OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

panfrost/midgard: Rework mir_adjust_constants() to make it type/size agnostic

Browse Source 
Right now, constant combining is not supported in 16 bit mode, and 64
bit mode is simply ignored. Let's rework the function to make it
type/bit-size agnostic.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3478>
This commit is contained in:
Boris Brezillon
2020-01-20 15:44:48 +01:00
committed by Marge Bot
parent 15c92d158c
commit 9566f26ed4
1 changed files with 70 additions and 95 deletions
Download Patch File Download Diff File Expand all files Collapse all files

165
src/panfrost/midgard/midgard_schedule.c
View File

@@ -333,7 +333,7 @@ struct midgard_predicate {
* will be adjusted to index into the constants array */ * will be adjusted to index into the constants array */
midgard_constants *constants; midgard_constants *constants;
unsigned constant_count; unsigned constant_mask;
bool blend_constant; bool blend_constant;
/* Exclude this destination (if not ~0) */ /* Exclude this destination (if not ~0) */
@@ -360,11 +360,11 @@ mir_adjust_constants(midgard_instruction *ins,
{ {
/* Blend constants dominate */ /* Blend constants dominate */
if (ins->has_blend_constant) { if (ins->has_blend_constant) {
if (pred->constant_count) if (pred->constant_mask)
return false; return false;
else if (destructive) { else if (destructive) {
pred->blend_constant = true; pred->blend_constant = true;
pred->constant_count = 16; pred->constant_mask = 0xffff;
return true; return true;
} }
} }
@@ -373,115 +373,90 @@ mir_adjust_constants(midgard_instruction *ins,
if (!ins->has_constants) if (!ins->has_constants)
return true; return true;
if (ins->alu.reg_mode != midgard_reg_mode_32) { unsigned r_constant = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
/* TODO: 16-bit constant combining */ midgard_reg_mode reg_mode = ins->alu.reg_mode;
if (pred->constant_count)
return false;
uint16_t *bundles = pred->constants->u16; midgard_vector_alu_src const_src = { };
const uint16_t *constants = ins->constants.u16;
/* Copy them wholesale */ if (ins->src[0] == r_constant)
for (unsigned i = 0; i < 4; ++i) const_src = vector_alu_from_unsigned(ins->alu.src1);
bundles[i] = constants[i]; else if (ins->src[1] == r_constant)
const_src = vector_alu_from_unsigned(ins->alu.src2);
pred->constant_count = 16; unsigned type_size = mir_bytes_for_mode(reg_mode);
} else {
/* Pack 32-bit constants */
uint32_t *bundles = pred->constants->u32;
const uint32_t *constants = ins->constants.u32;
unsigned r_constant = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
unsigned mask = mir_from_bytemask(mir_bytemask_of_read_components(ins, r_constant), midgard_reg_mode_32);
/* First, check if it fits */ /* If the ALU is converting up we need to divide type_size by 2 */
unsigned count = DIV_ROUND_UP(pred->constant_count, sizeof(uint32_t)); if (const_src.half)
unsigned existing_count = count; type_size /= 2;
for (unsigned i = 0; i < 4; ++i) { unsigned max_comp = 16 / type_size;
if (!(mask & (1 << i))) unsigned comp_mask = mir_from_bytemask(mir_bytemask_of_read_components(ins, r_constant),
continue; reg_mode);
unsigned type_mask = (1 << type_size) - 1;
unsigned bundle_constant_mask = pred->constant_mask;
unsigned comp_mapping[16] = { };
uint8_t bundle_constants[16];
bool ok = false; memcpy(bundle_constants, pred->constants, 16);
/* Look for existing constant */ /* Let's try to find a place for each active component of the constant
for (unsigned j = 0; j < existing_count; ++j) { * register.
if (bundles[j] == constants[i]) { */
ok = true; for (unsigned comp = 0; comp < max_comp; comp++) {
break; if (!(comp_mask & (1 << comp)))
} continue;
}
if (ok) uint8_t *constantp = ins->constants.u8 + (type_size * comp);
continue; unsigned best_reuse_bytes = 0;
signed best_place = -1;
unsigned i, j;
/* If the constant is new, check ourselves */ for (i = 0; i < 16; i += type_size) {
for (unsigned j = 0; j < i; ++j) { unsigned reuse_bytes = 0;
if (constants[j] == constants[i] && (mask & (1 << j))) {
ok = true;
break;
}
}
if (ok) for (j = 0; j < type_size; j++) {
continue; if (!(bundle_constant_mask & (1 << (i + j))))
/* Otherwise, this is a new constant */
count++;
}
/* Check if we have space */
if (count > 4)
return false;
/* If non-destructive, we're done */
if (!destructive)
return true;
/* If destructive, let's copy in the new constants and adjust
* swizzles to pack it in. */
unsigned indices[16] = { 0 };
/* Reset count */
count = existing_count;
for (unsigned i = 0; i < 4; ++i) {
if (!(mask & (1 << i)))
continue;
uint32_t cons = constants[i];
bool constant_found = false;
/* Search for the constant */
for (unsigned j = 0; j < count; ++j) {
if (bundles[j] != cons)
continue; continue;
if (constantp[j] != bundle_constants[i + j])
break;
/* We found it, reuse */ reuse_bytes++;
indices[i] = j; }
constant_found = true;
/* Select the place where existing bytes can be
* reused so we leave empty slots to others
*/
if (j == type_size &&
(reuse_bytes > best_reuse_bytes || best_place < 0)) {
best_reuse_bytes = reuse_bytes;
best_place = i;
break; break;
} }
if (constant_found)
continue;
/* We didn't find it, so allocate it */
unsigned idx = count++;
/* We have space, copy it in! */
bundles[idx] = cons;
indices[i] = idx;
} }
pred->constant_count = count * sizeof(uint32_t); /* This component couldn't fit in the remaining constant slot,
* no need check the remaining components, bail out now
*/
if (best_place < 0)
return false;
/* Use indices as a swizzle */ memcpy(&bundle_constants[i], constantp, type_size);
bundle_constant_mask |= type_mask << best_place;
comp_mapping[comp] = best_place / type_size;
}
mir_foreach_src(ins, s) { /* If non-destructive, we're done */
if (ins->src[s] == r_constant) if (!destructive)
mir_compose_swizzle(ins->swizzle[s], indices, ins->swizzle[s]); return true;
}
/* Otherwise update the constant_mask and constant values */
pred->constant_mask = bundle_constant_mask;
memcpy(pred->constants, bundle_constants, 16);
/* Use comp_mapping as a swizzle */
mir_foreach_src(ins, s) {
if (ins->src[s] == r_constant)
mir_compose_swizzle(ins->swizzle[s], comp_mapping, ins->swizzle[s]);
} }
return true; return true;
@@ -1028,7 +1003,7 @@ mir_schedule_alu(
mir_update_worklist(worklist, len, instructions, sadd); mir_update_worklist(worklist, len, instructions, sadd);
bundle.has_blend_constant = predicate.blend_constant; bundle.has_blend_constant = predicate.blend_constant;
bundle.has_embedded_constants = predicate.constant_count > 0; bundle.has_embedded_constants = predicate.constant_mask != 0;
unsigned padding = 0; unsigned padding = 0;