* src instruction and create a mov. This is easier for cp
* to eliminate.
*
+ * NOTE: a3xx definitely seen not working with flat bary.f. Same test
+ * uses ldlv on a4xx+, so not definitive. Seems rare enough to apply
+ * everywhere.
+ *
* TODO probably opc_cat==4 is ok too
*/
if (alu->src[0].src.is_ssa &&
+ src[0]->opc != OPC_BARY_F &&
(list_length(&alu->src[0].src.ssa->uses) == 1) &&
((opc_cat(src[0]->opc) == 2) || (opc_cat(src[0]->opc) == 3))) {
src[0]->flags |= IR3_INSTR_SAT;
{
struct ir3_block *b = ctx->block;
struct ir3_instruction *base_lo, *base_hi, *addr, *src0, *src1;
- /* UBO addresses are the first driver params, but subtract 2 here to
- * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
- * is the uniforms: */
struct ir3_const_state *const_state = &ctx->so->shader->const_state;
- unsigned ubo = regid(const_state->offsets.ubo, 0) - 2;
+ unsigned ubo = regid(const_state->offsets.ubo, 0);
const unsigned ptrsz = ir3_pointer_size(ctx->compiler);
int off = 0;
struct ir3_block *b = ctx->block;
struct ir3_instruction *stl, *offset;
struct ir3_instruction * const *value;
- unsigned base, wrmask;
+ unsigned base, wrmask, ncomp;
value = ir3_get_src(ctx, &intr->src[0]);
offset = ir3_get_src(ctx, &intr->src[1])[0];
base = nir_intrinsic_base(intr);
wrmask = nir_intrinsic_write_mask(intr);
+ ncomp = ffs(~wrmask) - 1;
- /* Combine groups of consecutive enabled channels in one write
- * message. We use ffs to find the first enabled channel and then ffs on
- * the bit-inverse, down-shifted writemask to determine the length of
- * the block of enabled bits.
- *
- * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
- */
- while (wrmask) {
- unsigned first_component = ffs(wrmask) - 1;
- unsigned length = ffs(~(wrmask >> first_component)) - 1;
-
- stl = ir3_STL(b, offset, 0,
- ir3_create_collect(ctx, &value[first_component], length), 0,
- create_immed(b, length), 0);
- stl->cat6.dst_offset = first_component + base;
- stl->cat6.type = utype_src(intr->src[0]);
- stl->barrier_class = IR3_BARRIER_SHARED_W;
- stl->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
-
- array_insert(b, b->keeps, stl);
-
- /* Clear the bits in the writemask that we just wrote, then try
- * again to see if more channels are left.
- */
- wrmask &= (15 << (first_component + length));
- }
+ assert(wrmask == BITFIELD_MASK(intr->num_components));
+
+ stl = ir3_STL(b, offset, 0,
+ ir3_create_collect(ctx, value, ncomp), 0,
+ create_immed(b, ncomp), 0);
+ stl->cat6.dst_offset = base;
+ stl->cat6.type = utype_src(intr->src[0]);
+ stl->barrier_class = IR3_BARRIER_SHARED_W;
+ stl->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
+
+ array_insert(b, b->keeps, stl);
}
/* src[] = { offset }. const_index[] = { base } */
ir3_split_dest(b, dst, load, 0, intr->num_components);
}
-/* src[] = { value, offset }. const_index[] = { base, write_mask } */
+/* src[] = { value, offset }. const_index[] = { base } */
static void
emit_intrinsic_store_shared_ir3(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
struct ir3_instruction *store, *offset;
struct ir3_instruction * const *value;
- unsigned base, wrmask;
value = ir3_get_src(ctx, &intr->src[0]);
offset = ir3_get_src(ctx, &intr->src[1])[0];
- base = nir_intrinsic_base(intr);
- wrmask = nir_intrinsic_write_mask(intr);
-
- /* Combine groups of consecutive enabled channels in one write
- * message. We use ffs to find the first enabled channel and then ffs on
- * the bit-inverse, down-shifted writemask to determine the length of
- * the block of enabled bits.
- *
- * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
- */
- while (wrmask) {
- unsigned first_component = ffs(wrmask) - 1;
- unsigned length = ffs(~(wrmask >> first_component)) - 1;
-
- store = ir3_STLW(b, offset, 0,
- ir3_create_collect(ctx, &value[first_component], length), 0,
- create_immed(b, length), 0);
+ store = ir3_STLW(b, offset, 0,
+ ir3_create_collect(ctx, value, intr->num_components), 0,
+ create_immed(b, intr->num_components), 0);
- store->cat6.dst_offset = first_component + base;
- store->cat6.type = utype_src(intr->src[0]);
- store->barrier_class = IR3_BARRIER_SHARED_W;
- store->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
+ store->cat6.dst_offset = nir_intrinsic_base(intr);
+ store->cat6.type = utype_src(intr->src[0]);
+ store->barrier_class = IR3_BARRIER_SHARED_W;
+ store->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
- array_insert(b, b->keeps, store);
-
- /* Clear the bits in the writemask that we just wrote, then try
- * again to see if more channels are left.
- */
- wrmask &= (15 << (first_component + length));
- }
+ array_insert(b, b->keeps, store);
}
/*
}
static struct ir3_instruction *
-get_frag_coord(struct ir3_context *ctx)
+get_frag_coord(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
if (!ctx->frag_coord) {
struct ir3_block *b = ctx->in_block;
}
ctx->frag_coord = ir3_create_collect(ctx, xyzw, 4);
- ctx->so->frag_coord = true;
}
+ ctx->so->fragcoord_compmask |=
+ nir_ssa_def_components_read(&intr->dest.ssa);
+
return ctx->frag_coord;
}
case nir_intrinsic_end_patch_ir3:
assert(ctx->so->type == MESA_SHADER_TESS_CTRL);
- struct ir3_instruction *end = ir3_ENDIF(b);
+ struct ir3_instruction *end = ir3_PREDE(b);
array_insert(b, b->keeps, end);
end->barrier_class = IR3_BARRIER_EVERYTHING;
emit_intrinsic_load_ubo_ldc(ctx, intr, dst);
break;
case nir_intrinsic_load_frag_coord:
- ir3_split_dest(b, dst, get_frag_coord(ctx), 0, 4);
+ ir3_split_dest(b, dst, get_frag_coord(ctx, intr), 0, 4);
break;
case nir_intrinsic_load_sample_pos_from_id: {
/* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
/* condition always goes in predicate register: */
cond->regs[0]->num = regid(REG_P0, 0);
- kill = ir3_IF(b, cond, 0);
+ kill = ir3_PREDT(b, cond, 0);
kill->barrier_class = IR3_BARRIER_EVERYTHING;
kill->barrier_conflict = IR3_BARRIER_EVERYTHING;
compile_assert(ctx, nir_tex_instr_src_index(tex, nir_tex_src_texture_offset) < 0);
compile_assert(ctx, nir_tex_instr_src_index(tex, nir_tex_src_sampler_offset) < 0);
- if (ctx->so->num_sampler_prefetch < IR3_MAX_SAMPLER_PREFETCH) {
+ if (ctx->so->num_sampler_prefetch < ctx->prefetch_limit) {
opc = OPC_META_TEX_PREFETCH;
ctx->so->num_sampler_prefetch++;
break;
struct ir3_instruction *base, *out, *stg;
base = bases[strmout->output[i].output_buffer];
- out = ctx->ir->outputs[regid(strmout->output[i].register_index, c)];
+ out = ctx->outputs[regid(strmout->output[i].register_index, c)];
stg = ir3_STG(ctx->block, base, 0, out, 0,
create_immed(ctx->block, 1), 0);
ctx->inputs[idx] = instr;
}
} else if (ctx->so->type == MESA_SHADER_VERTEX) {
- struct ir3_instruction *input = NULL, *in;
+ struct ir3_instruction *input = NULL;
struct ir3_instruction *components[4];
unsigned mask = (1 << (ncomp + frac)) - 1;
setup_output(struct ir3_context *ctx, nir_variable *out)
{
struct ir3_shader_variant *so = ctx->so;
- unsigned ncomp = glsl_get_components(out->type);
+ unsigned slots = glsl_count_vec4_slots(out->type, false, false);
+ unsigned ncomp = glsl_get_components(glsl_without_array(out->type));
unsigned n = out->data.driver_location;
unsigned frac = out->data.location_frac;
unsigned slot = out->data.location;
so->writes_smask = true;
break;
default:
+ slot += out->data.index; /* For dual-src blend */
if (slot >= FRAG_RESULT_DATA0)
break;
ir3_context_error(ctx, "unknown FS output name: %s\n",
ir3_context_error(ctx, "unknown shader type: %d\n", ctx->so->type);
}
- compile_assert(ctx, n < ARRAY_SIZE(so->outputs));
- so->outputs[n].slot = slot;
- so->outputs[n].regid = regid(n, 0);
- so->outputs_count = MAX2(so->outputs_count, n + 1);
+ so->outputs_count = out->data.driver_location + slots;
+ compile_assert(ctx, so->outputs_count < ARRAY_SIZE(so->outputs));
- for (int i = 0; i < ncomp; i++) {
- unsigned idx = (n * 4) + i + frac;
- compile_assert(ctx, idx < ctx->noutputs);
- ctx->outputs[idx] = create_immed(ctx->block, fui(0.0));
- }
+ for (int i = 0; i < slots; i++) {
+ int slot_base = n + i;
+ so->outputs[slot_base].slot = slot + i;
- /* if varying packing doesn't happen, we could end up in a situation
- * with "holes" in the output, and since the per-generation code that
- * sets up varying linkage registers doesn't expect to have more than
- * one varying per vec4 slot, pad the holes.
- *
- * Note that this should probably generate a performance warning of
- * some sort.
- */
- for (int i = 0; i < frac; i++) {
- unsigned idx = (n * 4) + i;
- if (!ctx->outputs[idx]) {
+ for (int i = 0; i < ncomp; i++) {
+ unsigned idx = (slot_base * 4) + i + frac;
+ compile_assert(ctx, idx < ctx->noutputs);
ctx->outputs[idx] = create_immed(ctx->block, fui(0.0));
}
+
+ /* if varying packing doesn't happen, we could end up in a situation
+ * with "holes" in the output, and since the per-generation code that
+ * sets up varying linkage registers doesn't expect to have more than
+ * one varying per vec4 slot, pad the holes.
+ *
+ * Note that this should probably generate a performance warning of
+ * some sort.
+ */
+ for (int i = 0; i < frac; i++) {
+ unsigned idx = (slot_base * 4) + i;
+ if (!ctx->outputs[idx]) {
+ ctx->outputs[idx] = create_immed(ctx->block, fui(0.0));
+ }
+ }
}
}
so->outputs[j] = so->outputs[i];
/* fixup outidx to point to new output table entry: */
- struct ir3_instruction *out;
foreach_output (out, ir) {
if (out->collect.outidx == i) {
out->collect.outidx = j;
struct ir3_context *ctx;
struct ir3 *ir;
int ret = 0, max_bary;
+ bool progress;
assert(!so->ir);
}
}
- /* at this point, for binning pass, throw away unneeded outputs: */
- if (so->binning_pass && (ctx->compiler->gpu_id < 600))
- fixup_binning_pass(ctx);
-
- ir3_debug_print(ir, "BEFORE CF");
-
- ir3_cf(ir);
-
- ir3_debug_print(ir, "BEFORE CP");
-
- ir3_cp(ir, so);
-
- /* at this point, for binning pass, throw away unneeded outputs:
- * Note that for a6xx and later, we do this after ir3_cp to ensure
- * that the uniform/constant layout for BS and VS matches, so that
- * we can re-use same VS_CONST state group.
- */
- if (so->binning_pass && (ctx->compiler->gpu_id >= 600))
- fixup_binning_pass(ctx);
-
/* for a6xx+, binning and draw pass VS use same VBO state, so we
* need to make sure not to remove any inputs that are used by
* the nonbinning VS.
}
}
- ir3_debug_print(ir, "BEFORE GROUPING");
+ /* at this point, for binning pass, throw away unneeded outputs: */
+ if (so->binning_pass && (ctx->compiler->gpu_id < 600))
+ fixup_binning_pass(ctx);
- ir3_sched_add_deps(ir);
+ ir3_debug_print(ir, "AFTER: nir->ir3");
- /* Group left/right neighbors, inserting mov's where needed to
- * solve conflicts:
- */
- ir3_group(ir);
+ do {
+ progress = false;
- ir3_debug_print(ir, "AFTER GROUPING");
+ progress |= IR3_PASS(ir, ir3_cf);
+ progress |= IR3_PASS(ir, ir3_cp, so);
+ progress |= IR3_PASS(ir, ir3_dce, so);
+ } while (progress);
- ir3_dce(ir, so);
+ /* at this point, for binning pass, throw away unneeded outputs:
+ * Note that for a6xx and later, we do this after ir3_cp to ensure
+ * that the uniform/constant layout for BS and VS matches, so that
+ * we can re-use same VS_CONST state group.
+ */
+ if (so->binning_pass && (ctx->compiler->gpu_id >= 600)) {
+ fixup_binning_pass(ctx);
+ /* cleanup the result of removing unneeded outputs: */
+ while (IR3_PASS(ir, ir3_dce, so)) {}
+ }
- ir3_debug_print(ir, "AFTER DCE");
+ IR3_PASS(ir, ir3_sched_add_deps);
- /* do Sethi–Ullman numbering before scheduling: */
- ir3_sun(ir);
+ /* Group left/right neighbors, inserting mov's where needed to
+ * solve conflicts:
+ */
+ IR3_PASS(ir, ir3_group);
+
+ /* At this point, all the dead code should be long gone: */
+ assert(!IR3_PASS(ir, ir3_dce, so));
ret = ir3_sched(ir);
if (ret) {
goto out;
}
- ir3_debug_print(ir, "AFTER SCHED");
+ ir3_debug_print(ir, "AFTER: ir3_sched");
/* Pre-assign VS inputs on a6xx+ binning pass shader, to align
* with draw pass VS, so binning and draw pass can both use the
ret = ir3_ra(so, precolor, ARRAY_SIZE(precolor));
} else if (so->num_sampler_prefetch) {
assert(so->type == MESA_SHADER_FRAGMENT);
- struct ir3_instruction *instr, *precolor[2];
+ struct ir3_instruction *precolor[2];
int idx = 0;
foreach_input (instr, ir) {
goto out;
}
- ir3_postsched(ctx);
- ir3_debug_print(ir, "AFTER POSTSCHED");
+ IR3_PASS(ir, ir3_postsched);
if (compiler->gpu_id >= 600) {
- if (ir3_a6xx_fixup_atomic_dests(ir, so)) {
- ir3_debug_print(ir, "AFTER ATOMIC FIXUP");
- }
+ IR3_PASS(ir, ir3_a6xx_fixup_atomic_dests, so);
}
if (so->type == MESA_SHADER_FRAGMENT)
for (unsigned i = 0; i < so->outputs_count; i++)
so->outputs[i].regid = INVALID_REG;
- struct ir3_instruction *out;
foreach_output (out, ir) {
assert(out->opc == OPC_META_COLLECT);
unsigned outidx = out->collect.outidx;
so->outputs[outidx].half = !!(out->regs[0]->flags & IR3_REG_HALF);
}
- struct ir3_instruction *in;
foreach_input (in, ir) {
assert(in->opc == OPC_META_INPUT);
unsigned inidx = in->input.inidx;
/* We need to do legalize after (for frag shader's) the "bary.f"
* offsets (inloc) have been assigned.
*/
- ir3_legalize(ir, so, &max_bary);
-
- ir3_debug_print(ir, "AFTER LEGALIZE");
+ IR3_PASS(ir, ir3_legalize, so, &max_bary);
/* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
* know what we might have to wait on when coming in from VS chsh.
if (so->type == MESA_SHADER_FRAGMENT)
so->total_in = max_bary + 1;
- so->max_sun = ir->max_sun;
-
/* Collect sampling instructions eligible for pre-dispatch. */
collect_tex_prefetches(ctx, ir);