From f5401cb8866cc74c5b3d0fde278fb1046779b415 Mon Sep 17 00:00:00 2001 From: Alyssa Rosenzweig Date: Thu, 27 Feb 2020 09:38:21 -0500 Subject: [PATCH] pan/midgard: Add address analysis framework Midgard has the ability to calculate addresses as part of the load/store pipeline. We'd like to make use of this to avoid doing this work on the ALU pipes. To do so, when emitting globals/SSBOs/shareds, we walk the tree looking for address arithmetic to try to parse out something the hardware can work with, letting the original instructions be DCE'd ideally. This analysis is done at the NIR level to properly account for some messy details of vectorization which we'd rather not poke at the backend level. (Originally I wrote this as a MIR pass but I'm fairly sure it was wrong.) Signed-off-by: Alyssa Rosenzweig Part-of: --- src/panfrost/midgard/midgard_address.c | 205 +++++++++++++++++++++++++ 1 file changed, 205 insertions(+) create mode 100644 src/panfrost/midgard/midgard_address.c diff --git a/src/panfrost/midgard/midgard_address.c b/src/panfrost/midgard/midgard_address.c new file mode 100644 index 00000000000..7b8e01811d2 --- /dev/null +++ b/src/panfrost/midgard/midgard_address.c @@ -0,0 +1,205 @@ +/* + * Copyright (C) 2019 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. + * + * Authors (Collabora): + * Alyssa Rosenzweig + */ + +#include "compiler.h" + +/* Midgard's generic load/store instructions, particularly to implement SSBOs + * and globals, have support for address arithmetic natively. In particularly, + * they take two indirect arguments A, B and two immediates #s, #c, calculating + * the address: + * + * A + (zext?(B) << #s) + #c + * + * This allows for fast indexing into arrays. This file tries to pattern match the offset in NIR with this form to reduce pressure on the ALU pipe. + */ + +struct mir_address { + nir_ssa_scalar A; + nir_ssa_scalar B; + + bool zext; + unsigned shift; + unsigned bias; +}; + +/* Matches a constant in either slot and moves it to the bias */ + +static void +mir_match_constant(struct mir_address *address) +{ + if (address->A.def && nir_ssa_scalar_is_const(address->A)) { + address->bias += nir_ssa_scalar_as_uint(address->A); + address->A.def = NULL; + } + + if (address->B.def && nir_ssa_scalar_is_const(address->B)) { + address->bias += nir_ssa_scalar_as_uint(address->B); + address->B.def = NULL; + } +} + +/* Matches an iadd when there is a free slot or constant */ + +static void +mir_match_iadd(struct mir_address *address, bool first_free) +{ + if (!address->B.def || !nir_ssa_scalar_is_alu(address->B)) + return; + + nir_op op = nir_ssa_scalar_alu_op(address->B); + + if (op != nir_op_iadd) return; + + nir_ssa_scalar op1 = nir_ssa_scalar_chase_alu_src(address->B, 0); + nir_ssa_scalar op2 = nir_ssa_scalar_chase_alu_src(address->B, 1); + + if (nir_ssa_scalar_is_const(op1)) { + address->bias += nir_ssa_scalar_as_uint(op1); + address->B = op2; + } else if (nir_ssa_scalar_is_const(op2)) { + address->bias += nir_ssa_scalar_as_uint(op2); + address->B = op1; + } else if (first_free && !address->A.def) { + address->A = op1; + address->B = op2; + } +} + +/* Matches u2u64 and sets zext */ + +static void +mir_match_u2u64(struct mir_address *address) +{ + if (!address->B.def || !nir_ssa_scalar_is_alu(address->B)) + return; + + nir_op op = nir_ssa_scalar_alu_op(address->B); + if (op != nir_op_u2u64) return; + nir_ssa_scalar arg = nir_ssa_scalar_chase_alu_src(address->B, 0); + + address->B = arg; + address->zext = true; +} + +/* Matches ishl to shift */ + +static void +mir_match_ishl(struct mir_address *address) +{ + if (!address->B.def || !nir_ssa_scalar_is_alu(address->B)) + return; + + nir_op op = nir_ssa_scalar_alu_op(address->B); + if (op != nir_op_ishl) return; + nir_ssa_scalar op1 = nir_ssa_scalar_chase_alu_src(address->B, 0); + nir_ssa_scalar op2 = nir_ssa_scalar_chase_alu_src(address->B, 1); + + if (!nir_ssa_scalar_is_const(op2)) return; + + unsigned shift = nir_ssa_scalar_as_uint(op2); + if (shift > 0x7) return; + + address->B = op1; + address->shift = shift; +} + +/* Strings through mov which can happen from NIR vectorization */ + +static void +mir_match_mov(struct mir_address *address) +{ + if (address->A.def && nir_ssa_scalar_is_alu(address->A)) { + nir_op op = nir_ssa_scalar_alu_op(address->A); + + if (op == nir_op_mov) + address->A = nir_ssa_scalar_chase_alu_src(address->A, 0); + } + + if (address->B.def && nir_ssa_scalar_is_alu(address->B)) { + nir_op op = nir_ssa_scalar_alu_op(address->B); + + if (op == nir_op_mov) + address->B = nir_ssa_scalar_chase_alu_src(address->B, 0); + } +} + +/* Tries to pattern match into mir_address */ + +static struct mir_address +mir_match_offset(nir_ssa_def *offset, bool first_free) +{ + struct mir_address address = { + .B = { .def = offset } + }; + + mir_match_mov(&address); + mir_match_constant(&address); + mir_match_mov(&address); + mir_match_iadd(&address, first_free); + mir_match_mov(&address); + mir_match_u2u64(&address); + mir_match_mov(&address); + mir_match_ishl(&address); + + return address; +} + +void +mir_set_offset(compiler_context *ctx, midgard_instruction *ins, nir_src *offset, bool is_shared) +{ + for(unsigned i = 0; i < 16; ++i) { + ins->swizzle[1][i] = 0; + ins->swizzle[2][i] = 0; + } + + if (!offset->is_ssa) { + ins->load_store.arg_1 |= is_shared ? 0x6E : 0x7E; + ins->src[2] = nir_src_index(ctx, offset); + return; + } + + struct mir_address match = mir_match_offset(offset->ssa, !is_shared); + + if (match.A.def) { + ins->src[1] = nir_ssa_index(match.A.def); + ins->swizzle[1][0] = match.A.comp; + } else + ins->load_store.arg_1 |= is_shared ? 0x6E : 0x7E; + + if (match.B.def) { + ins->src[2] = nir_ssa_index(match.B.def); + ins->swizzle[2][0] = match.B.comp; + } else + ins->load_store.arg_2 = 0x1E; + + if (match.zext) + ins->load_store.arg_1 |= 0x80; + + assert(match.shift <= 7); + ins->load_store.arg_2 |= (match.shift) << 5; + + ins->constants.u32[0] = match.bias; +} -- 2.30.2