#include "broadcom/common/v3d_device_info.h"
#include "v3d_compiler.h"
+#include "util/u_prim.h"
int
-vir_get_non_sideband_nsrc(struct qinst *inst)
+vir_get_nsrc(struct qinst *inst)
{
switch (inst->qpu.type) {
case V3D_QPU_INSTR_TYPE_BRANCH:
return 0;
}
-int
-vir_get_nsrc(struct qinst *inst)
-{
- int nsrc = vir_get_non_sideband_nsrc(inst);
-
- if (vir_has_implicit_uniform(inst))
- nsrc++;
-
- return nsrc;
-}
-
-bool
-vir_has_implicit_uniform(struct qinst *inst)
-{
- switch (inst->qpu.type) {
- case V3D_QPU_INSTR_TYPE_BRANCH:
- return true;
- case V3D_QPU_INSTR_TYPE_ALU:
- switch (inst->dst.file) {
- case QFILE_TLBU:
- return true;
- default:
- return inst->has_implicit_uniform;
- }
- }
- return false;
-}
-
-/* The sideband uniform for textures gets stored after the normal ALU
- * arguments.
- */
-int
-vir_get_implicit_uniform_src(struct qinst *inst)
-{
- return vir_get_nsrc(inst) - 1;
-}
-
/**
* Returns whether the instruction has any side effects that must be
* preserved.
case V3D_QPU_A_STVPMD:
case V3D_QPU_A_STVPMP:
case V3D_QPU_A_VPMWT:
+ case V3D_QPU_A_TMUWT:
return true;
default:
break;
}
}
- if (inst->qpu.sig.ldtmu || inst->qpu.sig.thrsw)
+ 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;
-
- return false;
-}
-
-bool
-vir_is_float_input(struct qinst *inst)
-{
- /* XXX: More instrs */
- switch (inst->qpu.type) {
- case V3D_QPU_INSTR_TYPE_BRANCH:
- return false;
- case V3D_QPU_INSTR_TYPE_ALU:
- switch (inst->qpu.alu.add.op) {
- case V3D_QPU_A_FADD:
- case V3D_QPU_A_FSUB:
- case V3D_QPU_A_FMIN:
- case V3D_QPU_A_FMAX:
- case V3D_QPU_A_FTOIN:
- return true;
- default:
- break;
- }
-
- switch (inst->qpu.alu.mul.op) {
- case V3D_QPU_M_FMOV:
- case V3D_QPU_M_VFMUL:
- case V3D_QPU_M_FMUL:
- return true;
- default:
- break;
- }
}
return false;
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;
if (inst->dst.file == QFILE_MAGIC)
return v3d_qpu_magic_waddr_is_tmu(inst->dst.index);
- return false;
-}
-
-bool
-vir_depends_on_flags(struct qinst *inst)
-{
- if (inst->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH) {
- return (inst->qpu.branch.cond != V3D_QPU_BRANCH_COND_ALWAYS);
- } else {
- return (inst->qpu.flags.ac != V3D_QPU_COND_NONE &&
- inst->qpu.flags.mc != V3D_QPU_COND_NONE);
+ if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
+ inst->qpu.alu.add.op == V3D_QPU_A_TMUWT) {
+ return true;
}
+
+ return false;
}
bool
{
for (int i = 0; i < vir_get_nsrc(inst); i++) {
switch (inst->src[i].file) {
- case QFILE_VARY:
case QFILE_VPM:
return true;
default:
}
}
+void
+vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf)
+{
+ 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 (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_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src)
+vir_branch_inst(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
{
struct qinst *inst = calloc(1, sizeof(*inst));
inst->qpu.branch.ub = true;
inst->qpu.branch.bdu = V3D_QPU_BRANCH_DEST_REL;
- inst->dst = vir_reg(QFILE_NULL, 0);
- inst->src[0] = src;
- inst->uniform = ~0;
+ inst->dst = vir_nop_reg();
+ inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, 0);
return inst;
}
static void
vir_emit(struct v3d_compile *c, struct qinst *inst)
{
- list_addtail(&inst->link, &c->cur_block->instructions);
+ 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. */
{
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)
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);
}
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)
{
struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
c->program_id = program_id;
c->variant_id = variant_id;
c->threads = 4;
+ c->debug_output = debug_output;
+ c->debug_output_data = debug_output_data;
s = nir_shader_clone(c, s);
c->s = s;
vir_set_emit_block(c, vir_new_block(c));
c->output_position_index = -1;
- c->output_point_size_index = -1;
c->output_sample_mask_index = -1;
c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
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_rect = false, /* XXX */
+ .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,
tex_options.saturate_t |= 1 << i;
if (c->key->tex[i].clamp_r)
tex_options.saturate_r |= 1 << i;
+ if (c->key->tex[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_variable(var, &c->s->uniforms) {
+ 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_V(c->s, nir_lower_tex, &tex_options);
-}
+ NIR_PASS_V(c->s, nir_lower_system_values);
-static void
-v3d_lower_nir_late(struct v3d_compile *c)
-{
- NIR_PASS_V(c->s, v3d_nir_lower_io, c);
- NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
- NIR_PASS_V(c->s, nir_lower_idiv);
+ NIR_PASS_V(c->s, nir_lower_vars_to_scratch,
+ nir_var_function_temp,
+ 0,
+ glsl_get_natural_size_align_bytes);
+ NIR_PASS_V(c->s, v3d_nir_lower_scratch);
}
static void
count * sizeof(*ulist->contents));
}
-/* Copy the compiler UBO range state to the compiled shader, dropping out
- * arrays that were never referenced by an indirect load.
- *
- * (Note that QIR dead code elimination of an array access still leaves that
- * array alive, though)
- */
static void
-v3d_set_prog_data_ubo(struct v3d_compile *c,
- struct v3d_prog_data *prog_data)
+v3d_vs_set_prog_data(struct v3d_compile *c,
+ struct v3d_vs_prog_data *prog_data)
{
- if (!c->num_ubo_ranges)
- return;
+ /* 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];
+ }
- prog_data->num_ubo_ranges = 0;
- prog_data->ubo_ranges = ralloc_array(prog_data, struct v3d_ubo_range,
- c->num_ubo_ranges);
- for (int i = 0; i < c->num_ubo_ranges; i++) {
- if (!c->ubo_range_used[i])
- continue;
+ prog_data->uses_vid = (c->s->info.system_values_read &
+ (1ull << SYSTEM_VALUE_VERTEX_ID));
+ prog_data->uses_iid = (c->s->info.system_values_read &
+ (1ull << SYSTEM_VALUE_INSTANCE_ID));
+
+ if (prog_data->uses_vid)
+ 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));
- struct v3d_ubo_range *range = &c->ubo_ranges[i];
- prog_data->ubo_ranges[prog_data->num_ubo_ranges++] = *range;
- prog_data->ubo_size += range->size;
+ /* 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 = (c->s->info.system_values_read &
+ (1ull << 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);
- if (prog_data->ubo_size) {
- if (V3D_DEBUG & V3D_DEBUG_SHADERDB) {
- fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
- vir_get_stage_name(c),
- c->program_id, c->variant_id,
- prog_data->ubo_size / 4);
- }
+ prog_data->out_prim_type = c->s->info.gs.output_primitive;
+ prog_data->num_invocations = c->s->info.gs.invocations;
+}
+
+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->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;
+}
+
+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.cs.shared_size;
+}
+
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_dirty_rcl = c->tmu_dirty_rcl;
v3d_set_prog_data_uniforms(c, prog_data);
- v3d_set_prog_data_ubo(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 *
return qpu_insts;
}
-uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
- struct v3d_vs_key *key,
- struct v3d_vs_prog_data *prog_data,
- nir_shader *s,
- int program_id, int variant_id,
- uint32_t *final_assembly_size)
+static void
+v3d_nir_lower_vs_early(struct v3d_compile *c)
{
- struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
- program_id, variant_id);
-
- c->vs_key = key;
+ /* Split our I/O vars and dead code eliminate the unused
+ * components.
+ */
+ NIR_PASS_V(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_V(c->s, nir_remove_unused_io_vars,
+ &c->s->outputs, used_outputs, NULL); /* demotes to globals */
+ NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
+ v3d_optimize_nir(c->s);
+ NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in);
- v3d_lower_nir(c);
+ /* This must go before nir_lower_io */
+ if (c->vs_key->per_vertex_point_size)
+ NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
- if (key->clamp_color)
- NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
+ NIR_PASS_V(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 */
+ NIR_PASS_V(c->s, nir_opt_dce);
+}
- if (key->base.ucp_enables) {
- NIR_PASS_V(c->s, nir_lower_clip_vs, key->base.ucp_enables);
- NIR_PASS_V(c->s, nir_lower_io_to_scalar,
- nir_var_shader_out);
+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_V(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;
}
-
- /* 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);
-
- v3d_lower_nir_late(c);
+ NIR_PASS_V(c->s, nir_remove_unused_io_vars,
+ &c->s->outputs, used_outputs, NULL); /* demotes to globals */
+ NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c->s);
- NIR_PASS_V(c->s, nir_convert_from_ssa, true);
+ NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in);
- v3d_nir_to_vir(c);
+ /* This must go before nir_lower_io */
+ if (c->gs_key->per_vertex_point_size)
+ NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
- v3d_set_prog_data(c, &prog_data->base);
+ NIR_PASS_V(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 */
+ NIR_PASS_V(c->s, nir_opt_dce);
+}
- prog_data->base.num_inputs = c->num_inputs;
+static void
+v3d_fixup_fs_output_types(struct v3d_compile *c)
+{
+ nir_foreach_variable(var, &c->s->outputs) {
+ uint32_t mask = 0;
- /* 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];
+ 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));
+ }
}
+}
- /* Input/output segment size are in 8x32-bit multiples. */
- prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
- prog_data->vpm_output_size = align(c->num_vpm_writes, 8) / 8;
+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);
- prog_data->uses_vid = (s->info.system_values_read &
- (1ull << SYSTEM_VALUE_VERTEX_ID));
- prog_data->uses_iid = (s->info.system_values_read &
- (1ull << SYSTEM_VALUE_INSTANCE_ID));
+ NIR_PASS_V(c->s, v3d_nir_lower_logic_ops, c);
- return v3d_return_qpu_insts(c, final_assembly_size);
+ /* If the shader has no non-TLB side effects, we can promote it to
+ * enabling early_fragment_tests even if the user didn't.
+ */
+ if (!(c->s->info.num_images ||
+ c->s->info.num_ssbos)) {
+ c->s->info.fs.early_fragment_tests = true;
+ }
}
static void
-v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
- struct v3d_fs_prog_data *prog_data)
+v3d_nir_lower_gs_late(struct v3d_compile *c)
{
- prog_data->base.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 (c->key->ucp_enables) {
+ NIR_PASS_V(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);
}
-uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
- struct v3d_fs_key *key,
- struct v3d_fs_prog_data *prog_data,
- nir_shader *s,
- int program_id, int variant_id,
- uint32_t *final_assembly_size)
+static void
+v3d_nir_lower_vs_late(struct v3d_compile *c)
{
- struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
- program_id, variant_id);
+ if (c->vs_key->clamp_color)
+ NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
- c->fs_key = key;
+ if (c->key->ucp_enables) {
+ NIR_PASS_V(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);
+ }
- v3d_lower_nir(c);
+ /* 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);
+}
- if (key->light_twoside)
+static void
+v3d_nir_lower_fs_late(struct v3d_compile *c)
+{
+ if (c->fs_key->light_twoside)
NIR_PASS_V(c->s, nir_lower_two_sided_color);
- if (key->clamp_color)
+ if (c->fs_key->clamp_color)
NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
- if (key->alpha_test) {
- NIR_PASS_V(c->s, nir_lower_alpha_test, key->alpha_test_func,
- false);
+ if (c->fs_key->alpha_test) {
+ NIR_PASS_V(c->s, nir_lower_alpha_test,
+ c->fs_key->alpha_test_func,
+ false, NULL);
}
- if (key->base.ucp_enables)
- NIR_PASS_V(c->s, nir_lower_clip_fs, key->base.ucp_enables);
+ if (c->key->ucp_enables)
+ NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables,
+ false);
/* Note: FS input scalarizing must happen after
* nir_lower_two_sided_color, which only handles a vec4 at a time.
*/
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;
+}
+
+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_prog_data *prog_data;
+ struct v3d_compile *c = vir_compile_init(compiler, key, s,
+ debug_output, debug_output_data,
+ program_id, variant_id);
+
+ switch (c->s->info.stage) {
+ case MESA_SHADER_VERTEX:
+ c->vs_key = (struct v3d_vs_key *)key;
+ prog_data = rzalloc_size(NULL, sizeof(struct v3d_vs_prog_data));
+ break;
+ case MESA_SHADER_GEOMETRY:
+ c->gs_key = (struct v3d_gs_key *)key;
+ prog_data = rzalloc_size(NULL, sizeof(struct v3d_gs_prog_data));
+ break;
+ case MESA_SHADER_FRAGMENT:
+ c->fs_key = (struct v3d_fs_key *)key;
+ prog_data = rzalloc_size(NULL, sizeof(struct v3d_fs_prog_data));
+ break;
+ case MESA_SHADER_COMPUTE:
+ prog_data = rzalloc_size(NULL,
+ sizeof(struct v3d_compute_prog_data));
+ 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_V(c->s, v3d_nir_lower_io, c);
+ NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
+ NIR_PASS_V(c->s, v3d_nir_lower_image_load_store);
+ NIR_PASS_V(c->s, nir_lower_idiv, nir_lower_idiv_fast);
- v3d_lower_nir_late(c);
v3d_optimize_nir(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_V(c->s, nir_opt_constant_folding);
+ NIR_PASS_V(c->s, nir_copy_prop);
+ NIR_PASS_V(c->s, nir_opt_dce);
+ NIR_PASS_V(c->s, nir_opt_cse);
+ }
+
+ NIR_PASS_V(c->s, nir_lower_bool_to_int32);
NIR_PASS_V(c->s, nir_convert_from_ssa, true);
+ /* Schedule for about half our register space, to enable more shaders
+ * to hit 4 threads.
+ */
+ NIR_PASS_V(c->s, nir_schedule, 24);
+
v3d_nir_to_vir(c);
- v3d_set_prog_data(c, &prog_data->base);
- v3d_set_fs_prog_data_inputs(c, prog_data);
- prog_data->writes_z = (c->s->info.outputs_written &
- (1 << FRAG_RESULT_DEPTH));
- prog_data->discard = c->s->info.fs.uses_discard;
+ v3d_set_prog_data(c, prog_data);
+
+ *out_prog_data = prog_data;
+
+ char *shaderdb;
+ int ret = asprintf(&shaderdb,
+ "%s shader: %d inst, %d threads, %d loops, "
+ "%d uniforms, %d max-temps, %d:%d spills:fills, "
+ "%d sfu-stalls, %d inst-and-stalls",
+ 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);
+ if (ret >= 0) {
+ if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
+ fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
+
+ c->debug_output(shaderdb, c->debug_output_data);
+ free(shaderdb);
+ }
- return v3d_return_qpu_insts(c, final_assembly_size);
+ return v3d_return_qpu_insts(c, final_assembly_size);
}
void
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
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_empty(&block->instructions)) {
+ while (!list_is_empty(&block->instructions)) {
struct qinst *qinst =
list_first_entry(&block->instructions,
struct qinst, link);
ralloc_free(c);
}
-struct qreg
-vir_uniform(struct v3d_compile *c,
- enum quniform_contents contents,
- uint32_t data)
+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 vir_reg(QFILE_UNIF, i);
+ return i;
}
}
c->uniform_contents[uniform] = contents;
c->uniform_data[uniform] = data;
- return vir_reg(QFILE_UNIF, uniform);
+ return uniform;
}
-void
-vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf)
+struct qreg
+vir_uniform(struct v3d_compile *c,
+ enum quniform_contents contents,
+ uint32_t data)
{
- struct qinst *last_inst = NULL;
-
- if (!list_empty(&c->cur_block->instructions))
- last_inst = (struct qinst *)c->cur_block->instructions.prev;
-
- if (src.file != QFILE_TEMP ||
- !c->defs[src.index] ||
- last_inst != c->defs[src.index]) {
- /* XXX: Make the MOV be the appropriate type */
- last_inst = vir_MOV_dest(c, vir_reg(QFILE_NULL, 0), src);
- last_inst = (struct qinst *)c->cur_block->instructions.prev;
- }
-
- vir_set_pf(last_inst, pf);
+ struct qinst *inst = vir_NOP(c);
+ inst->qpu.sig.ldunif = true;
+ inst->uniform = vir_get_uniform_index(c, contents, data);
+ inst->dst = vir_get_temp(c);
+ c->defs[inst->dst.index] = inst;
+ return inst->dst;
}
#define OPTPASS(func) \
bool progress = false;
OPTPASS(vir_opt_copy_propagate);
+ OPTPASS(vir_opt_redundant_flags);
OPTPASS(vir_opt_dead_code);
+ OPTPASS(vir_opt_small_immediates);
if (!progress)
break;
vir_get_stage_name(struct v3d_compile *c)
{
if (c->vs_key && c->vs_key->is_coord)
- return "MESA_SHADER_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);
}