freedreno/ir3: limit # of tex prefetch by shader size

[mesa.git] / src / freedreno / ir3 / ir3_context.c

/*
 * Copyright (C) 2015-2018 Rob Clark <robclark@freedesktop.org>
 *
 * 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:
 *    Rob Clark <robclark@freedesktop.org>
 */

#include "ir3_compiler.h"
#include "ir3_context.h"
#include "ir3_image.h"
#include "ir3_shader.h"
#include "ir3_nir.h"

struct ir3_context *
ir3_context_init(struct ir3_compiler *compiler,
                struct ir3_shader_variant *so)
{
        struct ir3_context *ctx = rzalloc(NULL, struct ir3_context);

        if (compiler->gpu_id >= 400) {
                if (so->type == MESA_SHADER_VERTEX) {
                        ctx->astc_srgb = so->key.vastc_srgb;
                } else if (so->type == MESA_SHADER_FRAGMENT) {
                        ctx->astc_srgb = so->key.fastc_srgb;
                }

        } else {
                if (so->type == MESA_SHADER_VERTEX) {
                        ctx->samples = so->key.vsamples;
                } else if (so->type == MESA_SHADER_FRAGMENT) {
                        ctx->samples = so->key.fsamples;
                }
        }

        if (compiler->gpu_id >= 600) {
                ctx->funcs = &ir3_a6xx_funcs;
        } else if (compiler->gpu_id >= 400) {
                ctx->funcs = &ir3_a4xx_funcs;
        }

        ctx->compiler = compiler;
        ctx->so = so;
        ctx->def_ht = _mesa_hash_table_create(ctx,
                        _mesa_hash_pointer, _mesa_key_pointer_equal);
        ctx->block_ht = _mesa_hash_table_create(ctx,
                        _mesa_hash_pointer, _mesa_key_pointer_equal);
        ctx->sel_cond_conversions = _mesa_hash_table_create(ctx,
                        _mesa_hash_pointer, _mesa_key_pointer_equal);

        /* TODO: maybe generate some sort of bitmask of what key
         * lowers vs what shader has (ie. no need to lower
         * texture clamp lowering if no texture sample instrs)..
         * although should be done further up the stack to avoid
         * creating duplicate variants..
         */

        ctx->s = nir_shader_clone(ctx, so->shader->nir);
        if (ir3_key_lowers_nir(&so->key))
                ir3_optimize_nir(so->shader, ctx->s, &so->key);

        /* this needs to be the last pass run, so do this here instead of
         * in ir3_optimize_nir():
         */
        bool progress = false;
        NIR_PASS(progress, ctx->s, nir_lower_locals_to_regs);

        /* we could need cleanup after lower_locals_to_regs */
        while (progress) {
                progress = false;
                NIR_PASS(progress, ctx->s, nir_opt_algebraic);
                NIR_PASS(progress, ctx->s, nir_opt_constant_folding);
        }

        /* We want to lower nir_op_imul as late as possible, to catch also
         * those generated by earlier passes (e.g, nir_lower_locals_to_regs).
         * However, we want a final swing of a few passes to have a chance
         * at optimizing the result.
         */
        progress = false;
        NIR_PASS(progress, ctx->s, ir3_nir_lower_imul);
        while (progress) {
                progress = false;
                NIR_PASS(progress, ctx->s, nir_opt_algebraic);
                NIR_PASS(progress, ctx->s, nir_opt_copy_prop_vars);
                NIR_PASS(progress, ctx->s, nir_opt_dead_write_vars);
                NIR_PASS(progress, ctx->s, nir_opt_dce);
                NIR_PASS(progress, ctx->s, nir_opt_constant_folding);
        }

        /* Enable the texture pre-fetch feature only a4xx onwards.  But
         * only enable it on generations that have been tested:
         */
        if ((so->type == MESA_SHADER_FRAGMENT) && (compiler->gpu_id >= 600))
                NIR_PASS_V(ctx->s, ir3_nir_lower_tex_prefetch);

        NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);

        /* Super crude heuristic to limit # of tex prefetch in small
         * shaders.  This completely ignores loops.. but that's really
         * not the worst of it's problems.  (A frag shader that has
         * loops is probably going to be big enough to not trigger a
         * lower threshold.)
         *
         *   1) probably want to do this in terms of ir3 instructions
         *   2) probably really want to decide this after scheduling
         *      (or at least pre-RA sched) so we have a rough idea about
         *      nops, and don't count things that get cp'd away
         *   3) blob seems to use higher thresholds with a mix of more
         *      SFU instructions.  Which partly makes sense, more SFU
         *      instructions probably means you want to get the real
         *      shader started sooner, but that considers where in the
         *      shader the SFU instructions are, which blob doesn't seem
         *      to do.
         *
         * This uses more conservative thresholds assuming a more alu
         * than sfu heavy instruction mix.
         */
        if (so->type == MESA_SHADER_FRAGMENT) {
                nir_function_impl *fxn = nir_shader_get_entrypoint(ctx->s);

                unsigned instruction_count = 0;
                nir_foreach_block (block, fxn) {
                        instruction_count += exec_list_length(&block->instr_list);
                }

                if (instruction_count < 50) {
                        ctx->prefetch_limit = 2;
                } else if (instruction_count < 70) {
                        ctx->prefetch_limit = 3;
                } else {
                        ctx->prefetch_limit = IR3_MAX_SAMPLER_PREFETCH;
                }
        }

        if (shader_debug_enabled(so->type)) {
                fprintf(stdout, "NIR (final form) for %s shader %s:\n",
                        ir3_shader_stage(so), so->shader->nir->info.name);
                nir_print_shader(ctx->s, stdout);
        }

        ir3_ibo_mapping_init(&so->image_mapping, ctx->s->info.num_textures);

        return ctx;
}

void
ir3_context_free(struct ir3_context *ctx)
{
        ralloc_free(ctx);
}

/*
 * Misc helpers
 */

/* allocate a n element value array (to be populated by caller) and
 * insert in def_ht
 */
struct ir3_instruction **
ir3_get_dst_ssa(struct ir3_context *ctx, nir_ssa_def *dst, unsigned n)
{
        struct ir3_instruction **value =
                ralloc_array(ctx->def_ht, struct ir3_instruction *, n);
        _mesa_hash_table_insert(ctx->def_ht, dst, value);
        return value;
}

struct ir3_instruction **
ir3_get_dst(struct ir3_context *ctx, nir_dest *dst, unsigned n)
{
        struct ir3_instruction **value;

        if (dst->is_ssa) {
                value = ir3_get_dst_ssa(ctx, &dst->ssa, n);
        } else {
                value = ralloc_array(ctx, struct ir3_instruction *, n);
        }

        /* NOTE: in non-ssa case, we don't really need to store last_dst
         * but this helps us catch cases where put_dst() call is forgotten
         */
        compile_assert(ctx, !ctx->last_dst);
        ctx->last_dst = value;
        ctx->last_dst_n = n;

        return value;
}

struct ir3_instruction * const *
ir3_get_src(struct ir3_context *ctx, nir_src *src)
{
        if (src->is_ssa) {
                struct hash_entry *entry;
                entry = _mesa_hash_table_search(ctx->def_ht, src->ssa);
                compile_assert(ctx, entry);
                return entry->data;
        } else {
                nir_register *reg = src->reg.reg;
                struct ir3_array *arr = ir3_get_array(ctx, reg);
                unsigned num_components = arr->r->num_components;
                struct ir3_instruction *addr = NULL;
                struct ir3_instruction **value =
                        ralloc_array(ctx, struct ir3_instruction *, num_components);

                if (src->reg.indirect)
                        addr = ir3_get_addr0(ctx, ir3_get_src(ctx, src->reg.indirect)[0],
                                        reg->num_components);

                for (unsigned i = 0; i < num_components; i++) {
                        unsigned n = src->reg.base_offset * reg->num_components + i;
                        compile_assert(ctx, n < arr->length);
                        value[i] = ir3_create_array_load(ctx, arr, n, addr, reg->bit_size);
                }

                return value;
        }
}

void
ir3_put_dst(struct ir3_context *ctx, nir_dest *dst)
{
        unsigned bit_size = nir_dest_bit_size(*dst);

        /* add extra mov if dst value is HIGH reg.. in some cases not all
         * instructions can read from HIGH regs, in cases where they can
         * ir3_cp will clean up the extra mov:
         */
        for (unsigned i = 0; i < ctx->last_dst_n; i++) {
                if (!ctx->last_dst[i])
                        continue;
                if (ctx->last_dst[i]->regs[0]->flags & IR3_REG_HIGH) {
                        ctx->last_dst[i] = ir3_MOV(ctx->block, ctx->last_dst[i], TYPE_U32);
                }
        }

        /* Note: 1-bit bools are stored in 32-bit regs */
        if (bit_size == 16) {
                for (unsigned i = 0; i < ctx->last_dst_n; i++) {
                        struct ir3_instruction *dst = ctx->last_dst[i];
                        dst->regs[0]->flags |= IR3_REG_HALF;
                        if (ctx->last_dst[i]->opc == OPC_META_SPLIT)
                                dst->regs[1]->instr->regs[0]->flags |= IR3_REG_HALF;
                }
        }

        if (!dst->is_ssa) {
                nir_register *reg = dst->reg.reg;
                struct ir3_array *arr = ir3_get_array(ctx, reg);
                unsigned num_components = ctx->last_dst_n;
                struct ir3_instruction *addr = NULL;

                if (dst->reg.indirect)
                        addr = ir3_get_addr0(ctx, ir3_get_src(ctx, dst->reg.indirect)[0],
                                        reg->num_components);

                for (unsigned i = 0; i < num_components; i++) {
                        unsigned n = dst->reg.base_offset * reg->num_components + i;
                        compile_assert(ctx, n < arr->length);
                        if (!ctx->last_dst[i])
                                continue;
                        ir3_create_array_store(ctx, arr, n, ctx->last_dst[i], addr);
                }

                ralloc_free(ctx->last_dst);
        }

        ctx->last_dst = NULL;
        ctx->last_dst_n = 0;
}

static unsigned
dest_flags(struct ir3_instruction *instr)
{
        return instr->regs[0]->flags & (IR3_REG_HALF | IR3_REG_HIGH);
}

struct ir3_instruction *
ir3_create_collect(struct ir3_context *ctx, struct ir3_instruction *const *arr,
                unsigned arrsz)
{
        struct ir3_block *block = ctx->block;
        struct ir3_instruction *collect;

        if (arrsz == 0)
                return NULL;

        unsigned flags = dest_flags(arr[0]);

        collect = ir3_instr_create2(block, OPC_META_COLLECT, 1 + arrsz);
        __ssa_dst(collect)->flags |= flags;
        for (unsigned i = 0; i < arrsz; i++) {
                struct ir3_instruction *elem = arr[i];

                /* Since arrays are pre-colored in RA, we can't assume that
                 * things will end up in the right place.  (Ie. if a collect
                 * joins elements from two different arrays.)  So insert an
                 * extra mov.
                 *
                 * We could possibly skip this if all the collected elements
                 * are contiguous elements in a single array.. not sure how
                 * likely that is to happen.
                 *
                 * Fixes a problem with glamor shaders, that in effect do
                 * something like:
                 *
                 *   if (foo)
                 *     texcoord = ..
                 *   else
                 *     texcoord = ..
                 *   color = texture2D(tex, texcoord);
                 *
                 * In this case, texcoord will end up as nir registers (which
                 * translate to ir3 array's of length 1.  And we can't assume
                 * the two (or more) arrays will get allocated in consecutive
                 * scalar registers.
                 *
                 */
                if (elem->regs[0]->flags & IR3_REG_ARRAY) {
                        type_t type = (flags & IR3_REG_HALF) ? TYPE_U16 : TYPE_U32;
                        elem = ir3_MOV(block, elem, type);
                }

                compile_assert(ctx, dest_flags(elem) == flags);
                __ssa_src(collect, elem, flags);
        }

        collect->regs[0]->wrmask = MASK(arrsz);

        return collect;
}

/* helper for instructions that produce multiple consecutive scalar
 * outputs which need to have a split meta instruction inserted
 */
void
ir3_split_dest(struct ir3_block *block, struct ir3_instruction **dst,
                struct ir3_instruction *src, unsigned base, unsigned n)
{
        struct ir3_instruction *prev = NULL;

        if ((n == 1) && (src->regs[0]->wrmask == 0x1)) {
                dst[0] = src;
                return;
        }

        if (src->opc == OPC_META_COLLECT) {
                debug_assert((base + n) < src->regs_count);

                for (int i = 0; i < n; i++) {
                        dst[i] = ssa(src->regs[i + base + 1]);
                }

                return;
        }

        unsigned flags = dest_flags(src);

        for (int i = 0, j = 0; i < n; i++) {
                struct ir3_instruction *split =
                                ir3_instr_create(block, OPC_META_SPLIT);
                __ssa_dst(split)->flags |= flags;
                __ssa_src(split, src, flags);
                split->split.off = i + base;

                if (prev) {
                        split->cp.left = prev;
                        split->cp.left_cnt++;
                        prev->cp.right = split;
                        prev->cp.right_cnt++;
                }
                prev = split;

                if (src->regs[0]->wrmask & (1 << (i + base)))
                        dst[j++] = split;
        }
}

NORETURN void
ir3_context_error(struct ir3_context *ctx, const char *format, ...)
{
        struct hash_table *errors = NULL;
        va_list ap;
        va_start(ap, format);
        if (ctx->cur_instr) {
                errors = _mesa_hash_table_create(NULL,
                                _mesa_hash_pointer,
                                _mesa_key_pointer_equal);
                char *msg = ralloc_vasprintf(errors, format, ap);
                _mesa_hash_table_insert(errors, ctx->cur_instr, msg);
        } else {
                _debug_vprintf(format, ap);
        }
        va_end(ap);
        nir_print_shader_annotated(ctx->s, stdout, errors);
        ralloc_free(errors);
        ctx->error = true;
        unreachable("");
}

static struct ir3_instruction *
create_addr0(struct ir3_block *block, struct ir3_instruction *src, int align)
{
        struct ir3_instruction *instr, *immed;

        /* TODO in at least some cases, the backend could probably be
         * made clever enough to propagate IR3_REG_HALF..
         */
        instr = ir3_COV(block, src, TYPE_U32, TYPE_S16);
        instr->regs[0]->flags |= IR3_REG_HALF;

        switch(align){
        case 1:
                /* src *= 1: */
                break;
        case 2:
                /* src *= 2     => src <<= 1: */
                immed = create_immed(block, 1);
                immed->regs[0]->flags |= IR3_REG_HALF;

                instr = ir3_SHL_B(block, instr, 0, immed, 0);
                instr->regs[0]->flags |= IR3_REG_HALF;
                instr->regs[1]->flags |= IR3_REG_HALF;
                break;
        case 3:
                /* src *= 3: */
                immed = create_immed(block, 3);
                immed->regs[0]->flags |= IR3_REG_HALF;

                instr = ir3_MULL_U(block, instr, 0, immed, 0);
                instr->regs[0]->flags |= IR3_REG_HALF;
                instr->regs[1]->flags |= IR3_REG_HALF;
                break;
        case 4:
                /* src *= 4 => src <<= 2: */
                immed = create_immed(block, 2);
                immed->regs[0]->flags |= IR3_REG_HALF;

                instr = ir3_SHL_B(block, instr, 0, immed, 0);
                instr->regs[0]->flags |= IR3_REG_HALF;
                instr->regs[1]->flags |= IR3_REG_HALF;
                break;
        default:
                unreachable("bad align");
                return NULL;
        }

        instr = ir3_MOV(block, instr, TYPE_S16);
        instr->regs[0]->num = regid(REG_A0, 0);
        instr->regs[0]->flags &= ~IR3_REG_SSA;
        instr->regs[0]->flags |= IR3_REG_HALF;
        instr->regs[1]->flags |= IR3_REG_HALF;

        return instr;
}

static struct ir3_instruction *
create_addr1(struct ir3_block *block, unsigned const_val)
{

        struct ir3_instruction *immed = create_immed(block, const_val);
        struct ir3_instruction *instr = ir3_MOV(block, immed, TYPE_S16);
        instr->regs[0]->num = regid(REG_A0, 1);
        instr->regs[0]->flags &= ~IR3_REG_SSA;
        instr->regs[0]->flags |= IR3_REG_HALF;
        instr->regs[1]->flags |= IR3_REG_HALF;
        return instr;
}

/* caches addr values to avoid generating multiple cov/shl/mova
 * sequences for each use of a given NIR level src as address
 */
struct ir3_instruction *
ir3_get_addr0(struct ir3_context *ctx, struct ir3_instruction *src, int align)
{
        struct ir3_instruction *addr;
        unsigned idx = align - 1;

        compile_assert(ctx, idx < ARRAY_SIZE(ctx->addr0_ht));

        if (!ctx->addr0_ht[idx]) {
                ctx->addr0_ht[idx] = _mesa_hash_table_create(ctx,
                                _mesa_hash_pointer, _mesa_key_pointer_equal);
        } else {
                struct hash_entry *entry;
                entry = _mesa_hash_table_search(ctx->addr0_ht[idx], src);
                if (entry)
                        return entry->data;
        }

        addr = create_addr0(ctx->block, src, align);
        _mesa_hash_table_insert(ctx->addr0_ht[idx], src, addr);

        return addr;
}

/* Similar to ir3_get_addr0, but for a1.x. */
struct ir3_instruction *
ir3_get_addr1(struct ir3_context *ctx, unsigned const_val)
{
        struct ir3_instruction *addr;

        if (!ctx->addr1_ht) {
                ctx->addr1_ht = _mesa_hash_table_u64_create(ctx);
        } else {
                addr = _mesa_hash_table_u64_search(ctx->addr1_ht, const_val);
                if (addr)
                        return addr;
        }

        addr = create_addr1(ctx->block, const_val);
        _mesa_hash_table_u64_insert(ctx->addr1_ht, const_val, addr);

        return addr;
}

struct ir3_instruction *
ir3_get_predicate(struct ir3_context *ctx, struct ir3_instruction *src)
{
        struct ir3_block *b = ctx->block;
        struct ir3_instruction *cond;

        /* NOTE: only cmps.*.* can write p0.x: */
        cond = ir3_CMPS_S(b, src, 0, create_immed(b, 0), 0);
        cond->cat2.condition = IR3_COND_NE;

        /* condition always goes in predicate register: */
        cond->regs[0]->num = regid(REG_P0, 0);
        cond->regs[0]->flags &= ~IR3_REG_SSA;

        return cond;
}

/*
 * Array helpers
 */

void
ir3_declare_array(struct ir3_context *ctx, nir_register *reg)
{
        struct ir3_array *arr = rzalloc(ctx, struct ir3_array);
        arr->id = ++ctx->num_arrays;
        /* NOTE: sometimes we get non array regs, for example for arrays of
         * length 1.  See fs-const-array-of-struct-of-array.shader_test.  So
         * treat a non-array as if it was an array of length 1.
         *
         * It would be nice if there was a nir pass to convert arrays of
         * length 1 to ssa.
         */
        arr->length = reg->num_components * MAX2(1, reg->num_array_elems);
        compile_assert(ctx, arr->length > 0);
        arr->r = reg;
        list_addtail(&arr->node, &ctx->ir->array_list);
}

struct ir3_array *
ir3_get_array(struct ir3_context *ctx, nir_register *reg)
{
        foreach_array (arr, &ctx->ir->array_list) {
                if (arr->r == reg)
                        return arr;
        }
        ir3_context_error(ctx, "bogus reg: %s\n", reg->name);
        return NULL;
}

/* relative (indirect) if address!=NULL */
struct ir3_instruction *
ir3_create_array_load(struct ir3_context *ctx, struct ir3_array *arr, int n,
                struct ir3_instruction *address, unsigned bitsize)
{
        struct ir3_block *block = ctx->block;
        struct ir3_instruction *mov;
        struct ir3_register *src;
        unsigned flags = 0;

        mov = ir3_instr_create(block, OPC_MOV);
        if (bitsize == 16) {
                mov->cat1.src_type = TYPE_U16;
                mov->cat1.dst_type = TYPE_U16;
                flags |= IR3_REG_HALF;
                arr->half = true;
        } else {
                mov->cat1.src_type = TYPE_U32;
                mov->cat1.dst_type = TYPE_U32;
        }

        mov->barrier_class = IR3_BARRIER_ARRAY_R;
        mov->barrier_conflict = IR3_BARRIER_ARRAY_W;
        __ssa_dst(mov)->flags |= flags;
        src = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
                        COND(address, IR3_REG_RELATIV) | flags);
        src->instr = arr->last_write;
        src->size  = arr->length;
        src->array.id = arr->id;
        src->array.offset = n;

        if (address)
                ir3_instr_set_address(mov, address);

        return mov;
}

/* relative (indirect) if address!=NULL */
void
ir3_create_array_store(struct ir3_context *ctx, struct ir3_array *arr, int n,
                struct ir3_instruction *src, struct ir3_instruction *address)
{
        struct ir3_block *block = ctx->block;
        struct ir3_instruction *mov;
        struct ir3_register *dst;

        /* if not relative store, don't create an extra mov, since that
         * ends up being difficult for cp to remove.
         *
         * Also, don't skip the mov if the src is meta (like fanout/split),
         * since that creates a situation that RA can't really handle properly.
         */
        if (!address && !is_meta(src)) {
                dst = src->regs[0];

                src->barrier_class |= IR3_BARRIER_ARRAY_W;
                src->barrier_conflict |= IR3_BARRIER_ARRAY_R | IR3_BARRIER_ARRAY_W;

                dst->flags |= IR3_REG_ARRAY;
                dst->instr = arr->last_write;
                dst->size = arr->length;
                dst->array.id = arr->id;
                dst->array.offset = n;

                arr->last_write = src;

                array_insert(block, block->keeps, src);

                return;
        }

        mov = ir3_instr_create(block, OPC_MOV);
        mov->cat1.src_type = TYPE_U32;
        mov->cat1.dst_type = TYPE_U32;
        mov->barrier_class = IR3_BARRIER_ARRAY_W;
        mov->barrier_conflict = IR3_BARRIER_ARRAY_R | IR3_BARRIER_ARRAY_W;
        dst = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
                        COND(address, IR3_REG_RELATIV));
        dst->instr = arr->last_write;
        dst->size  = arr->length;
        dst->array.id = arr->id;
        dst->array.offset = n;
        ir3_reg_create(mov, 0, IR3_REG_SSA)->instr = src;

        if (address)
                ir3_instr_set_address(mov, address);

        arr->last_write = mov;

        /* the array store may only matter to something in an earlier
         * block (ie. loops), but since arrays are not in SSA, depth
         * pass won't know this.. so keep all array stores:
         */
        array_insert(block, block->keeps, mov);
}