pan/bi: Use FCLAMP pseudo op for clamp prop

Map nir_op_fsat/etc to FCLAMP pseudo ops, instead of FADD. There are
significantly fewer knobs on FCLAMP, meaning significantly fewer things
to get wrong.

This fixes two(!) classes of bugs:

* Swizzles (failing to lower/compose swizzles on clamps)
* Numerical bugs (incorrectly treating +0.0 as an additive identity)

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12205>
This commit is contained in:
Alyssa Rosenzweig
2021-08-04 14:49:30 -04:00
parent a98790fa96
commit ac636f5adb
4 changed files with 47 additions and 14 deletions

View File

@@ -8313,4 +8313,24 @@
</mod>
</ins>
<ins name="*FCLAMP.f32" pseudo="true">
<src start="0" mask="0xfb"/>
<mod name="clamp" start="15" size="2">
<opt>none</opt>
<opt>clamp_0_inf</opt>
<opt>clamp_m1_1</opt>
<opt>clamp_0_1</opt>
</mod>
</ins>
<ins name="*FCLAMP.v2f16" pseudo="true">
<src start="0" mask="0xfb"/>
<mod name="clamp" start="15" size="2">
<opt>none</opt>
<opt>clamp_0_inf</opt>
<opt>clamp_m1_1</opt>
<opt>clamp_0_1</opt>
</mod>
</ins>
</bifrost>

View File

@@ -33,6 +33,10 @@
static void
bi_lower_swizzle_16(bi_context *ctx, bi_instr *ins, unsigned src)
{
/* Identity is ok */
if (ins->src[src].swizzle == BI_SWIZZLE_H01)
return;
/* TODO: Use the opcode table and be a lot more methodical about this... */
switch (ins->op) {
/* Some instructions used with 16-bit data never have swizzles */
@@ -66,14 +70,24 @@ bi_lower_swizzle_16(bi_context *ctx, bi_instr *ins, unsigned src)
return;
else
break;
/* We don't want to deal with reswizzling logic in modifier prop. Move
* the swizzle outside, it's easier for clamp propagation. */
case BI_OPCODE_FCLAMP_V2F16:
{
bi_builder b = bi_init_builder(ctx, bi_after_instr(ins));
bi_index dest = ins->dest[0];
bi_index tmp = bi_temp(ctx);
ins->dest[0] = tmp;
bi_swz_v2i16_to(&b, dest, bi_replace_index(ins->src[0], tmp));
return;
}
default:
return;
}
/* Identity is ok (TODO: what about replicate only?) */
if (ins->src[src].swizzle == BI_SWIZZLE_H01)
return;
/* If the instruction is scalar we can ignore the other component */
if (ins->dest[0].swizzle == BI_SWIZZLE_H00 &&
ins->src[src].swizzle == BI_SWIZZLE_H00)

View File

@@ -157,19 +157,16 @@ bi_takes_clamp(bi_instr *I)
}
static bool
bi_is_fclamp(bi_instr *I)
bi_is_fclamp(enum bi_opcode op, enum bi_size size)
{
return (I->op == BI_OPCODE_FADD_F32 || I->op == BI_OPCODE_FADD_V2F16) &&
(!I->src[0].abs && !I->src[0].neg) &&
(I->src[1].type == BI_INDEX_CONSTANT && I->src[1].value == 0) &&
(I->clamp != BI_CLAMP_NONE);
return (size == BI_SIZE_32 && op == BI_OPCODE_FCLAMP_F32) ||
(size == BI_SIZE_16 && op == BI_OPCODE_FCLAMP_V2F16);
}
static bool
bi_optimizer_clamp(bi_instr *I, bi_instr *use)
{
if (bi_opcode_props[use->op].size != bi_opcode_props[I->op].size) return false;
if (!bi_is_fclamp(use)) return false;
if (!bi_is_fclamp(use->op, bi_opcode_props[I->op].size)) return false;
if (!bi_takes_clamp(I)) return false;
/* Clamps are bitfields (clamp_m1_1/clamp_0_inf) so composition is OR */
@@ -260,6 +257,8 @@ bi_lower_opt_instruction(bi_instr *I)
switch (I->op) {
case BI_OPCODE_FABSNEG_F32:
case BI_OPCODE_FABSNEG_V2F16:
case BI_OPCODE_FCLAMP_F32:
case BI_OPCODE_FCLAMP_V2F16:
I->op = (bi_opcode_props[I->op].size == BI_SIZE_32) ?
BI_OPCODE_FADD_F32 : BI_OPCODE_FADD_V2F16;

View File

@@ -1862,19 +1862,19 @@ bi_emit_alu(bi_builder *b, nir_alu_instr *instr)
break;
case nir_op_fsat: {
bi_instr *I = bi_fadd_to(b, sz, dst, s0, bi_negzero(), BI_ROUND_NONE);
bi_instr *I = bi_fclamp_to(b, sz, dst, s0);
I->clamp = BI_CLAMP_CLAMP_0_1;
break;
}
case nir_op_fsat_signed_mali: {
bi_instr *I = bi_fadd_to(b, sz, dst, s0, bi_negzero(), BI_ROUND_NONE);
bi_instr *I = bi_fclamp_to(b, sz, dst, s0);
I->clamp = BI_CLAMP_CLAMP_M1_1;
break;
}
case nir_op_fclamp_pos_mali: {
bi_instr *I = bi_fadd_to(b, sz, dst, s0, bi_negzero(), BI_ROUND_NONE);
bi_instr *I = bi_fclamp_to(b, sz, dst, s0);
I->clamp = BI_CLAMP_CLAMP_0_INF;
break;
}