KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
mesa/src/broadcom/compiler/vir.c

2310 lines
81 KiB
C
Raw Permalink Blame History

/*
* Copyright © 2016-2017 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.
*/
#include "broadcom/common/v3d_device_info.h"
#include "v3d_compiler.h"
#include "util/u_prim.h"
#include "compiler/nir/nir_schedule.h"
#include "compiler/nir/nir_builder.h"
int
vir_get_nsrc(struct qinst *inst)
{
switch (inst->qpu.type) {
case V3D_QPU_INSTR_TYPE_BRANCH:
return 0;
case V3D_QPU_INSTR_TYPE_ALU:
if (inst->qpu.alu.add.op != V3D_QPU_A_NOP)
return v3d_qpu_add_op_num_src(inst->qpu.alu.add.op);
else
return v3d_qpu_mul_op_num_src(inst->qpu.alu.mul.op);
}
return 0;
}
/**
* Returns whether the instruction has any side effects that must be
* preserved.
*/
bool
vir_has_side_effects(struct v3d_compile *c, struct qinst *inst)
{
switch (inst->qpu.type) {
case V3D_QPU_INSTR_TYPE_BRANCH:
return true;
case V3D_QPU_INSTR_TYPE_ALU:
switch (inst->qpu.alu.add.op) {
case V3D_QPU_A_SETREVF:
case V3D_QPU_A_SETMSF:
case V3D_QPU_A_VPMSETUP:
case V3D_QPU_A_STVPMV:
case V3D_QPU_A_STVPMD:
case V3D_QPU_A_STVPMP:
case V3D_QPU_A_VPMWT:
case V3D_QPU_A_TMUWT:
return true;
default:
break;
}
switch (inst->qpu.alu.mul.op) {
case V3D_QPU_M_MULTOP:
return true;
default:
break;
}
}
if (inst->qpu.sig.ldtmu ||
inst->qpu.sig.ldvary ||
inst->qpu.sig.ldtlbu ||
inst->qpu.sig.ldtlb ||
inst->qpu.sig.wrtmuc ||
inst->qpu.sig.thrsw) {
return true;
}
/* ldunifa works like ldunif: it reads an element and advances the
* pointer, so each read has a side effect (we don't care for ldunif
* because we reconstruct the uniform stream buffer after compiling
* with the surviving uniforms), so allowing DCE to remove
* one would break follow-up loads. We could fix this by emiting a
* unifa for each ldunifa, but each unifa requires 3 delay slots
* before a ldunifa, so that would be quite expensive.
*/
if (inst->qpu.sig.ldunifa || inst->qpu.sig.ldunifarf)
return true;
return false;
}
bool
vir_is_raw_mov(struct qinst *inst)
{
if (inst->qpu.type != V3D_QPU_INSTR_TYPE_ALU ||
(inst->qpu.alu.mul.op != V3D_QPU_M_FMOV &&
inst->qpu.alu.mul.op != V3D_QPU_M_MOV)) {
return false;
}
if (inst->qpu.alu.add.output_pack != V3D_QPU_PACK_NONE ||
inst->qpu.alu.mul.output_pack != V3D_QPU_PACK_NONE) {
return false;
}
if (inst->qpu.alu.add.a_unpack != V3D_QPU_UNPACK_NONE ||
inst->qpu.alu.add.b_unpack != V3D_QPU_UNPACK_NONE ||
inst->qpu.alu.mul.a_unpack != V3D_QPU_UNPACK_NONE ||
inst->qpu.alu.mul.b_unpack != V3D_QPU_UNPACK_NONE) {
return false;
}
if (inst->qpu.flags.ac != V3D_QPU_COND_NONE ||
inst->qpu.flags.mc != V3D_QPU_COND_NONE)
return false;
return true;
}
bool
vir_is_add(struct qinst *inst)
{
return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
inst->qpu.alu.add.op != V3D_QPU_A_NOP);
}
bool
vir_is_mul(struct qinst *inst)
{
return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
inst->qpu.alu.mul.op != V3D_QPU_M_NOP);
}
bool
vir_is_tex(const struct v3d_device_info *devinfo, struct qinst *inst)
{
if (inst->dst.file == QFILE_MAGIC)
return v3d_qpu_magic_waddr_is_tmu(devinfo, inst->dst.index);
if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
inst->qpu.alu.add.op == V3D_QPU_A_TMUWT) {
return true;
}
return false;
}
bool
vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst)
{
for (int i = 0; i < vir_get_nsrc(inst); i++) {
switch (inst->src[i].file) {
case QFILE_VPM:
return true;
default:
break;
}
}
if (devinfo->ver < 41 && (inst->qpu.sig.ldvary ||
inst->qpu.sig.ldtlb ||
inst->qpu.sig.ldtlbu ||
inst->qpu.sig.ldvpm)) {
return true;
}
return false;
}
bool
vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst)
{
switch (inst->dst.file) {
case QFILE_MAGIC:
switch (inst->dst.index) {
case V3D_QPU_WADDR_RECIP:
case V3D_QPU_WADDR_RSQRT:
case V3D_QPU_WADDR_EXP:
case V3D_QPU_WADDR_LOG:
case V3D_QPU_WADDR_SIN:
return true;
}
break;
default:
break;
}
if (devinfo->ver < 41 && inst->qpu.sig.ldtmu)
return true;
return false;
}
void
vir_set_unpack(struct qinst *inst, int src,
enum v3d_qpu_input_unpack unpack)
{
assert(src == 0 || src == 1);
if (vir_is_add(inst)) {
if (src == 0)
inst->qpu.alu.add.a_unpack = unpack;
else
inst->qpu.alu.add.b_unpack = unpack;
} else {
assert(vir_is_mul(inst));
if (src == 0)
inst->qpu.alu.mul.a_unpack = unpack;
else
inst->qpu.alu.mul.b_unpack = unpack;
}
}
void
vir_set_pack(struct qinst *inst, enum v3d_qpu_output_pack pack)
{
if (vir_is_add(inst)) {
inst->qpu.alu.add.output_pack = pack;
} else {
assert(vir_is_mul(inst));
inst->qpu.alu.mul.output_pack = pack;
}
}
void
vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond)
{
if (vir_is_add(inst)) {
inst->qpu.flags.ac = cond;
} else {
assert(vir_is_mul(inst));
inst->qpu.flags.mc = cond;
}
}
enum v3d_qpu_cond
vir_get_cond(struct qinst *inst)
{
assert(inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU);
if (vir_is_add(inst))
return inst->qpu.flags.ac;
else if (vir_is_mul(inst))
return inst->qpu.flags.mc;
else /* NOP */
return V3D_QPU_COND_NONE;
}
void
vir_set_pf(struct v3d_compile *c, struct qinst *inst, enum v3d_qpu_pf pf)
{
c->flags_temp = -1;
if (vir_is_add(inst)) {
inst->qpu.flags.apf = pf;
} else {
assert(vir_is_mul(inst));
inst->qpu.flags.mpf = pf;
}
}
void
vir_set_uf(struct v3d_compile *c, struct qinst *inst, enum v3d_qpu_uf uf)
{
c->flags_temp = -1;
if (vir_is_add(inst)) {
inst->qpu.flags.auf = uf;
} else {
assert(vir_is_mul(inst));
inst->qpu.flags.muf = uf;
}
}
#if 0
uint8_t
vir_channels_written(struct qinst *inst)
{
if (vir_is_mul(inst)) {
switch (inst->dst.pack) {
case QPU_PACK_MUL_NOP:
case QPU_PACK_MUL_8888:
return 0xf;
case QPU_PACK_MUL_8A:
return 0x1;
case QPU_PACK_MUL_8B:
return 0x2;
case QPU_PACK_MUL_8C:
return 0x4;
case QPU_PACK_MUL_8D:
return 0x8;
}
} else {
switch (inst->dst.pack) {
case QPU_PACK_A_NOP:
case QPU_PACK_A_8888:
case QPU_PACK_A_8888_SAT:
case QPU_PACK_A_32_SAT:
return 0xf;
case QPU_PACK_A_8A:
case QPU_PACK_A_8A_SAT:
return 0x1;
case QPU_PACK_A_8B:
case QPU_PACK_A_8B_SAT:
return 0x2;
case QPU_PACK_A_8C:
case QPU_PACK_A_8C_SAT:
return 0x4;
case QPU_PACK_A_8D:
case QPU_PACK_A_8D_SAT:
return 0x8;
case QPU_PACK_A_16A:
case QPU_PACK_A_16A_SAT:
return 0x3;
case QPU_PACK_A_16B:
case QPU_PACK_A_16B_SAT:
return 0xc;
}
}
unreachable("Bad pack field");
}
#endif
struct qreg
vir_get_temp(struct v3d_compile *c)
{
struct qreg reg;
reg.file = QFILE_TEMP;
reg.index = c->num_temps++;
if (c->num_temps > c->defs_array_size) {
uint32_t old_size = c->defs_array_size;
c->defs_array_size = MAX2(old_size * 2, 16);
c->defs = reralloc(c, c->defs, struct qinst *,
c->defs_array_size);
memset(&c->defs[old_size], 0,
sizeof(c->defs[0]) * (c->defs_array_size - old_size));
c->spillable = reralloc(c, c->spillable,
BITSET_WORD,
BITSET_WORDS(c->defs_array_size));
for (int i = old_size; i < c->defs_array_size; i++)
BITSET_SET(c->spillable, i);
}
return reg;
}
struct qinst *
vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst, struct qreg src0, struct qreg src1)
{
struct qinst *inst = calloc(1, sizeof(*inst));
inst->qpu = v3d_qpu_nop();
inst->qpu.alu.add.op = op;
inst->dst = dst;
inst->src[0] = src0;
inst->src[1] = src1;
inst->uniform = ~0;
inst->ip = -1;
return inst;
}
struct qinst *
vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst, struct qreg src0, struct qreg src1)
{
struct qinst *inst = calloc(1, sizeof(*inst));
inst->qpu = v3d_qpu_nop();
inst->qpu.alu.mul.op = op;
inst->dst = dst;
inst->src[0] = src0;
inst->src[1] = src1;
inst->uniform = ~0;
inst->ip = -1;
return inst;
}
struct qinst *
vir_branch_inst(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
{
struct qinst *inst = calloc(1, sizeof(*inst));
inst->qpu = v3d_qpu_nop();
inst->qpu.type = V3D_QPU_INSTR_TYPE_BRANCH;
inst->qpu.branch.cond = cond;
inst->qpu.branch.msfign = V3D_QPU_MSFIGN_NONE;
inst->qpu.branch.bdi = V3D_QPU_BRANCH_DEST_REL;
inst->qpu.branch.ub = true;
inst->qpu.branch.bdu = V3D_QPU_BRANCH_DEST_REL;
inst->dst = vir_nop_reg();
inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, 0);
inst->ip = -1;
return inst;
}
static void
vir_emit(struct v3d_compile *c, struct qinst *inst)
{
inst->ip = -1;
switch (c->cursor.mode) {
case vir_cursor_add:
list_add(&inst->link, c->cursor.link);
break;
case vir_cursor_addtail:
list_addtail(&inst->link, c->cursor.link);
break;
}
c->cursor = vir_after_inst(inst);
c->live_intervals_valid = false;
}
/* Updates inst to write to a new temporary, emits it, and notes the def. */
struct qreg
vir_emit_def(struct v3d_compile *c, struct qinst *inst)
{
assert(inst->dst.file == QFILE_NULL);
/* If we're emitting an instruction that's a def, it had better be
* writing a register.
*/
if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP ||
v3d_qpu_add_op_has_dst(inst->qpu.alu.add.op));
assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP ||
v3d_qpu_mul_op_has_dst(inst->qpu.alu.mul.op));
}
inst->dst = vir_get_temp(c);
if (inst->dst.file == QFILE_TEMP)
c->defs[inst->dst.index] = inst;
vir_emit(c, inst);
return inst->dst;
}
struct qinst *
vir_emit_nondef(struct v3d_compile *c, struct qinst *inst)
{
if (inst->dst.file == QFILE_TEMP)
c->defs[inst->dst.index] = NULL;
vir_emit(c, inst);
return inst;
}
struct qblock *
vir_new_block(struct v3d_compile *c)
{
struct qblock *block = rzalloc(c, struct qblock);
list_inithead(&block->instructions);
block->predecessors = _mesa_set_create(block,
_mesa_hash_pointer,
_mesa_key_pointer_equal);
block->index = c->next_block_index++;
return block;
}
void
vir_set_emit_block(struct v3d_compile *c, struct qblock *block)
{
c->cur_block = block;
c->cursor = vir_after_block(block);
list_addtail(&block->link, &c->blocks);
}
struct qblock *
vir_entry_block(struct v3d_compile *c)
{
return list_first_entry(&c->blocks, struct qblock, link);
}
struct qblock *
vir_exit_block(struct v3d_compile *c)
{
return list_last_entry(&c->blocks, struct qblock, link);
}
void
vir_link_blocks(struct qblock *predecessor, struct qblock *successor)
{
_mesa_set_add(successor->predecessors, predecessor);
if (predecessor->successors[0]) {
assert(!predecessor->successors[1]);
predecessor->successors[1] = successor;
} else {
predecessor->successors[0] = successor;
}
}
const struct v3d_compiler *
v3d_compiler_init(const struct v3d_device_info *devinfo,
uint32_t max_inline_uniform_buffers)
{
struct v3d_compiler *compiler = rzalloc(NULL, struct v3d_compiler);
if (!compiler)
return NULL;
compiler->devinfo = devinfo;
compiler->max_inline_uniform_buffers = max_inline_uniform_buffers;
if (!vir_init_reg_sets(compiler)) {
ralloc_free(compiler);
return NULL;
}
return compiler;
}
void
v3d_compiler_free(const struct v3d_compiler *compiler)
{
ralloc_free((void *)compiler);
}
static struct v3d_compile *
vir_compile_init(const struct v3d_compiler *compiler,
struct v3d_key *key,
nir_shader *s,
void (*debug_output)(const char *msg,
void *debug_output_data),
void *debug_output_data,
int program_id, int variant_id,
uint32_t compile_strategy_idx,
uint32_t max_threads,
uint32_t min_threads_for_reg_alloc,
uint32_t max_tmu_spills,
bool disable_general_tmu_sched,
bool disable_loop_unrolling,
bool disable_constant_ubo_load_sorting,
bool disable_tmu_pipelining,
bool fallback_scheduler)
{
struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
c->compiler = compiler;
c->devinfo = compiler->devinfo;
c->key = key;
c->program_id = program_id;
c->variant_id = variant_id;
c->compile_strategy_idx = compile_strategy_idx;
c->threads = max_threads;
c->debug_output = debug_output;
c->debug_output_data = debug_output_data;
c->compilation_result = V3D_COMPILATION_SUCCEEDED;
c->min_threads_for_reg_alloc = min_threads_for_reg_alloc;
c->max_tmu_spills = max_tmu_spills;
c->fallback_scheduler = fallback_scheduler;
c->disable_general_tmu_sched = disable_general_tmu_sched;
c->disable_tmu_pipelining = disable_tmu_pipelining;
c->disable_constant_ubo_load_sorting = disable_constant_ubo_load_sorting;
c->disable_loop_unrolling = V3D_DEBUG & V3D_DEBUG_NO_LOOP_UNROLL
? true : disable_loop_unrolling;
s = nir_shader_clone(c, s);
c->s = s;
list_inithead(&c->blocks);
vir_set_emit_block(c, vir_new_block(c));
c->output_position_index = -1;
c->output_sample_mask_index = -1;
c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
_mesa_key_pointer_equal);
c->tmu.outstanding_regs = _mesa_pointer_set_create(c);
c->flags_temp = -1;
return c;
}
static int
type_size_vec4(const struct glsl_type *type, bool bindless)
{
return glsl_count_attribute_slots(type, false);
}
static void
v3d_lower_nir(struct v3d_compile *c)
{
struct nir_lower_tex_options tex_options = {
.lower_txd = true,
.lower_tg4_broadcom_swizzle = true,
.lower_rect = false, /* XXX: Use this on V3D 3.x */
.lower_txp = ~0,
/* Apply swizzles to all samplers. */
.swizzle_result = ~0,
.lower_invalid_implicit_lod = true,
};
/* Lower the format swizzle and (for 32-bit returns)
* ARB_texture_swizzle-style swizzle.
*/
assert(c->key->num_tex_used <= ARRAY_SIZE(c->key->tex));
for (int i = 0; i < c->key->num_tex_used; i++) {
for (int j = 0; j < 4; j++)
tex_options.swizzles[i][j] = c->key->tex[i].swizzle[j];
}
assert(c->key->num_samplers_used <= ARRAY_SIZE(c->key->sampler));
for (int i = 0; i < c->key->num_samplers_used; i++) {
if (c->key->sampler[i].return_size == 16) {
tex_options.lower_tex_packing[i] =
nir_lower_tex_packing_16;
}
}
/* CS textures may not have return_size reflecting the shadow state. */
nir_foreach_uniform_variable(var, c->s) {
const struct glsl_type *type = glsl_without_array(var->type);
unsigned array_len = MAX2(glsl_get_length(var->type), 1);
if (!glsl_type_is_sampler(type) ||
!glsl_sampler_type_is_shadow(type))
continue;
for (int i = 0; i < array_len; i++) {
tex_options.lower_tex_packing[var->data.binding + i] =
nir_lower_tex_packing_16;
}
}
NIR_PASS(_, c->s, nir_lower_tex, &tex_options);
NIR_PASS(_, c->s, nir_lower_system_values);
NIR_PASS(_, c->s, nir_lower_compute_system_values, NULL);
NIR_PASS(_, c->s, nir_lower_vars_to_scratch,
nir_var_function_temp,
0,
glsl_get_natural_size_align_bytes);
NIR_PASS(_, c->s, v3d_nir_lower_scratch);
}
static void
v3d_set_prog_data_uniforms(struct v3d_compile *c,
struct v3d_prog_data *prog_data)
{
int count = c->num_uniforms;
struct v3d_uniform_list *ulist = &prog_data->uniforms;
ulist->count = count;
ulist->data = ralloc_array(prog_data, uint32_t, count);
memcpy(ulist->data, c->uniform_data,
count * sizeof(*ulist->data));
ulist->contents = ralloc_array(prog_data, enum quniform_contents, count);
memcpy(ulist->contents, c->uniform_contents,
count * sizeof(*ulist->contents));
}
static void
v3d_vs_set_prog_data(struct v3d_compile *c,
struct v3d_vs_prog_data *prog_data)
{
/* The vertex data gets format converted by the VPM so that
* each attribute channel takes up a VPM column. Precompute
* the sizes for the shader record.
*/
for (int i = 0; i < ARRAY_SIZE(prog_data->vattr_sizes); i++) {
prog_data->vattr_sizes[i] = c->vattr_sizes[i];
prog_data->vpm_input_size += c->vattr_sizes[i];
}
memset(prog_data->driver_location_map, -1,
sizeof(prog_data->driver_location_map));
nir_foreach_shader_in_variable(var, c->s) {
prog_data->driver_location_map[var->data.location] =
var->data.driver_location;
}
prog_data->uses_vid = BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_VERTEX_ID) ||
BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
prog_data->uses_biid = BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_BASE_INSTANCE);
prog_data->uses_iid = BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_INSTANCE_ID) ||
BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_INSTANCE_INDEX);
if (prog_data->uses_vid)
prog_data->vpm_input_size++;
if (prog_data->uses_biid)
prog_data->vpm_input_size++;
if (prog_data->uses_iid)
prog_data->vpm_input_size++;
/* Input/output segment size are in sectors (8 rows of 32 bits per
* channel).
*/
prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
/* Set us up for shared input/output segments. This is apparently
* necessary for our VCM setup to avoid varying corruption.
*/
prog_data->separate_segments = false;
prog_data->vpm_output_size = MAX2(prog_data->vpm_output_size,
prog_data->vpm_input_size);
prog_data->vpm_input_size = 0;
/* Compute VCM cache size. We set up our program to take up less than
* half of the VPM, so that any set of bin and render programs won't
* run out of space. We need space for at least one input segment,
* and then allocate the rest to output segments (one for the current
* program, the rest to VCM). The valid range of the VCM cache size
* field is 1-4 16-vertex batches, but GFXH-1744 limits us to 2-4
* batches.
*/
assert(c->devinfo->vpm_size);
int sector_size = V3D_CHANNELS * sizeof(uint32_t) * 8;
int vpm_size_in_sectors = c->devinfo->vpm_size / sector_size;
int half_vpm = vpm_size_in_sectors / 2;
int vpm_output_sectors = half_vpm - prog_data->vpm_input_size;
int vpm_output_batches = vpm_output_sectors / prog_data->vpm_output_size;
assert(vpm_output_batches >= 2);
prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
}
static void
v3d_gs_set_prog_data(struct v3d_compile *c,
struct v3d_gs_prog_data *prog_data)
{
prog_data->num_inputs = c->num_inputs;
memcpy(prog_data->input_slots, c->input_slots,
c->num_inputs * sizeof(*c->input_slots));
/* gl_PrimitiveIdIn is written by the GBG into the first word of the
* VPM output header automatically and the shader will overwrite
* it after reading it if necessary, so it doesn't add to the VPM
* size requirements.
*/
prog_data->uses_pid = BITSET_TEST(c->s->info.system_values_read,
SYSTEM_VALUE_PRIMITIVE_ID);
/* Output segment size is in sectors (8 rows of 32 bits per channel) */
prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
/* Compute SIMD dispatch width and update VPM output size accordingly
* to ensure we can fit our program in memory. Available widths are
* 16, 8, 4, 1.
*
* Notice that at draw time we will have to consider VPM memory
* requirements from other stages and choose a smaller dispatch
* width if needed to fit the program in VPM memory.
*/
prog_data->simd_width = 16;
while ((prog_data->simd_width > 1 && prog_data->vpm_output_size > 16) ||
prog_data->simd_width == 2) {
prog_data->simd_width >>= 1;
prog_data->vpm_output_size =
align(prog_data->vpm_output_size, 2) / 2;
}
assert(prog_data->vpm_output_size <= 16);
assert(prog_data->simd_width != 2);
prog_data->out_prim_type = c->s->info.gs.output_primitive;
prog_data->num_invocations = c->s->info.gs.invocations;
prog_data->writes_psiz =
c->s->info.outputs_written & (1 << VARYING_SLOT_PSIZ);
}
static void
v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
struct v3d_fs_prog_data *prog_data)
{
prog_data->num_inputs = c->num_inputs;
memcpy(prog_data->input_slots, c->input_slots,
c->num_inputs * sizeof(*c->input_slots));
STATIC_ASSERT(ARRAY_SIZE(prog_data->flat_shade_flags) >
(V3D_MAX_FS_INPUTS - 1) / 24);
for (int i = 0; i < V3D_MAX_FS_INPUTS; i++) {
if (BITSET_TEST(c->flat_shade_flags, i))
prog_data->flat_shade_flags[i / 24] |= 1 << (i % 24);
if (BITSET_TEST(c->noperspective_flags, i))
prog_data->noperspective_flags[i / 24] |= 1 << (i % 24);
if (BITSET_TEST(c->centroid_flags, i))
prog_data->centroid_flags[i / 24] |= 1 << (i % 24);
}
}
static void
v3d_fs_set_prog_data(struct v3d_compile *c,
struct v3d_fs_prog_data *prog_data)
{
v3d_set_fs_prog_data_inputs(c, prog_data);
prog_data->writes_z = c->writes_z;
prog_data->writes_z_from_fep = c->writes_z_from_fep;
prog_data->disable_ez = !c->s->info.fs.early_fragment_tests;
prog_data->uses_center_w = c->uses_center_w;
prog_data->uses_implicit_point_line_varyings =
c->uses_implicit_point_line_varyings;
prog_data->lock_scoreboard_on_first_thrsw =
c->lock_scoreboard_on_first_thrsw;
prog_data->force_per_sample_msaa = c->force_per_sample_msaa;
prog_data->uses_pid = c->fs_uses_primitive_id;
}
static void
v3d_cs_set_prog_data(struct v3d_compile *c,
struct v3d_compute_prog_data *prog_data)
{
prog_data->shared_size = c->s->info.shared_size;
prog_data->local_size[0] = c->s->info.workgroup_size[0];
prog_data->local_size[1] = c->s->info.workgroup_size[1];
prog_data->local_size[2] = c->s->info.workgroup_size[2];
prog_data->has_subgroups = c->has_subgroups;
}
static void
v3d_set_prog_data(struct v3d_compile *c,
struct v3d_prog_data *prog_data)
{
prog_data->threads = c->threads;
prog_data->single_seg = !c->last_thrsw;
prog_data->spill_size = c->spill_size;
prog_data->tmu_spills = c->spills;
prog_data->tmu_fills = c->fills;
prog_data->qpu_read_stalls = c->qpu_inst_stalled_count;
prog_data->compile_strategy_idx = c->compile_strategy_idx;
prog_data->tmu_dirty_rcl = c->tmu_dirty_rcl;
prog_data->has_control_barrier = c->s->info.uses_control_barrier;
prog_data->has_global_address = c->has_global_address;
v3d_set_prog_data_uniforms(c, prog_data);
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
v3d_vs_set_prog_data(c, (struct v3d_vs_prog_data *)prog_data);
break;
case MESA_SHADER_GEOMETRY:
v3d_gs_set_prog_data(c, (struct v3d_gs_prog_data *)prog_data);
break;
case MESA_SHADER_FRAGMENT:
v3d_fs_set_prog_data(c, (struct v3d_fs_prog_data *)prog_data);
break;
case MESA_SHADER_COMPUTE:
v3d_cs_set_prog_data(c, (struct v3d_compute_prog_data *)prog_data);
break;
default:
unreachable("unsupported shader stage");
}
}
static uint64_t *
v3d_return_qpu_insts(struct v3d_compile *c, uint32_t *final_assembly_size)
{
*final_assembly_size = c->qpu_inst_count * sizeof(uint64_t);
uint64_t *qpu_insts = malloc(*final_assembly_size);
if (!qpu_insts)
return NULL;
memcpy(qpu_insts, c->qpu_insts, *final_assembly_size);
vir_compile_destroy(c);
return qpu_insts;
}
static void
v3d_nir_lower_vs_early(struct v3d_compile *c)
{
/* Split our I/O vars and dead code eliminate the unused
* components.
*/
NIR_PASS(_, c->s, nir_lower_io_to_scalar_early,
nir_var_shader_in | nir_var_shader_out);
uint64_t used_outputs[4] = {0};
for (int i = 0; i < c->vs_key->num_used_outputs; i++) {
int slot = v3d_slot_get_slot(c->vs_key->used_outputs[i]);
int comp = v3d_slot_get_component(c->vs_key->used_outputs[i]);
used_outputs[comp] |= 1ull << slot;
}
NIR_PASS(_, c->s, nir_remove_unused_io_vars,
nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
NIR_PASS(_, c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c, c->s);
NIR_PASS(_, c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
/* This must go before nir_lower_io */
if (c->vs_key->per_vertex_point_size)
NIR_PASS(_, c->s, nir_lower_point_size, 1.0f, 0.0f);
NIR_PASS(_, c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
type_size_vec4,
(nir_lower_io_options)0);
/* clean up nir_lower_io's deref_var remains and do a constant folding pass
* on the code it generated.
*/
NIR_PASS(_, c->s, nir_opt_dce);
NIR_PASS(_, c->s, nir_opt_constant_folding);
}
static void
v3d_nir_lower_gs_early(struct v3d_compile *c)
{
/* Split our I/O vars and dead code eliminate the unused
* components.
*/
NIR_PASS(_, c->s, nir_lower_io_to_scalar_early,
nir_var_shader_in | nir_var_shader_out);
uint64_t used_outputs[4] = {0};
for (int i = 0; i < c->gs_key->num_used_outputs; i++) {
int slot = v3d_slot_get_slot(c->gs_key->used_outputs[i]);
int comp = v3d_slot_get_component(c->gs_key->used_outputs[i]);
used_outputs[comp] |= 1ull << slot;
}
NIR_PASS(_, c->s, nir_remove_unused_io_vars,
nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
NIR_PASS(_, c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c, c->s);
NIR_PASS(_, c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
/* This must go before nir_lower_io */
if (c->gs_key->per_vertex_point_size)
NIR_PASS(_, c->s, nir_lower_point_size, 1.0f, 0.0f);
NIR_PASS(_, c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
type_size_vec4,
(nir_lower_io_options)0);
/* clean up nir_lower_io's deref_var remains and do a constant folding pass
* on the code it generated.
*/
NIR_PASS(_, c->s, nir_opt_dce);
NIR_PASS(_, c->s, nir_opt_constant_folding);
}
static void
v3d_fixup_fs_output_types(struct v3d_compile *c)
{
nir_foreach_shader_out_variable(var, c->s) {
uint32_t mask = 0;
switch (var->data.location) {
case FRAG_RESULT_COLOR:
mask = ~0;
break;
case FRAG_RESULT_DATA0:
case FRAG_RESULT_DATA1:
case FRAG_RESULT_DATA2:
case FRAG_RESULT_DATA3:
mask = 1 << (var->data.location - FRAG_RESULT_DATA0);
break;
}
if (c->fs_key->int_color_rb & mask) {
var->type =
glsl_vector_type(GLSL_TYPE_INT,
glsl_get_components(var->type));
} else if (c->fs_key->uint_color_rb & mask) {
var->type =
glsl_vector_type(GLSL_TYPE_UINT,
glsl_get_components(var->type));
}
}
}
static void
v3d_nir_lower_fs_early(struct v3d_compile *c)
{
if (c->fs_key->int_color_rb || c->fs_key->uint_color_rb)
v3d_fixup_fs_output_types(c);
NIR_PASS(_, c->s, v3d_nir_lower_logic_ops, c);
if (c->fs_key->line_smoothing) {
NIR_PASS(_, c->s, v3d_nir_lower_line_smooth);
NIR_PASS(_, c->s, nir_lower_global_vars_to_local);
/* The lowering pass can introduce new sysval reads */
nir_shader_gather_info(c->s, nir_shader_get_entrypoint(c->s));
}
}
static void
v3d_nir_lower_gs_late(struct v3d_compile *c)
{
if (c->key->ucp_enables) {
NIR_PASS(_, c->s, nir_lower_clip_gs, c->key->ucp_enables,
false, NULL);
}
/* Note: GS output scalarizing must happen after nir_lower_clip_gs. */
NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
}
static void
v3d_nir_lower_vs_late(struct v3d_compile *c)
{
if (c->key->ucp_enables) {
NIR_PASS(_, c->s, nir_lower_clip_vs, c->key->ucp_enables,
false, false, NULL);
NIR_PASS_V(c->s, nir_lower_io_to_scalar,
nir_var_shader_out);
}
/* Note: VS output scalarizing must happen after nir_lower_clip_vs. */
NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
}
static void
v3d_nir_lower_fs_late(struct v3d_compile *c)
{
/* In OpenGL the fragment shader can't read gl_ClipDistance[], but
* Vulkan allows it, in which case the SPIR-V compiler will declare
* VARING_SLOT_CLIP_DIST0 as compact array variable. Pass true as
* the last parameter to always operate with a compact array in both
* OpenGL and Vulkan so we do't have to care about the API we
* are using.
*/
if (c->key->ucp_enables)
NIR_PASS(_, c->s, nir_lower_clip_fs, c->key->ucp_enables, true);
NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
}
static uint32_t
vir_get_max_temps(struct v3d_compile *c)
{
int max_ip = 0;
vir_for_each_inst_inorder(inst, c)
max_ip++;
uint32_t *pressure = rzalloc_array(NULL, uint32_t, max_ip);
for (int t = 0; t < c->num_temps; t++) {
for (int i = c->temp_start[t]; (i < c->temp_end[t] &&
i < max_ip); i++) {
if (i > max_ip)
break;
pressure[i]++;
}
}
uint32_t max_temps = 0;
for (int i = 0; i < max_ip; i++)
max_temps = MAX2(max_temps, pressure[i]);
ralloc_free(pressure);
return max_temps;
}
enum v3d_dependency_class {
V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0
};
static bool
v3d_intrinsic_dependency_cb(nir_intrinsic_instr *intr,
nir_schedule_dependency *dep,
void *user_data)
{
struct v3d_compile *c = user_data;
switch (intr->intrinsic) {
case nir_intrinsic_store_output:
/* Writing to location 0 overwrites the value passed in for
* gl_PrimitiveID on geometry shaders
*/
if (c->s->info.stage != MESA_SHADER_GEOMETRY ||
nir_intrinsic_base(intr) != 0)
break;
nir_const_value *const_value =
nir_src_as_const_value(intr->src[1]);
if (const_value == NULL)
break;
uint64_t offset =
nir_const_value_as_uint(*const_value,
nir_src_bit_size(intr->src[1]));
if (offset != 0)
break;
dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
dep->type = NIR_SCHEDULE_WRITE_DEPENDENCY;
return true;
case nir_intrinsic_load_primitive_id:
if (c->s->info.stage != MESA_SHADER_GEOMETRY)
break;
dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
dep->type = NIR_SCHEDULE_READ_DEPENDENCY;
return true;
default:
break;
}
return false;
}
static unsigned
v3d_instr_delay_cb(nir_instr *instr, void *data)
{
struct v3d_compile *c = (struct v3d_compile *) data;
switch (instr->type) {
case nir_instr_type_ssa_undef:
case nir_instr_type_load_const:
case nir_instr_type_alu:
case nir_instr_type_deref:
case nir_instr_type_jump:
case nir_instr_type_parallel_copy:
case nir_instr_type_call:
case nir_instr_type_phi:
return 1;
/* We should not use very large delays for TMU instructions. Typically,
* thread switches will be sufficient to hide all or most of the latency,
* so we typically only need a little bit of extra room. If we over-estimate
* the latency here we may end up unnecesarily delaying the critical path in
* the shader, which would have a negative effect in performance, so here
* we are trying to strike a balance based on empirical testing.
*/
case nir_instr_type_intrinsic: {
if (!c->disable_general_tmu_sched) {
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
switch (intr->intrinsic) {
case nir_intrinsic_load_ssbo:
case nir_intrinsic_load_scratch:
case nir_intrinsic_load_shared:
case nir_intrinsic_image_load:
return 3;
case nir_intrinsic_load_ubo:
if (nir_src_is_divergent(intr->src[1]))
return 3;
FALLTHROUGH;
default:
return 1;
}
} else {
return 1;
}
break;
}
case nir_instr_type_tex:
return 5;
}
return 0;
}
static bool
should_split_wrmask(const nir_instr *instr, const void *data)
{
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
switch (intr->intrinsic) {
case nir_intrinsic_store_ssbo:
case nir_intrinsic_store_shared:
case nir_intrinsic_store_global:
case nir_intrinsic_store_scratch:
return true;
default:
return false;
}
}
static nir_intrinsic_instr *
nir_instr_as_constant_ubo_load(nir_instr *inst)
{
if (inst->type != nir_instr_type_intrinsic)
return NULL;
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(inst);
if (intr->intrinsic != nir_intrinsic_load_ubo)
return NULL;
assert(nir_src_is_const(intr->src[0]));
if (!nir_src_is_const(intr->src[1]))
return NULL;
return intr;
}
static bool
v3d_nir_sort_constant_ubo_load(nir_block *block, nir_intrinsic_instr *ref)
{
bool progress = false;
nir_instr *ref_inst = &ref->instr;
uint32_t ref_offset = nir_src_as_uint(ref->src[1]);
uint32_t ref_index = nir_src_as_uint(ref->src[0]);
/* Go through all instructions after ref searching for constant UBO
* loads for the same UBO index.
*/
bool seq_break = false;
nir_instr *inst = &ref->instr;
nir_instr *next_inst = NULL;
while (true) {
inst = next_inst ? next_inst : nir_instr_next(inst);
if (!inst)
break;
next_inst = NULL;
if (inst->type != nir_instr_type_intrinsic)
continue;
nir_intrinsic_instr *intr = nir_instr_as_intrinsic(inst);
if (intr->intrinsic != nir_intrinsic_load_ubo)
continue;
/* We only produce unifa sequences for non-divergent loads */
if (nir_src_is_divergent(intr->src[1]))
continue;
/* If there are any UBO loads that are not constant or that
* use a different UBO index in between the reference load and
* any other constant load for the same index, they would break
* the unifa sequence. We will flag that so we can then move
* all constant UBO loads for the reference index before these
* and not just the ones that are not ordered to avoid breaking
* the sequence and reduce unifa writes.
*/
if (!nir_src_is_const(intr->src[1])) {
seq_break = true;
continue;
}
uint32_t offset = nir_src_as_uint(intr->src[1]);
assert(nir_src_is_const(intr->src[0]));
uint32_t index = nir_src_as_uint(intr->src[0]);
if (index != ref_index) {
seq_break = true;
continue;
}
/* Only move loads with an offset that is close enough to the
* reference offset, since otherwise we would not be able to
* skip the unifa write for them. See ntq_emit_load_ubo_unifa.
*/
if (abs((int)(ref_offset - offset)) > MAX_UNIFA_SKIP_DISTANCE)
continue;
/* We will move this load if its offset is smaller than ref's
* (in which case we will move it before ref) or if the offset
* is larger than ref's but there are sequence breakers in
* in between (in which case we will move it after ref and
* before the sequence breakers).
*/
if (!seq_break && offset >= ref_offset)
continue;
/* Find where exactly we want to move this load:
*
* If we are moving it before ref, we want to check any other
* UBO loads we placed before ref and make sure we insert this
* one properly ordered with them. Likewise, if we are moving
* it after ref.
*/
nir_instr *pos = ref_inst;
nir_instr *tmp = pos;
do {
if (offset < ref_offset)
tmp = nir_instr_prev(tmp);
else
tmp = nir_instr_next(tmp);
if (!tmp || tmp == inst)
break;
/* Ignore non-unifa UBO loads */
if (tmp->type != nir_instr_type_intrinsic)
continue;
nir_intrinsic_instr *tmp_intr =
nir_instr_as_intrinsic(tmp);
if (tmp_intr->intrinsic != nir_intrinsic_load_ubo)
continue;
if (nir_src_is_divergent(tmp_intr->src[1]))
continue;
/* Stop if we find a unifa UBO load that breaks the
* sequence.
*/
if (!nir_src_is_const(tmp_intr->src[1]))
break;
if (nir_src_as_uint(tmp_intr->src[0]) != index)
break;
uint32_t tmp_offset = nir_src_as_uint(tmp_intr->src[1]);
if (offset < ref_offset) {
if (tmp_offset < offset ||
tmp_offset >= ref_offset) {
break;
} else {
pos = tmp;
}
} else {
if (tmp_offset > offset ||
tmp_offset <= ref_offset) {
break;
} else {
pos = tmp;
}
}
} while (true);
/* We can't move the UBO load before the instruction that
* defines its constant offset. If that instruction is placed
* in between the new location (pos) and the current location
* of this load, we will have to move that instruction too.
*
* We don't care about the UBO index definition because that
* is optimized to be reused by all UBO loads for the same
* index and therefore is certain to be defined before the
* first UBO load that uses it.
*/
nir_instr *offset_inst = NULL;
tmp = inst;
while ((tmp = nir_instr_prev(tmp)) != NULL) {
if (pos == tmp) {
/* We reached the target location without
* finding the instruction that defines the
* offset, so that instruction must be before
* the new position and we don't have to fix it.
*/
break;
}
if (intr->src[1].ssa->parent_instr == tmp) {
offset_inst = tmp;
break;
}
}
if (offset_inst) {
exec_node_remove(&offset_inst->node);
exec_node_insert_node_before(&pos->node,
&offset_inst->node);
}
/* Since we are moving the instruction before its current
* location, grab its successor before the move so that
* we can continue the next iteration of the main loop from
* that instruction.
*/
next_inst = nir_instr_next(inst);
/* Move this load to the selected location */
exec_node_remove(&inst->node);
if (offset < ref_offset)
exec_node_insert_node_before(&pos->node, &inst->node);
else
exec_node_insert_after(&pos->node, &inst->node);
progress = true;
}
return progress;
}
static bool
v3d_nir_sort_constant_ubo_loads_block(struct v3d_compile *c,
nir_block *block)
{
bool progress = false;
bool local_progress;
do {
local_progress = false;
nir_foreach_instr_safe(inst, block) {
nir_intrinsic_instr *intr =
nir_instr_as_constant_ubo_load(inst);
if (intr) {
local_progress |=
v3d_nir_sort_constant_ubo_load(block, intr);
}
}
progress |= local_progress;
} while (local_progress);
return progress;
}
/**
* Sorts constant UBO loads in each block by offset to maximize chances of
* skipping unifa writes when converting to VIR. This can increase register
* pressure.
*/
static bool
v3d_nir_sort_constant_ubo_loads(nir_shader *s, struct v3d_compile *c)
{
nir_foreach_function(function, s) {
if (function->impl) {
nir_foreach_block(block, function->impl) {
c->sorted_any_ubo_loads |=
v3d_nir_sort_constant_ubo_loads_block(c, block);
}
nir_metadata_preserve(function->impl,
nir_metadata_block_index |
nir_metadata_dominance);
}
}
return c->sorted_any_ubo_loads;
}
static void
lower_load_num_subgroups(struct v3d_compile *c,
nir_builder *b,
nir_intrinsic_instr *intr)
{
assert(c->s->info.stage == MESA_SHADER_COMPUTE);
assert(intr->intrinsic == nir_intrinsic_load_num_subgroups);
b->cursor = nir_after_instr(&intr->instr);
uint32_t num_subgroups =
DIV_ROUND_UP(c->s->info.workgroup_size[0] *
c->s->info.workgroup_size[1] *
c->s->info.workgroup_size[2], V3D_CHANNELS);
nir_ssa_def *result = nir_imm_int(b, num_subgroups);
nir_ssa_def_rewrite_uses(&intr->dest.ssa, result);
nir_instr_remove(&intr->instr);
}
static bool
lower_subgroup_intrinsics(struct v3d_compile *c,
nir_block *block, nir_builder *b)
{
bool progress = false;
nir_foreach_instr_safe(inst, block) {
if (inst->type != nir_instr_type_intrinsic)
continue;;
nir_intrinsic_instr *intr =
nir_instr_as_intrinsic(inst);
if (!intr)
continue;
switch (intr->intrinsic) {
case nir_intrinsic_load_num_subgroups:
lower_load_num_subgroups(c, b, intr);
progress = true;
FALLTHROUGH;
case nir_intrinsic_load_subgroup_id:
case nir_intrinsic_load_subgroup_size:
case nir_intrinsic_load_subgroup_invocation:
case nir_intrinsic_elect:
c->has_subgroups = true;
break;
default:
break;
}
}
return progress;
}
static bool
v3d_nir_lower_subgroup_intrinsics(nir_shader *s, struct v3d_compile *c)
{
bool progress = false;
nir_foreach_function(function, s) {
if (function->impl) {
nir_builder b;
nir_builder_init(&b, function->impl);
nir_foreach_block(block, function->impl)
progress |= lower_subgroup_intrinsics(c, block, &b);
nir_metadata_preserve(function->impl,
nir_metadata_block_index |
nir_metadata_dominance);
}
}
return progress;
}
static void
v3d_attempt_compile(struct v3d_compile *c)
{
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
c->vs_key = (struct v3d_vs_key *) c->key;
break;
case MESA_SHADER_GEOMETRY:
c->gs_key = (struct v3d_gs_key *) c->key;
break;
case MESA_SHADER_FRAGMENT:
c->fs_key = (struct v3d_fs_key *) c->key;
break;
case MESA_SHADER_COMPUTE:
break;
default:
unreachable("unsupported shader stage");
}
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
v3d_nir_lower_vs_early(c);
break;
case MESA_SHADER_GEOMETRY:
v3d_nir_lower_gs_early(c);
break;
case MESA_SHADER_FRAGMENT:
v3d_nir_lower_fs_early(c);
break;
default:
break;
}
v3d_lower_nir(c);
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
v3d_nir_lower_vs_late(c);
break;
case MESA_SHADER_GEOMETRY:
v3d_nir_lower_gs_late(c);
break;
case MESA_SHADER_FRAGMENT:
v3d_nir_lower_fs_late(c);
break;
default:
break;
}
NIR_PASS(_, c->s, v3d_nir_lower_io, c);
NIR_PASS(_, c->s, v3d_nir_lower_txf_ms, c);
NIR_PASS(_, c->s, v3d_nir_lower_image_load_store);
nir_lower_idiv_options idiv_options = {
.imprecise_32bit_lowering = true,
.allow_fp16 = true,
};
NIR_PASS(_, c->s, nir_lower_idiv, &idiv_options);
if (c->key->robust_buffer_access) {
/* v3d_nir_lower_robust_buffer_access assumes constant buffer
* indices on ubo/ssbo intrinsics so run copy propagation and
* constant folding passes before we run the lowering to warrant
* this. We also want to run the lowering before v3d_optimize to
* clean-up redundant get_buffer_size calls produced in the pass.
*/
NIR_PASS(_, c->s, nir_copy_prop);
NIR_PASS(_, c->s, nir_opt_constant_folding);
NIR_PASS(_, c->s, v3d_nir_lower_robust_buffer_access, c);
}
NIR_PASS(_, c->s, nir_lower_wrmasks, should_split_wrmask, c->s);
NIR_PASS(_, c->s, v3d_nir_lower_load_store_bitsize, c);
NIR_PASS(_, c->s, v3d_nir_lower_subgroup_intrinsics, c);
v3d_optimize_nir(c, c->s);
/* Do late algebraic optimization to turn add(a, neg(b)) back into
* subs, then the mandatory cleanup after algebraic. Note that it may
* produce fnegs, and if so then we need to keep running to squash
* fneg(fneg(a)).
*/
bool more_late_algebraic = true;
while (more_late_algebraic) {
more_late_algebraic = false;
NIR_PASS(more_late_algebraic, c->s, nir_opt_algebraic_late);
NIR_PASS(_, c->s, nir_opt_constant_folding);
NIR_PASS(_, c->s, nir_copy_prop);
NIR_PASS(_, c->s, nir_opt_dce);
NIR_PASS(_, c->s, nir_opt_cse);
}
NIR_PASS(_, c->s, nir_lower_bool_to_int32);
NIR_PASS(_, c->s, nir_convert_to_lcssa, true, true);
NIR_PASS_V(c->s, nir_divergence_analysis);
NIR_PASS(_, c->s, nir_convert_from_ssa, true);
struct nir_schedule_options schedule_options = {
/* Schedule for about half our register space, to enable more
* shaders to hit 4 threads.
*/
.threshold = c->threads == 4 ? 24 : 48,
/* Vertex shaders share the same memory for inputs and outputs,
* fragement and geometry shaders do not.
*/
.stages_with_shared_io_memory =
(((1 << MESA_ALL_SHADER_STAGES) - 1) &
~((1 << MESA_SHADER_FRAGMENT) |
(1 << MESA_SHADER_GEOMETRY))),
.fallback = c->fallback_scheduler,
.intrinsic_cb = v3d_intrinsic_dependency_cb,
.intrinsic_cb_data = c,
.instr_delay_cb = v3d_instr_delay_cb,
.instr_delay_cb_data = c,
};
NIR_PASS_V(c->s, nir_schedule, &schedule_options);
if (!c->disable_constant_ubo_load_sorting)
NIR_PASS(_, c->s, v3d_nir_sort_constant_ubo_loads, c);
NIR_PASS(_, c->s, nir_opt_move, nir_move_load_uniform |
nir_move_const_undef);
v3d_nir_to_vir(c);
}
uint32_t
v3d_prog_data_size(gl_shader_stage stage)
{
static const int prog_data_size[] = {
[MESA_SHADER_VERTEX] = sizeof(struct v3d_vs_prog_data),
[MESA_SHADER_GEOMETRY] = sizeof(struct v3d_gs_prog_data),
[MESA_SHADER_FRAGMENT] = sizeof(struct v3d_fs_prog_data),
[MESA_SHADER_COMPUTE] = sizeof(struct v3d_compute_prog_data),
};
assert(stage >= 0 &&
stage < ARRAY_SIZE(prog_data_size) &&
prog_data_size[stage]);
return prog_data_size[stage];
}
int v3d_shaderdb_dump(struct v3d_compile *c,
char **shaderdb_str)
{
if (c == NULL || c->compilation_result != V3D_COMPILATION_SUCCEEDED)
return -1;
return asprintf(shaderdb_str,
"%s shader: %d inst, %d threads, %d loops, "
"%d uniforms, %d max-temps, %d:%d spills:fills, "
"%d sfu-stalls, %d inst-and-stalls, %d nops",
vir_get_stage_name(c),
c->qpu_inst_count,
c->threads,
c->loops,
c->num_uniforms,
vir_get_max_temps(c),
c->spills,
c->fills,
c->qpu_inst_stalled_count,
c->qpu_inst_count + c->qpu_inst_stalled_count,
c->nop_count);
}
/* This is a list of incremental changes to the compilation strategy
* that will be used to try to compile the shader successfully. The
* default strategy is to enable all optimizations which will have
* the highest register pressure but is expected to produce most
* optimal code. Following strategies incrementally disable specific
* optimizations that are known to contribute to register pressure
* in order to be able to compile the shader successfully while meeting
* thread count requirements.
*
* V3D 4.1+ has a min thread count of 2, but we can use 1 here to also
* cover previous hardware as well (meaning that we are not limiting
* register allocation to any particular thread count). This is fine
* because v3d_nir_to_vir will cap this to the actual minimum.
*/
struct v3d_compiler_strategy {
const char *name;
uint32_t max_threads;
uint32_t min_threads;
bool disable_general_tmu_sched;
bool disable_loop_unrolling;
bool disable_ubo_load_sorting;
bool disable_tmu_pipelining;
uint32_t max_tmu_spills;
} static const strategies[] = {
/*0*/ { "default", 4, 4, false, false, false, false, 0 },
/*1*/ { "disable general TMU sched", 4, 4, true, false, false, false, 0 },
/*2*/ { "disable loop unrolling", 4, 4, true, true, false, false, 0 },
/*3*/ { "disable UBO load sorting", 4, 4, true, true, true, false, 0 },
/*4*/ { "disable TMU pipelining", 4, 4, true, true, true, true, 0 },
/*5*/ { "lower thread count", 2, 1, false, false, false, false, -1 },
/*6*/ { "disable general TMU sched (2t)", 2, 1, true, false, false, false, -1 },
/*7*/ { "disable loop unrolling (2t)", 2, 1, true, true, false, false, -1 },
/*8*/ { "disable UBO load sorting (2t)", 2, 1, true, true, true, false, -1 },
/*9*/ { "disable TMU pipelining (2t)", 2, 1, true, true, true, true, -1 },
/*10*/ { "fallback scheduler", 2, 1, true, true, true, true, -1 }
};
/**
* If a particular optimization didn't make any progress during a compile
* attempt disabling it alone won't allow us to compile the shader successfuly,
* since we'll end up with the same code. Detect these scenarios so we can
* avoid wasting time with useless compiles. We should also consider if the
* gy changes other aspects of the compilation process though, like
* spilling, and not skip it in that case.
*/
static bool
skip_compile_strategy(struct v3d_compile *c, uint32_t idx)
{
/* We decide if we can skip a strategy based on the optimizations that
* were active in the previous strategy, so we should only be calling this
* for strategies after the first.
*/
assert(idx > 0);
/* Don't skip a strategy that changes spilling behavior */
if (strategies[idx].max_tmu_spills !=
strategies[idx - 1].max_tmu_spills) {
return false;
}
switch (idx) {
/* General TMU sched.: skip if we didn't emit any TMU loads */
case 1:
case 6:
return !c->has_general_tmu_load;
/* Loop unrolling: skip if we didn't unroll any loops */
case 2:
case 7:
return !c->unrolled_any_loops;
/* UBO load sorting: skip if we didn't sort any loads */
case 3:
case 8:
return !c->sorted_any_ubo_loads;
/* TMU pipelining: skip if we didn't pipeline any TMU ops */
case 4:
case 9:
return !c->pipelined_any_tmu;
/* Lower thread count: skip if we already tried less that 4 threads */
case 5:
return c->threads < 4;
default:
return false;
};
}
uint64_t *v3d_compile(const struct v3d_compiler *compiler,
struct v3d_key *key,
struct v3d_prog_data **out_prog_data,
nir_shader *s,
void (*debug_output)(const char *msg,
void *debug_output_data),
void *debug_output_data,
int program_id, int variant_id,
uint32_t *final_assembly_size)
{
struct v3d_compile *c = NULL;
uint32_t best_spill_fill_count = UINT32_MAX;
struct v3d_compile *best_c = NULL;
for (int32_t strat = 0; strat < ARRAY_SIZE(strategies); strat++) {
/* Fallback strategy */
if (strat > 0) {
assert(c);
if (skip_compile_strategy(c, strat))
continue;
char *debug_msg;
int ret = asprintf(&debug_msg,
"Falling back to strategy '%s' "
"for %s prog %d/%d",
strategies[strat].name,
vir_get_stage_name(c),
c->program_id, c->variant_id);
if (ret >= 0) {
if (unlikely(V3D_DEBUG & V3D_DEBUG_PERF))
fprintf(stderr, "%s\n", debug_msg);
c->debug_output(debug_msg, c->debug_output_data);
free(debug_msg);
}
if (c != best_c)
vir_compile_destroy(c);
}
c = vir_compile_init(compiler, key, s,
debug_output, debug_output_data,
program_id, variant_id,
strat,
strategies[strat].max_threads,
strategies[strat].min_threads,
strategies[strat].max_tmu_spills,
strategies[strat].disable_general_tmu_sched,
strategies[strat].disable_loop_unrolling,
strategies[strat].disable_ubo_load_sorting,
strategies[strat].disable_tmu_pipelining,
strat == ARRAY_SIZE(strategies) - 1);
v3d_attempt_compile(c);
/* Broken shader or driver bug */
if (c->compilation_result == V3D_COMPILATION_FAILED)
break;
/* If we compiled without spills, choose this.
* Otherwise if this is a 4-thread compile, choose this (these
* have a very low cap on the allowed TMU spills so we assume
* it will be better than a 2-thread compile without spills).
* Otherwise, keep going while tracking the strategy with the
* lowest spill count.
*/
if (c->compilation_result == V3D_COMPILATION_SUCCEEDED) {
if (c->spills == 0 ||
strategies[strat].min_threads == 4) {
best_c = c;
break;
} else if (c->spills + c->fills <
best_spill_fill_count) {
best_c = c;
best_spill_fill_count = c->spills + c->fills;
}
if (unlikely(V3D_DEBUG & V3D_DEBUG_PERF)) {
char *debug_msg;
int ret = asprintf(&debug_msg,
"Compiled %s prog %d/%d with %d "
"spills and %d fills. Will try "
"more strategies.",
vir_get_stage_name(c),
c->program_id, c->variant_id,
c->spills, c->fills);
if (ret >= 0) {
fprintf(stderr, "%s\n", debug_msg);
c->debug_output(debug_msg, c->debug_output_data);
free(debug_msg);
}
}
}
/* Only try next streategy if we failed to register allocate
* or we had to spill.
*/
assert(c->compilation_result ==
V3D_COMPILATION_FAILED_REGISTER_ALLOCATION ||
c->spills > 0);
}
/* If the best strategy was not the last, choose that */
if (best_c && c != best_c) {
vir_compile_destroy(c);
c = best_c;
}
if (unlikely(V3D_DEBUG & V3D_DEBUG_PERF) &&
c->compilation_result !=
V3D_COMPILATION_FAILED_REGISTER_ALLOCATION &&
c->spills > 0) {
char *debug_msg;
int ret = asprintf(&debug_msg,
"Compiled %s prog %d/%d with %d "
"spills and %d fills",
vir_get_stage_name(c),
c->program_id, c->variant_id,
c->spills, c->fills);
fprintf(stderr, "%s\n", debug_msg);
if (ret >= 0) {
c->debug_output(debug_msg, c->debug_output_data);
free(debug_msg);
}
}
if (c->compilation_result != V3D_COMPILATION_SUCCEEDED) {
fprintf(stderr, "Failed to compile %s prog %d/%d "
"with any strategy.\n",
vir_get_stage_name(c), c->program_id, c->variant_id);
}
struct v3d_prog_data *prog_data;
prog_data = rzalloc_size(NULL, v3d_prog_data_size(c->s->info.stage));
v3d_set_prog_data(c, prog_data);
*out_prog_data = prog_data;
char *shaderdb;
int ret = v3d_shaderdb_dump(c, &shaderdb);
if (ret >= 0) {
if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
fprintf(stderr, "SHADER-DB-%s - %s\n", s->info.name, shaderdb);
c->debug_output(shaderdb, c->debug_output_data);
free(shaderdb);
}
return v3d_return_qpu_insts(c, final_assembly_size);
}
void
vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst)
{
if (qinst->dst.file == QFILE_TEMP)
c->defs[qinst->dst.index] = NULL;
assert(&qinst->link != c->cursor.link);
list_del(&qinst->link);
free(qinst);
c->live_intervals_valid = false;
}
struct qreg
vir_follow_movs(struct v3d_compile *c, struct qreg reg)
{
/* XXX
int pack = reg.pack;
while (reg.file == QFILE_TEMP &&
c->defs[reg.index] &&
(c->defs[reg.index]->op == QOP_MOV ||
c->defs[reg.index]->op == QOP_FMOV) &&
!c->defs[reg.index]->dst.pack &&
!c->defs[reg.index]->src[0].pack) {
reg = c->defs[reg.index]->src[0];
}
reg.pack = pack;
*/
return reg;
}
void
vir_compile_destroy(struct v3d_compile *c)
{
/* Defuse the assert that we aren't removing the cursor's instruction.
*/
c->cursor.link = NULL;
vir_for_each_block(block, c) {
while (!list_is_empty(&block->instructions)) {
struct qinst *qinst =
list_first_entry(&block->instructions,
struct qinst, link);
vir_remove_instruction(c, qinst);
}
}
ralloc_free(c);
}
uint32_t
vir_get_uniform_index(struct v3d_compile *c,
enum quniform_contents contents,
uint32_t data)
{
for (int i = 0; i < c->num_uniforms; i++) {
if (c->uniform_contents[i] == contents &&
c->uniform_data[i] == data) {
return i;
}
}
uint32_t uniform = c->num_uniforms++;
if (uniform >= c->uniform_array_size) {
c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
c->uniform_array_size * 2);
c->uniform_data = reralloc(c, c->uniform_data,
uint32_t,
c->uniform_array_size);
c->uniform_contents = reralloc(c, c->uniform_contents,
enum quniform_contents,
c->uniform_array_size);
}
c->uniform_contents[uniform] = contents;
c->uniform_data[uniform] = data;
return uniform;
}
/* Looks back into the current block to find the ldunif that wrote the uniform
* at the requested index. If it finds it, it returns true and writes the
* destination register of the ldunif instruction to 'unif'.
*
* This can impact register pressure and end up leading to worse code, so we
* limit the number of instructions we are willing to look back through to
* strike a good balance.
*/
static bool
try_opt_ldunif(struct v3d_compile *c, uint32_t index, struct qreg *unif)
{
uint32_t count = 20;
struct qinst *prev_inst = NULL;
assert(c->cur_block);
#ifdef DEBUG
/* We can only reuse a uniform if it was emitted in the same block,
* so callers must make sure the current instruction is being emitted
* in the current block.
*/
bool found = false;
vir_for_each_inst(inst, c->cur_block) {
if (&inst->link == c->cursor.link) {
found = true;
break;
}
}
assert(found || &c->cur_block->instructions == c->cursor.link);
#endif
list_for_each_entry_from_rev(struct qinst, inst, c->cursor.link->prev,
&c->cur_block->instructions, link) {
if ((inst->qpu.sig.ldunif || inst->qpu.sig.ldunifrf) &&
inst->uniform == index) {
prev_inst = inst;
break;
}
if (--count == 0)
break;
}
if (!prev_inst)
return false;
list_for_each_entry_from(struct qinst, inst, prev_inst->link.next,
&c->cur_block->instructions, link) {
if (inst->dst.file == prev_inst->dst.file &&
inst->dst.index == prev_inst->dst.index) {
return false;
}
}
*unif = prev_inst->dst;
return true;
}
struct qreg
vir_uniform(struct v3d_compile *c,
enum quniform_contents contents,
uint32_t data)
{
const int num_uniforms = c->num_uniforms;
const int index = vir_get_uniform_index(c, contents, data);
/* If this is not the first time we see this uniform try to reuse the
* result of the last ldunif that loaded it.
*/
const bool is_new_uniform = num_uniforms != c->num_uniforms;
if (!is_new_uniform && !c->disable_ldunif_opt) {
struct qreg ldunif_dst;
if (try_opt_ldunif(c, index, &ldunif_dst))
return ldunif_dst;
}
struct qinst *inst = vir_NOP(c);
inst->qpu.sig.ldunif = true;
inst->uniform = index;
inst->dst = vir_get_temp(c);
c->defs[inst->dst.index] = inst;
return inst->dst;
}
#define OPTPASS(func) \
do { \
bool stage_progress = func(c); \
if (stage_progress) { \
progress = true; \
if (print_opt_debug) { \
fprintf(stderr, \
"VIR opt pass %2d: %s progress\n", \
pass, #func); \
} \
/*XXX vir_validate(c);*/ \
} \
} while (0)
void
vir_optimize(struct v3d_compile *c)
{
bool print_opt_debug = false;
int pass = 1;
while (true) {
bool progress = false;
OPTPASS(vir_opt_copy_propagate);
OPTPASS(vir_opt_redundant_flags);
OPTPASS(vir_opt_dead_code);
OPTPASS(vir_opt_small_immediates);
OPTPASS(vir_opt_constant_alu);
if (!progress)
break;
pass++;
}
}
const char *
vir_get_stage_name(struct v3d_compile *c)
{
if (c->vs_key && c->vs_key->is_coord)
return "MESA_SHADER_VERTEX_BIN";
else if (c->gs_key && c->gs_key->is_coord)
return "MESA_SHADER_GEOMETRY_BIN";
else
return gl_shader_stage_name(c->s->info.stage);
}
static inline uint32_t
compute_vpm_size_in_sectors(const struct v3d_device_info *devinfo)
{
assert(devinfo->vpm_size > 0);
const uint32_t sector_size = V3D_CHANNELS * sizeof(uint32_t) * 8;
return devinfo->vpm_size / sector_size;
}
/* Computes various parameters affecting VPM memory configuration for programs
* involving geometry shaders to ensure the program fits in memory and honors
* requirements described in section "VPM usage" of the programming manual.
*/
static bool
compute_vpm_config_gs(struct v3d_device_info *devinfo,
struct v3d_vs_prog_data *vs,
struct v3d_gs_prog_data *gs,
struct vpm_config *vpm_cfg_out)
{
const uint32_t A = vs->separate_segments ? 1 : 0;
const uint32_t Ad = vs->vpm_input_size;
const uint32_t Vd = vs->vpm_output_size;
const uint32_t vpm_size = compute_vpm_size_in_sectors(devinfo);
/* Try to fit program into our VPM memory budget by adjusting
* configurable parameters iteratively. We do this in two phases:
* the first phase tries to fit the program into the total available
* VPM memory. If we succeed at that, then the second phase attempts
* to fit the program into half of that budget so we can run bin and
* render programs in parallel.
*/
struct vpm_config vpm_cfg[2];
struct vpm_config *final_vpm_cfg = NULL;
uint32_t phase = 0;
vpm_cfg[phase].As = 1;
vpm_cfg[phase].Gs = 1;
vpm_cfg[phase].Gd = gs->vpm_output_size;
vpm_cfg[phase].gs_width = gs->simd_width;
/* While there is a requirement that Vc >= [Vn / 16], this is
* always the case when tessellation is not present because in that
* case Vn can only be 6 at most (when input primitive is triangles
* with adjacency).
*
* We always choose Vc=2. We can't go lower than this due to GFXH-1744,
* and Broadcom has not found it worth it to increase it beyond this
* in general. Increasing Vc also increases VPM memory pressure which
* can turn up being detrimental for performance in some scenarios.
*/
vpm_cfg[phase].Vc = 2;
/* Gv is a constraint on the hardware to not exceed the
* specified number of vertex segments per GS batch. If adding a
* new primitive to a GS batch would result in a range of more
* than Gv vertex segments being referenced by the batch, then
* the hardware will flush the batch and start a new one. This
* means that we can choose any value we want, we just need to
* be aware that larger values improve GS batch utilization
* at the expense of more VPM memory pressure (which can affect
* other performance aspects, such as GS dispatch width).
* We start with the largest value, and will reduce it if we
* find that total memory pressure is too high.
*/
vpm_cfg[phase].Gv = 3;
do {
/* When GS is present in absence of TES, then we need to satisfy
* that Ve >= Gv. We go with the smallest value of Ve to avoid
* increasing memory pressure.
*/
vpm_cfg[phase].Ve = vpm_cfg[phase].Gv;
uint32_t vpm_sectors =
A * vpm_cfg[phase].As * Ad +
(vpm_cfg[phase].Vc + vpm_cfg[phase].Ve) * Vd +
vpm_cfg[phase].Gs * vpm_cfg[phase].Gd;
/* Ideally we want to use no more than half of the available
* memory so we can execute a bin and render program in parallel
* without stalls. If we achieved that then we are done.
*/
if (vpm_sectors <= vpm_size / 2) {
final_vpm_cfg = &vpm_cfg[phase];
break;
}
/* At the very least, we should not allocate more than the
* total available VPM memory. If we have a configuration that
* succeeds at this we save it and continue to see if we can
* meet the half-memory-use criteria too.
*/
if (phase == 0 && vpm_sectors <= vpm_size) {
vpm_cfg[1] = vpm_cfg[0];
phase = 1;
}
/* Try lowering Gv */
if (vpm_cfg[phase].Gv > 0) {
vpm_cfg[phase].Gv--;
continue;
}
/* Try lowering GS dispatch width */
if (vpm_cfg[phase].gs_width > 1) {
do {
vpm_cfg[phase].gs_width >>= 1;
vpm_cfg[phase].Gd = align(vpm_cfg[phase].Gd, 2) / 2;
} while (vpm_cfg[phase].gs_width == 2);
/* Reset Gv to max after dropping dispatch width */
vpm_cfg[phase].Gv = 3;
continue;
}
/* We ran out of options to reduce memory pressure. If we
* are at phase 1 we have at least a valid configuration, so we
* we use that.
*/
if (phase == 1)
final_vpm_cfg = &vpm_cfg[0];
break;
} while (true);
if (!final_vpm_cfg)
return false;
assert(final_vpm_cfg);
assert(final_vpm_cfg->Gd <= 16);
assert(final_vpm_cfg->Gv < 4);
assert(final_vpm_cfg->Ve < 4);
assert(final_vpm_cfg->Vc >= 2 && final_vpm_cfg->Vc <= 4);
assert(final_vpm_cfg->gs_width == 1 ||
final_vpm_cfg->gs_width == 4 ||
final_vpm_cfg->gs_width == 8 ||
final_vpm_cfg->gs_width == 16);
*vpm_cfg_out = *final_vpm_cfg;
return true;
}
bool
v3d_compute_vpm_config(struct v3d_device_info *devinfo,
struct v3d_vs_prog_data *vs_bin,
struct v3d_vs_prog_data *vs,
struct v3d_gs_prog_data *gs_bin,
struct v3d_gs_prog_data *gs,
struct vpm_config *vpm_cfg_bin,
struct vpm_config *vpm_cfg)
{
assert(vs && vs_bin);
assert((gs != NULL) == (gs_bin != NULL));
if (!gs) {
vpm_cfg_bin->As = 1;
vpm_cfg_bin->Ve = 0;
vpm_cfg_bin->Vc = vs_bin->vcm_cache_size;
vpm_cfg->As = 1;
vpm_cfg->Ve = 0;
vpm_cfg->Vc = vs->vcm_cache_size;
} else {
if (!compute_vpm_config_gs(devinfo, vs_bin, gs_bin, vpm_cfg_bin))
return false;
if (!compute_vpm_config_gs(devinfo, vs, gs, vpm_cfg))
return false;
}
return true;
}
Reference in New Issue View Git Blame Copy Permalink