_FPCLASS_F16: case BI_OPCODE_FPOW_SC_DET_F16: return false; default: break; } /* Replication analysis only makes sense for ALU instructions */ if (bi_opcode_props[I->op].message != BIFROST_MESSAGE_NONE) return false; /* We only analyze 16-bit instructions for 16-bit replication. We could * maybe do better. */ if (bi_opcode_props[I->op].size != BI_SIZE_16) return false; bi_foreach_src(I, s) { if (bi_is_null(I->src[s])) continue; /* Replicated swizzles */ if (bi_swizzle_replicates_16(I->src[s].swizzle)) continue; /* Replicated values */ if (bi_is_ssa(I->src[s]) && BITSET_TEST(replicates_16, I->src[s].value)) continue; /* Replicated constants */ if (I->src[s].type == BI_INDEX_CONSTANT && (I->src[s].value & 0xFFFF) == (I->src[s].value >> 16)) continue; return false; } return true; } void bi_lower_swizzle(bi_context *ctx) { bi_foreach_instr_global_safe(ctx, ins) { bi_foreach_src(ins, s) { if (bi_is_null(ins->src[s])) continue; if (ins->src[s].swizzle == BI_SWIZZLE_H01) continue; lower_swizzle(ctx, ins, s); } } /* Now that we've lowered swizzles, clean up the mess */ BITSET_WORD *replicates_16 = calloc(sizeof(bi_index), ctx->ssa_alloc); bi_foreach_instr_global(ctx, ins) { if (ins->nr_dests && bi_instr_replicates(ins, replicates_16)) BITSET_SET(replicates_16, ins->dest[0].value); if (ins->op == BI_OPCODE_SWZ_V2I16 && bi_is_ssa(ins->src[0]) && BITSET_TEST(replicates_16, ins->src[0].value)) { ins->op = BI_OPCODE_MOV_I32; ins->src[0].swizzle = BI_SWIZZLE_H01; } /* The above passes rely on replicating destinations. For * Valhall, we will want to optimize this. For now, default * to Bifrost compatible behaviour. */ if (ins->nr_dests) ins->dest[0].swizzle = BI_SWIZZLE_H01; } free(replicates_16); }