From 14dc281c1332518b6144718e1fb3845abbe23ff7 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Thu, 19 Feb 2015 12:58:53 -0800 Subject: [PATCH] vc4: Enforce one-uniform-per-instruction after optimization. This lets us more intelligently decide which uniform values should be put into temporaries, by choosing the most reused values to push to temps first. total uniforms in shared programs: 13457 -> 13433 (-0.18%) uniforms in affected programs: 1524 -> 1500 (-1.57%) total instructions in shared programs: 40198 -> 40019 (-0.45%) instructions in affected programs: 6027 -> 5848 (-2.97%) I noticed this opportunity because with the NIR work, some programs were happening to make different uniform copy propagation choices that significantly increased instruction counts. --- src/gallium/drivers/vc4/Makefile.sources | 1 + src/gallium/drivers/vc4/vc4_opt_copy_propagation.c | 15 +- src/gallium/drivers/vc4/vc4_program.c | 50 ++---- src/gallium/drivers/vc4/vc4_qir.c | 15 +- src/gallium/drivers/vc4/vc4_qir.h | 2 + src/gallium/drivers/vc4/vc4_qir_lower_uniforms.c | 176 +++++++++++++++++++++ 6 files changed, 209 insertions(+), 50 deletions(-) create mode 100644 src/gallium/drivers/vc4/vc4_qir_lower_uniforms.c diff --git a/src/gallium/drivers/vc4/Makefile.sources b/src/gallium/drivers/vc4/Makefile.sources index 95f1a340ab3..c7254ea1473 100644 --- a/src/gallium/drivers/vc4/Makefile.sources +++ b/src/gallium/drivers/vc4/Makefile.sources @@ -20,6 +20,7 @@ C_SOURCES := \ vc4_packet.h \ vc4_program.c \ vc4_qir.c \ + vc4_qir_lower_uniforms.c \ vc4_qir.h \ vc4_qpu.c \ vc4_qpu_defines.h \ diff --git a/src/gallium/drivers/vc4/vc4_opt_copy_propagation.c b/src/gallium/drivers/vc4/vc4_opt_copy_propagation.c index 07e1cb14b0d..f8c49a44bd3 100644 --- a/src/gallium/drivers/vc4/vc4_opt_copy_propagation.c +++ b/src/gallium/drivers/vc4/vc4_opt_copy_propagation.c @@ -49,22 +49,11 @@ qir_opt_copy_propagation(struct vc4_compile *c) if (inst->dst.file == QFILE_TEMP) defs[inst->dst.index] = inst; - /* A single instruction can only read one uniform value. (It - * could maybe read the same uniform value in two operands, - * but that doesn't seem important to do). - */ - bool reads_a_uniform = false; - for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) { - if (inst->src[i].file == QFILE_UNIF) - reads_a_uniform = true; - } - for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) { int index = inst->src[i].index; if (inst->src[i].file == QFILE_TEMP && (movs[index].file == QFILE_TEMP || - (movs[index].file == QFILE_UNIF && - !reads_a_uniform))) { + movs[index].file == QFILE_UNIF)) { if (debug) { fprintf(stderr, "Copy propagate: "); qir_dump_inst(c, inst); @@ -72,8 +61,6 @@ qir_opt_copy_propagation(struct vc4_compile *c) } inst->src[i] = movs[index]; - if (movs[index].file == QFILE_UNIF) - reads_a_uniform = true; if (debug) { fprintf(stderr, "to: "); diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c index a1d9a7f064c..b904679ef00 100644 --- a/src/gallium/drivers/vc4/vc4_program.c +++ b/src/gallium/drivers/vc4/vc4_program.c @@ -143,24 +143,15 @@ qir_uniform(struct vc4_compile *c, } static struct qreg -get_temp_for_uniform(struct vc4_compile *c, enum quniform_contents contents, - uint32_t data) -{ - struct qreg u = qir_uniform(c, contents, data); - struct qreg t = qir_MOV(c, u); - return t; -} - -static struct qreg qir_uniform_ui(struct vc4_compile *c, uint32_t ui) { - return get_temp_for_uniform(c, QUNIFORM_CONSTANT, ui); + return qir_uniform(c, QUNIFORM_CONSTANT, ui); } static struct qreg qir_uniform_f(struct vc4_compile *c, float f) { - return qir_uniform_ui(c, fui(f)); + return qir_uniform(c, QUNIFORM_CONSTANT, fui(f)); } static struct qreg @@ -232,8 +223,7 @@ get_src(struct vc4_compile *c, unsigned tgsi_op, if (src->Indirect) { r = indirect_uniform_load(c, full_src, s); } else { - r = get_temp_for_uniform(c, QUNIFORM_UNIFORM, - src->Index * 4 + s); + r = qir_uniform(c, QUNIFORM_UNIFORM, src->Index * 4 + s); } break; case TGSI_FILE_INPUT: @@ -660,13 +650,9 @@ tgsi_to_qir_tex(struct vc4_compile *c, if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_RECT || tgsi_inst->Texture.Texture == TGSI_TEXTURE_SHADOWRECT) { s = qir_FMUL(c, s, - get_temp_for_uniform(c, - QUNIFORM_TEXRECT_SCALE_X, - unit)); + qir_uniform(c, QUNIFORM_TEXRECT_SCALE_X, unit)); t = qir_FMUL(c, t, - get_temp_for_uniform(c, - QUNIFORM_TEXRECT_SCALE_Y, - unit)); + qir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y, unit)); } if (tgsi_inst->Texture.Texture == TGSI_TEXTURE_CUBE || @@ -689,7 +675,7 @@ tgsi_to_qir_tex(struct vc4_compile *c, c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP || c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER || c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) { - qir_TEX_R(c, get_temp_for_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit), + qir_TEX_R(c, qir_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit), texture_u[next_texture_u++]); } @@ -1504,14 +1490,11 @@ vc4_blend_channel(struct vc4_compile *c, } case PIPE_BLENDFACTOR_CONST_COLOR: return qir_FMUL(c, val, - get_temp_for_uniform(c, - QUNIFORM_BLEND_CONST_COLOR, - channel)); + qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR, + channel)); case PIPE_BLENDFACTOR_CONST_ALPHA: return qir_FMUL(c, val, - get_temp_for_uniform(c, - QUNIFORM_BLEND_CONST_COLOR, - 3)); + qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR, 3)); case PIPE_BLENDFACTOR_ZERO: return qir_uniform_f(c, 0.0); case PIPE_BLENDFACTOR_INV_SRC_COLOR: @@ -1529,15 +1512,15 @@ vc4_blend_channel(struct vc4_compile *c, case PIPE_BLENDFACTOR_INV_CONST_COLOR: return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0), - get_temp_for_uniform(c, - QUNIFORM_BLEND_CONST_COLOR, - channel))); + qir_uniform(c, + QUNIFORM_BLEND_CONST_COLOR, + channel))); case PIPE_BLENDFACTOR_INV_CONST_ALPHA: return qir_FMUL(c, val, qir_FSUB(c, qir_uniform_f(c, 1.0), - get_temp_for_uniform(c, - QUNIFORM_BLEND_CONST_COLOR, - 3))); + qir_uniform(c, + QUNIFORM_BLEND_CONST_COLOR, + 3))); default: case PIPE_BLENDFACTOR_SRC1_COLOR: @@ -1661,7 +1644,7 @@ static void alpha_test_discard(struct vc4_compile *c) { struct qreg src_alpha; - struct qreg alpha_ref = get_temp_for_uniform(c, QUNIFORM_ALPHA_REF, 0); + struct qreg alpha_ref = qir_uniform(c, QUNIFORM_ALPHA_REF, 0); if (!c->fs_key->alpha_test) return; @@ -2171,6 +2154,7 @@ vc4_shader_tgsi_to_qir(struct vc4_context *vc4, enum qstage stage, } qir_optimize(c); + qir_lower_uniforms(c); if (vc4_debug & VC4_DEBUG_QIR) { fprintf(stderr, "%s prog %d/%d QIR:\n", diff --git a/src/gallium/drivers/vc4/vc4_qir.c b/src/gallium/drivers/vc4/vc4_qir.c index 9e0ee1f0ae5..5c1fdbddfb6 100644 --- a/src/gallium/drivers/vc4/vc4_qir.c +++ b/src/gallium/drivers/vc4/vc4_qir.c @@ -174,6 +174,12 @@ qir_is_multi_instruction(struct qinst *inst) } bool +qir_is_tex(struct qinst *inst) +{ + return inst->op >= QOP_TEX_S && inst->op <= QOP_TEX_DIRECT; +} + +bool qir_depends_on_flags(struct qinst *inst) { switch (inst->op) { @@ -420,9 +426,12 @@ qir_get_stage_name(enum qstage stage) void qir_SF(struct vc4_compile *c, struct qreg src) { - assert(!is_empty_list(&c->instructions)); - struct qinst *last_inst = (struct qinst *)c->instructions.prev; - if (last_inst->dst.file != src.file || + struct qinst *last_inst = NULL; + if (!is_empty_list(&c->instructions)) + last_inst = (struct qinst *)c->instructions.prev; + + if (!last_inst || + last_inst->dst.file != src.file || last_inst->dst.index != src.index || qir_is_multi_instruction(last_inst)) { src = qir_MOV(c, src); diff --git a/src/gallium/drivers/vc4/vc4_qir.h b/src/gallium/drivers/vc4/vc4_qir.h index 6da6ff6542e..a1b55605584 100644 --- a/src/gallium/drivers/vc4/vc4_qir.h +++ b/src/gallium/drivers/vc4/vc4_qir.h @@ -376,6 +376,7 @@ bool qir_reg_equals(struct qreg a, struct qreg b); bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst); bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst); bool qir_is_multi_instruction(struct qinst *inst); +bool qir_is_tex(struct qinst *inst); bool qir_depends_on_flags(struct qinst *inst); bool qir_writes_r4(struct qinst *inst); bool qir_reads_r4(struct qinst *inst); @@ -393,6 +394,7 @@ bool qir_opt_cse(struct vc4_compile *c); bool qir_opt_dead_code(struct vc4_compile *c); bool qir_opt_small_immediates(struct vc4_compile *c); bool qir_opt_vpm_writes(struct vc4_compile *c); +void qir_lower_uniforms(struct vc4_compile *c); void qpu_schedule_instructions(struct vc4_compile *c); diff --git a/src/gallium/drivers/vc4/vc4_qir_lower_uniforms.c b/src/gallium/drivers/vc4/vc4_qir_lower_uniforms.c new file mode 100644 index 00000000000..d527889e76f --- /dev/null +++ b/src/gallium/drivers/vc4/vc4_qir_lower_uniforms.c @@ -0,0 +1,176 @@ +/* + * Copyright © 2014 Broadcom + * + * 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. + */ + +/** + * @file vc4_opt_algebraic.c + * + * This is the optimization pass for miscellaneous changes to instructions + * where we can simplify the operation by some knowledge about the specific + * operations. + * + * Mostly this will be a matter of turning things into MOVs so that they can + * later be copy-propagated out. + */ + +#include "vc4_qir.h" +#include "util/hash_table.h" +#include "util/u_math.h" + +static inline uint32_t +index_hash(const void *key) +{ + return (uintptr_t)key; +} + +static inline bool +index_compare(const void *a, const void *b) +{ + return a == b; +} + +static void +add_uniform(struct hash_table *ht, struct qreg reg) +{ + struct hash_entry *entry; + void *key = (void *)(uintptr_t)reg.index; + + entry = _mesa_hash_table_search(ht, key); + if (entry) { + entry->data++; + } else { + _mesa_hash_table_insert(ht, key, (void *)(uintptr_t)1); + } +} + +static void +remove_uniform(struct hash_table *ht, struct qreg reg) +{ + struct hash_entry *entry; + void *key = (void *)(uintptr_t)reg.index; + + entry = _mesa_hash_table_search(ht, key); + assert(entry); + entry->data--; + if (entry->data == NULL) + _mesa_hash_table_remove(ht, entry); +} + +static bool +is_lowerable_uniform(struct qinst *inst, int i) +{ + if (inst->src[i].file != QFILE_UNIF) + return false; + if (qir_is_tex(inst)) + return i != 1; + return true; +} + +void +qir_lower_uniforms(struct vc4_compile *c) +{ + struct simple_node *node; + struct hash_table *ht = + _mesa_hash_table_create(c, index_hash, index_compare); + + /* Walk the instruction list, finding which instructions have more + * than one uniform referenced, and add those uniform values to the + * ht. + */ + foreach(node, &c->instructions) { + struct qinst *inst = (struct qinst *)node; + uint32_t nsrc = qir_get_op_nsrc(inst->op); + + uint32_t count = 0; + for (int i = 0; i < nsrc; i++) { + if (inst->src[i].file == QFILE_UNIF) + count++; + } + + if (count <= 1) + continue; + + for (int i = 0; i < nsrc; i++) { + if (is_lowerable_uniform(inst, i)) + add_uniform(ht, inst->src[i]); + } + } + + while (ht->entries) { + /* Find the most commonly used uniform in instructions that + * need a uniform lowered. + */ + uint32_t max_count = 0; + uint32_t max_index = 0; + struct hash_entry *entry; + hash_table_foreach(ht, entry) { + uint32_t count = (uintptr_t)entry->data; + uint32_t index = (uintptr_t)entry->key; + if (count > max_count) { + max_count = count; + max_index = index; + } + } + + /* Now, find the instructions using this uniform and make them + * reference a temp instead. + */ + struct qreg temp = qir_get_temp(c); + struct qreg unif = { QFILE_UNIF, max_index }; + struct qinst *mov = qir_inst(QOP_MOV, temp, unif, c->undef); + insert_at_head(&c->instructions, &mov->link); + foreach(node, &c->instructions) { + struct qinst *inst = (struct qinst *)node; + uint32_t nsrc = qir_get_op_nsrc(inst->op); + + uint32_t count = 0; + for (int i = 0; i < nsrc; i++) { + if (inst->src[i].file == QFILE_UNIF) + count++; + } + + if (count <= 1) + continue; + + for (int i = 0; i < nsrc; i++) { + if (is_lowerable_uniform(inst, i) && + inst->src[i].index == max_index) { + inst->src[i] = temp; + remove_uniform(ht, unif); + count--; + } + } + + /* If the instruction doesn't need lowering any more, + * then drop it from the list. + */ + if (count <= 1) { + for (int i = 0; i < nsrc; i++) { + if (is_lowerable_uniform(inst, i)) + remove_uniform(ht, inst->src[i]); + } + } + } + } + + _mesa_hash_table_destroy(ht, NULL); +} -- 2.11.0