Add support to SEL instruction and add an assert to detect unsupported
instructions than do d2x conversions.
Signed-off-by: Samuel Iglesias Gonsálvez <sigles...@igalia.com>
---
Curro, this patch legalizes SEL instruction too. If other optimizations
modify later any SEL's (or any other instruction's) destination type
(hence, producing a non-lowered d2x conversion), we can call it again
around the end of fs_visitor::optimize(). Possibly together with
lower_simd_width() just in case it was added later.
For that reason there is the inst->dst.stride > 1 condition in the
test. This detects if either we emitted a strided destination in
purpose or it was as a result of a previous lower_d2x run, we don't
want to lowered it.
However, as I have not hit that case yet, I prefer to wait for your
opinion. What do you think?
src/mesa/drivers/dri/i965/brw_fs_lower_d2x.cpp | 57 ++++++++++++++++++--------
1 file changed, 41 insertions(+), 16 deletions(-)
diff --git a/src/mesa/drivers/dri/i965/brw_fs_lower_d2x.cpp
b/src/mesa/drivers/dri/i965/brw_fs_lower_d2x.cpp
index a2db1154615..330f2552929 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_lower_d2x.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_lower_d2x.cpp
@@ -33,17 +33,9 @@ fs_visitor::lower_d2x()
bool progress = false;
foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
- if (inst->opcode != BRW_OPCODE_MOV)
- continue;
-
- if (inst->dst.type != BRW_REGISTER_TYPE_F &&
- inst->dst.type != BRW_REGISTER_TYPE_D &&
- inst->dst.type != BRW_REGISTER_TYPE_UD)
- continue;
-
- if (inst->src[0].type != BRW_REGISTER_TYPE_DF &&
- inst->src[0].type != BRW_REGISTER_TYPE_UQ &&
- inst->src[0].type != BRW_REGISTER_TYPE_Q)
+ if (get_exec_type_size(inst) != 8 ||
+ type_sz(inst->dst.type) >= get_exec_type_size(inst) ||
+ inst->dst.stride > 1)
continue;
assert(inst->dst.file == VGRF);
@@ -61,13 +53,46 @@ fs_visitor::lower_d2x()
* So we need to allocate a temporary that's two registers, and then do
* a strided MOV to get the lower DWord of every Qword that has the
* result.
+ *
+ * This pass legalizes all the DF conversions to narrower types.
*/
- fs_reg temp = ibld.vgrf(inst->src[0].type, 1);
- fs_reg strided_temp = subscript(temp, inst->dst.type, 0);
- ibld.MOV(strided_temp, inst->src[0]);
- ibld.MOV(dst, strided_temp);
+ switch (inst->opcode) {
+ case SHADER_OPCODE_MOV_INDIRECT:
+ case BRW_OPCODE_MOV: {
+ fs_reg temp = ibld.vgrf(inst->src[0].type, 1);
+ fs_reg strided_temp = subscript(temp, inst->dst.type, 0);
+ /* We clone the original instruction as we are going to modify it
+ * and emit another one after it.
+ */
+ fs_inst *strided_inst = new(ibld.shader->mem_ctx) fs_inst(*inst);
+ strided_inst->dst = strided_temp;
+ /* As it is an strided destination, we write n-times more
+ * being n the size difference between source and destination types.
+ */
+ strided_inst->size_written *= (type_sz(inst->src[0].type) /
type_sz(inst->dst.type));
+ ibld.emit(strided_inst);
+ ibld.MOV(dst, strided_temp);
+ /* Remove original instruction, now that is superseded. */
+ inst->remove(block);
+ break;
+ }
+ case BRW_OPCODE_SEL: {
+ fs_reg temp0 = ibld.vgrf(inst->src[0].type, 1);
+ fs_reg strided_temp0 = subscript(temp0, inst->dst.type, 0);
+ fs_reg temp1 = ibld.vgrf(inst->src[1].type, 1);
+ fs_reg strided_temp1 = subscript(temp1, inst->dst.type, 0);
- inst->remove(block);
+ /* Let's convert the operands to the destination type first */
+ ibld.MOV(strided_temp0, inst->src[0]);
+ ibld.MOV(strided_temp1, inst->src[1]);
+ inst->src[0] = strided_temp0;
+ inst->src[1] = strided_temp1;
+ break;
+ }
+ default:
+ /* FIXME: If this is not a supported instruction, then we need to
support it. */
+ unreachable("Not supported instruction");
+ }
progress = true;
}
--
2.11.0
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
