<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
<html>
<body>
<div dir="auto">
<div dir="auto"><br></div>
<div id="aqm-original" style="color: black;">
<!-- body start -->
<div style="text-align:left; direction:ltr;" class="aqm-original-body"><div style="color: black;">
<p style="color: black; font-size: 10pt; font-family: sans-serif; margin: 8pt 0;">On January 29, 2019 05:27:50 Iago Toral <itoral@igalia.com> wrote:</p>
<blockquote type="cite" class="gmail_quote" style="margin: 0 0 0 0.75ex; border-left: 1px solid #808080; padding-left: 0.75ex;">
<div>On Thu, 2019-01-17 at 18:18 -0600, Jason Ekstrand wrote:</div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div dir="ltr">On Tue, Jan 15, 2019 at 7:55 AM Iago Toral Quiroga <<a href="mailto:itoral@igalia.com">itoral@igalia.com</a>> wrote:<br></div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex">Reviewed-by: Topi Pohjolainen <<a href="mailto:topi.pohjolainen@intel.com" target="_blank">topi.pohjolainen@intel.com</a>><br>
---<br>
.../compiler/brw_fs_combine_constants.cpp | 60 +++++++++++++++----<br>
1 file changed, 49 insertions(+), 11 deletions(-)<br>
<br>
diff --git a/src/intel/compiler/brw_fs_combine_constants.cpp b/src/intel/compiler/brw_fs_combine_constants.cpp<br>
index 7343f77bb45..54017e5668b 100644<br>
--- a/src/intel/compiler/brw_fs_combine_constants.cpp<br>
+++ b/src/intel/compiler/brw_fs_combine_constants.cpp<br>
@@ -36,6 +36,7 @@<br>
<br>
#include "brw_fs.h"<br>
#include "brw_cfg.h"<br>
+#include "util/half_float.h"<br>
<br>
using namespace brw;<br>
<br>
@@ -114,8 +115,9 @@ struct imm {<br>
*/<br>
exec_list *uses;<br>
<br>
- /** The immediate value. We currently only handle floats. */<br>
+ /** The immediate value. We currently only handle float and half-float. */<br>
float val;<br>
+ brw_reg_type type;<br></blockquote><div><br></div><div>I had a brief chat with Matt today and I think that this may be going in the wrong direction. In particular, I'd like us to eventually (maybe we can do it now?) generalize the combine_constants pass to more data types; in particular, integers. I recently came across a shader where the fact that we couldn't do combine_constants on integers was causing significant register pressure problems and spilling. (The test was doing a bunch of BFI2/BFE with constant sources.) It could also be a huge win for 8-bit and 64-bit where we can't put immediates in regular 2-src instructions.</div></div></div><div dir="ltr"><div class="gmail_quote"><div>What does this mean for the pass? I suspect we want a bit size instead of a type and a simple char[8] for the data and just make it a blob of bits. We may also want some sort of heuristic so we don't burn constant table space for things that are only used once or maybe even twice.</div></div></div></blockquote><div><br></div><div><div>I started working on it yesterday and have a skeleton that is currently capable of handling F, HF, D, UD, W, UW immediates. Seems to be working going by a bunch of selected tests from CTS I have around but I'd like to discuss a few things before I move forward:</div><div><br></div><div>The combine constants pass is very specific about the instructions it acts on. Currently it only acts in two scenarios:</div><div> 1. When we want to take advantage of co-issuing (this is gen7 only apparently).</div><div> 2. When we need to fix instructions that can't take an immediate argument (such as MAD or LRP)</div><div><br></div><div>In order to fix the issue you mention above with BFI2/BFE, we only need to let constant propagation happen on these instructions and the add them to the list of opcodes under must_promote() like we do for LRP and MAD (plus add the code to support combinationof D/UD immediates of course). That's what I am doing right now. The moment we decide to let constant propagation happen for these instructions then we must promote the immediates to GRF by the time we run the combine pass, so heuristics won't play any role in this, as they don't for MAD or LRP. However, you mentioned heuristics specifically, so that makes me wonder if your idea was to have the pass promote immediates for any instruction in the program (that support immediates, since that would be the only case in which we're free ot make a choice), but I don't think there is any benefit in doing that, right?</div></div></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">I see no reason why we shouldn't do constant promotion for all 3src instructions. We could also do it for 2src where the constant is required to be in the source that doesn't such as 1 << x. Of course, we'll have to run shader-db on whatever we do to see the effect.</div><div dir="auto"><br></div><div id="aqm-original" style="color: black;" dir="auto"><div style="text-align:left; direction:ltr;" class="aqm-original-body"><div style="color: black;"><blockquote type="cite" class="gmail_quote" style="margin: 0 0 0 0.75ex; border-left: 1px solid #808080; padding-left: 0.75ex;"><div><div></div><div>Also, I think we still need a type instead of a size, I see a few reasons for this:</div><div> - The pass stores the absolute value of the constants it finds and then relies on the signbit function (which only works for 32-bit and 64-bit floats) to decide whether it should negate the source whe it rewrites it. This allows us to promote an immediate and its negated version at the same time. As we start to support more types, including unsigned integer types, I think we need to know the ty¡pe to do the right thing. I think that we should at least know if the type is unsigned.</div></div></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Really, you only need to consider the sign relative to the type being used to read the constant. If someone actually wanted the integer 0x3f800000, the same constant could be used to store that, 1.0f, and -1.0f.</div><div dir="auto"><br></div><div id="aqm-original" style="color: black;" dir="auto"><div style="text-align:left; direction:ltr;" class="aqm-original-body"><div style="color: black;"><blockquote type="cite" class="gmail_quote" style="margin: 0 0 0 0.75ex; border-left: 1px solid #808080; padding-left: 0.75ex;"><div><div></div><div> - There is an assertion in the pass that is floating-point specific: before it rewrites an immediate source to read from the promoted constant, it checks that the absolute value of the original immediate source matches what we stored in the table, or that both are NaN. This assertion doesn't make sense as is for integers, so we need to either know the type to have a type-specific assertion or drop the assertion.</div></div></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">This sounds like it's just a sanity check for "in this constant the thing I stored the first time through the pass" type of thing (I'm not looking at the code right now) to make sure it didn't messes up its internal bookkeeping. If that's the case, it really just needs to be updated.</div><div dir="auto"><br></div><div id="aqm-original" style="color: black;" dir="auto"><div style="text-align:left; direction:ltr;" class="aqm-original-body"><div style="color: black;"><blockquote type="cite" class="gmail_quote" style="margin: 0 0 0 0.75ex; border-left: 1px solid #808080; padding-left: 0.75ex;"><div><div></div><div> - Some restrictions take place only on specific types. For example, there is a restriction on BDW that affects ALIGN16 with half-float, and we need to account for that in the pass when we promote the constant, but there is no such restriction for W or UW.</div></div></blockquote></div></div></div><div dir="auto"><br></div><div dir="auto">Uh... That's annoying. My initial reaction is to apply the restriction to all 16-bit things and hope it doesn't hurt too much. We could also add a bit to the constant table entry that says "I need the half-float restriction" and set it whenever an entry is used as HF.</div><div dir="auto"><br></div><div dir="auto">--Jason</div><div dir="auto"><br></div><div id="aqm-original" style="color: black;" dir="auto"><div style="text-align:left; direction:ltr;" class="aqm-original-body"><div style="color: black;"><blockquote type="cite" class="gmail_quote" style="margin: 0 0 0 0.75ex; border-left: 1px solid #808080; padding-left: 0.75ex;"><div><div></div><div>Thoughts?</div><div><br></div><div>Iago</div><div><br></div></div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><div>Normally, I would say "do it as a fixup" but if we go the direction of having a float and using _mesa_half_to_float and _mesa_float_to_half, I suspect it'll be harder to go for the bag-of-bits approach.<br></div><div><br></div><div>Thoughts?</div><div><br></div><div>--Jason<br></div><div> </div><blockquote type="cite" style="margin:0 0 0 .8ex; border-left:2px #729fcf solid;padding-left:1ex">
<br>
/**<br>
* The GRF register and subregister number where we've decided to store the<br>
@@ -145,10 +147,10 @@ struct table {<br>
};<br>
<br>
static struct imm *<br>
-find_imm(struct table *table, float val)<br>
+find_imm(struct table *table, float val, brw_reg_type type)<br>
{<br>
for (int i = 0; i < table->len; i++) {<br>
- if (table->imm[i].val == val) {<br>
+ if (table->imm[i].val == val && table->imm[i].type == type) {<br>
return &table->imm[i];<br>
}<br>
}<br>
@@ -190,6 +192,20 @@ compare(const void *_a, const void *_b)<br>
return a->first_use_ip - b->first_use_ip;<br>
}<br>
<br>
+static bool<br>
+needs_negate(float reg_val, float imm_val, brw_reg_type type)<br>
+{<br>
+ /* reg_val represents the immediate value in the register in its original<br>
+ * bit-size, while imm_val is always a valid 32-bit float value.<br>
+ */<br>
+ if (type == BRW_REGISTER_TYPE_HF) {<br>
+ uint32_t reg_val_ud = *((uint32_t *) ®_val);<br>
+ reg_val = _mesa_half_to_float(reg_val_ud & 0xffff);<br>
+ }<br>
+<br>
+ return signbit(imm_val) != signbit(reg_val);<br>
+}<br>
+<br>
bool<br>
fs_visitor::opt_combine_constants()<br>
{<br>
@@ -215,12 +231,20 @@ fs_visitor::opt_combine_constants()<br>
<br>
for (int i = 0; i < inst->sources; i++) {<br>
if (inst->src[i].file != IMM ||<br>
- inst->src[i].type != BRW_REGISTER_TYPE_F)<br>
+ (inst->src[i].type != BRW_REGISTER_TYPE_F &&<br>
+ inst->src[i].type != BRW_REGISTER_TYPE_HF))<br>
continue;<br>
<br>
- float val = !inst->can_do_source_mods(devinfo) ? inst->src[i].f :<br>
- fabs(inst->src[i].f);<br>
- struct imm *imm = find_imm(&table, val);<br>
+ float val;<br>
+ if (inst->src[i].type == BRW_REGISTER_TYPE_F) {<br>
+ val = !inst->can_do_source_mods(devinfo) ? inst->src[i].f :<br>
+ fabs(inst->src[i].f);<br>
+ } else {<br>
+ val = !inst->can_do_source_mods(devinfo) ?<br>
+ _mesa_half_to_float(inst->src[i].d & 0xffff) :<br>
+ fabs(_mesa_half_to_float(inst->src[i].d & 0xffff));<br>
+ }<br>
+ struct imm *imm = find_imm(&table, val, inst->src[i].type);<br>
<br>
if (imm) {<br>
bblock_t *intersection = cfg_t::intersect(block, imm->block);<br>
@@ -238,6 +262,7 @@ fs_visitor::opt_combine_constants()<br>
imm->uses = new(const_ctx) exec_list();<br>
imm->uses->push_tail(link(const_ctx, &inst->src[i]));<br>
imm->val = val;<br>
+ imm->type = inst->src[i].type;<br>
imm->uses_by_coissue = could_coissue(devinfo, inst);<br>
imm->must_promote = must_promote_imm(devinfo, inst);<br>
imm->first_use_ip = ip;<br>
@@ -278,12 +303,23 @@ fs_visitor::opt_combine_constants()<br>
imm->block->last_non_control_flow_inst()->next);<br>
const fs_builder ibld = <a href="http://bld.at" rel="noreferrer" target="_blank">bld.at</a>(imm->block, n).exec_all().group(1, 0);<br>
<br>
- ibld.MOV(reg, brw_imm_f(imm->val));<br>
+ reg = retype(reg, imm->type);<br>
+ if (imm->type == BRW_REGISTER_TYPE_F) {<br>
+ ibld.MOV(reg, brw_imm_f(imm->val));<br>
+ } else {<br>
+ const uint16_t val_hf = _mesa_float_to_half(imm->val);<br>
+ ibld.MOV(reg, retype(brw_imm_uw(val_hf), BRW_REGISTER_TYPE_HF));<br>
+ }<br>
imm->nr = <a href="http://reg.nr" rel="noreferrer" target="_blank">reg.nr</a>;<br>
imm->subreg_offset = reg.offset;<br>
<br>
+ /* Keep offsets 32-bit aligned since we are mixing 32-bit and 16-bit<br>
+ * constants into the same register<br>
+ *<br>
+ * TODO: try to pack pairs of HF constants into each 32-bit slot<br>
+ */<br>
reg.offset += sizeof(float);<br>
- if (reg.offset == 8 * sizeof(float)) {<br>
+ if (reg.offset == REG_SIZE) {<br>
<a href="http://reg.nr" rel="noreferrer" target="_blank">reg.nr</a> = alloc.allocate(1);<br>
reg.offset = 0;<br>
}<br>
@@ -295,12 +331,14 @@ fs_visitor::opt_combine_constants()<br>
foreach_list_typed(reg_link, link, link, table.imm[i].uses) {<br>
fs_reg *reg = link->reg;<br>
assert((isnan(reg->f) && isnan(table.imm[i].val)) ||<br>
- fabsf(reg->f) == fabs(table.imm[i].val));<br>
+ fabsf(reg->f) == fabs(table.imm[i].val) ||<br>
+ table.imm[i].type == BRW_REGISTER_TYPE_HF);<br>
<br>
reg->file = VGRF;<br>
+ reg->type = table.imm[i].type;<br>
reg->offset = table.imm[i].subreg_offset;<br>
reg->stride = 0;<br>
- reg->negate = signbit(reg->f) != signbit(table.imm[i].val);<br>
+ reg->negate = needs_negate(reg->f, table.imm[i].val, table.imm[i].type);<br>
reg->nr = table.imm[i].nr;<br>
}<br>
}<br>
</blockquote></div></div></blockquote></blockquote>
</div>
</div>
<!-- body end -->
</div><div dir="auto"><br></div>
</div></body>
</html>