int
vir_get_implicit_uniform_src(struct qinst *inst)
{
+ if (!vir_has_implicit_uniform(inst))
+ return -1;
return vir_get_nsrc(inst) - 1;
}
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->dst.file == QFILE_MAGIC)
return v3d_qpu_magic_waddr_is_tmu(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;
}
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;
}
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)
{
struct nir_lower_tex_options tex_options = {
.lower_txd = true,
- .lower_rect = false, /* XXX */
+ .lower_rect = false, /* XXX: Use this on V3D 3.x */
.lower_txp = ~0,
/* Apply swizzles to all samplers. */
.swizzle_result = ~0,
{
prog_data->threads = c->threads;
prog_data->single_seg = !c->last_thrsw;
+ prog_data->spill_size = c->spill_size;
v3d_set_prog_data_uniforms(c, prog_data);
v3d_set_prog_data_ubo(c, prog_data);
prog_data->vpm_input_size += c->vattr_sizes[i];
}
- /* 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;
-
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));
+ 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->num_vpm_writes, 8) / 8;
+
+ /* 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 = 16 * 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_batches = half_vpm - prog_data->vpm_input_size;
+ assert(vpm_output_batches >= 2);
+ prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
+
return v3d_return_qpu_insts(c, final_assembly_size);
}
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_fixup_fs_output_types(struct v3d_compile *c)
+{
+ nir_foreach_variable(var, &c->s->outputs) {
+ 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));
+ }
}
}
c->fs_key = key;
+ if (key->int_color_rb || key->uint_color_rb)
+ v3d_fixup_fs_output_types(c);
+
v3d_lower_nir(c);
if (key->light_twoside)
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;
+ prog_data->discard = (c->s->info.fs.uses_discard ||
+ c->fs_key->sample_alpha_to_coverage);
+ prog_data->uses_center_w = c->uses_center_w;
return v3d_return_qpu_insts(c, final_assembly_size);
}
list_del(&qinst->link);
free(qinst);
+
+ c->live_intervals_valid = false;
}
struct qreg
return vir_reg(QFILE_UNIF, uniform);
}
+static bool
+vir_can_set_flags(struct v3d_compile *c, struct qinst *inst)
+{
+ if (c->devinfo->ver >= 40 && (v3d_qpu_reads_vpm(&inst->qpu) ||
+ v3d_qpu_uses_sfu(&inst->qpu))) {
+ return false;
+ }
+
+ return true;
+}
+
void
vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf)
{
if (src.file != QFILE_TEMP ||
!c->defs[src.index] ||
- last_inst != c->defs[src.index]) {
+ last_inst != c->defs[src.index] ||
+ !vir_can_set_flags(c, last_inst)) {
/* XXX: Make the MOV be the appropriate type */
last_inst = vir_MOV_dest(c, vir_reg(QFILE_NULL, 0), src);
}
OPTPASS(vir_opt_copy_propagate);
OPTPASS(vir_opt_dead_code);
+ OPTPASS(vir_opt_small_immediates);
if (!progress)
break;