-
Notifications
You must be signed in to change notification settings - Fork 1.1k
fix(simplifier): VectorReduce(Broadcast(Vector)) used wrong semantics of Broadcast(Vector). #9074
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Changes from 3 commits
2f810de
88900c9
0526414
c2bef73
a8442de
01fe0d6
b4eeeef
8e25c70
4f85ddd
7680287
e1e02da
ae4024a
c5f81fc
08bfbd6
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -130,7 +130,10 @@ Expr Simplify::visit(const VectorReduce *op, ExprInfo *info) { | |
| auto rewrite = IRMatcher::rewriter(IRMatcher::h_add(value, lanes), op->type); | ||
| if (rewrite(h_add(x * broadcast(y, arg_lanes), lanes), h_add(x, lanes) * broadcast(y, lanes)) || | ||
| rewrite(h_add(broadcast(x, arg_lanes) * y, lanes), h_add(y, lanes) * broadcast(x, lanes)) || | ||
| rewrite(h_add(broadcast(x, arg_lanes), lanes), broadcast(x * factor, lanes))) { | ||
| rewrite(h_add(broadcast(x, arg_lanes), lanes), broadcast(x * factor, lanes)) || | ||
| rewrite(h_add(broadcast(x, c0), lanes), broadcast(h_add(x, lanes / c0), c0), lanes % c0 == 0) || | ||
| rewrite(h_add(broadcast(x, c0), lanes), broadcast(h_add(x, 1) * (c0 / lanes), lanes), c0 % lanes == 0) || | ||
| false) { | ||
| return mutate(rewrite.result, info); | ||
| } | ||
| break; | ||
|
|
@@ -142,8 +145,10 @@ Expr Simplify::visit(const VectorReduce *op, ExprInfo *info) { | |
| rewrite(h_min(max(x, broadcast(y, arg_lanes)), lanes), max(h_min(x, lanes), broadcast(y, lanes))) || | ||
| rewrite(h_min(max(broadcast(x, arg_lanes), y), lanes), max(h_min(y, lanes), broadcast(x, lanes))) || | ||
| rewrite(h_min(broadcast(x, arg_lanes), lanes), broadcast(x, lanes)) || | ||
| rewrite(h_min(broadcast(x, c0), lanes), h_min(x, lanes), factor % c0 == 0) || | ||
| rewrite(h_min(ramp(x, y, arg_lanes), lanes), x + min(y * (arg_lanes - 1), 0)) || | ||
| rewrite(h_min(broadcast(x, c0), 1), h_min(x, 1)) || | ||
| rewrite(h_min(broadcast(x, c0), lanes), broadcast(h_min(x, lanes / c0), c0), lanes % c0 == 0) || | ||
| rewrite(h_min(ramp(x, y, arg_lanes), 1), x + min(y * (arg_lanes - 1), 0)) || | ||
| rewrite(h_min(ramp(x, y, arg_lanes), lanes), ramp(x + min(y * (factor - 1), 0), y * factor, lanes)) || | ||
|
Member
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I think these two rules need to be guarded with no_overflow_type, because ramps are only strictly decreasing or increasing if there's no overflow. |
||
| false) { | ||
| return mutate(rewrite.result, info); | ||
| } | ||
|
|
@@ -156,8 +161,10 @@ Expr Simplify::visit(const VectorReduce *op, ExprInfo *info) { | |
| rewrite(h_max(max(x, broadcast(y, arg_lanes)), lanes), max(h_max(x, lanes), broadcast(y, lanes))) || | ||
| rewrite(h_max(max(broadcast(x, arg_lanes), y), lanes), max(h_max(y, lanes), broadcast(x, lanes))) || | ||
| rewrite(h_max(broadcast(x, arg_lanes), lanes), broadcast(x, lanes)) || | ||
| rewrite(h_max(broadcast(x, c0), lanes), h_max(x, lanes), factor % c0 == 0) || | ||
| rewrite(h_max(ramp(x, y, arg_lanes), lanes), x + max(y * (arg_lanes - 1), 0)) || | ||
| rewrite(h_max(broadcast(x, c0), 1), h_max(x, 1)) || | ||
| rewrite(h_max(broadcast(x, c0), lanes), broadcast(h_max(x, lanes / c0), c0), lanes % c0 == 0) || | ||
|
Member
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Same missing rule as above |
||
| rewrite(h_max(ramp(x, y, arg_lanes), 1), x + max(y * (arg_lanes - 1), 0)) || | ||
| rewrite(h_max(ramp(x, y, arg_lanes), lanes), ramp(x + max(y * (factor - 1), 0), y * factor, lanes)) || | ||
|
Member
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Needs to be guarded with no_overflow_type, as above |
||
| false) { | ||
| return mutate(rewrite.result, info); | ||
| } | ||
|
|
@@ -170,15 +177,16 @@ Expr Simplify::visit(const VectorReduce *op, ExprInfo *info) { | |
| rewrite(h_and(x && broadcast(y, arg_lanes), lanes), h_and(x, lanes) && broadcast(y, lanes)) || | ||
| rewrite(h_and(broadcast(x, arg_lanes) && y, lanes), h_and(y, lanes) && broadcast(x, lanes)) || | ||
| rewrite(h_and(broadcast(x, arg_lanes), lanes), broadcast(x, lanes)) || | ||
| rewrite(h_and(broadcast(x, c0), lanes), h_and(x, lanes), factor % c0 == 0) || | ||
| rewrite(h_and(ramp(x, y, arg_lanes) < broadcast(z, arg_lanes), lanes), | ||
| x + max(y * (arg_lanes - 1), 0) < z) || | ||
| rewrite(h_and(ramp(x, y, arg_lanes) <= broadcast(z, arg_lanes), lanes), | ||
| x + max(y * (arg_lanes - 1), 0) <= z) || | ||
| rewrite(h_and(broadcast(x, arg_lanes) < ramp(y, z, arg_lanes), lanes), | ||
| x < y + min(z * (arg_lanes - 1), 0)) || | ||
| rewrite(h_and(broadcast(x, arg_lanes) < ramp(y, z, arg_lanes), lanes), | ||
| x <= y + min(z * (arg_lanes - 1), 0)) || | ||
| rewrite(h_and(broadcast(x, c0), lanes), broadcast(h_and(x, lanes / c0), c0), lanes % c0 == 0) || | ||
| rewrite(h_and(broadcast(x, c0), lanes), broadcast(h_and(x, 1), lanes), c0 >= lanes) || | ||
| (lanes == 1 && rewrite(h_and(ramp(x, y, arg_lanes) < broadcast(z, arg_lanes), lanes), | ||
alexreinking marked this conversation as resolved.
Outdated
Show resolved
Hide resolved
|
||
| x + max(y * (arg_lanes - 1), 0) < z)) || | ||
| (lanes == 1 && rewrite(h_and(ramp(x, y, arg_lanes) <= broadcast(z, arg_lanes), lanes), | ||
| x + max(y * (arg_lanes - 1), 0) <= z)) || | ||
| (lanes == 1 && rewrite(h_and(broadcast(x, arg_lanes) < ramp(y, z, arg_lanes), lanes), | ||
| x < y + min(z * (arg_lanes - 1), 0))) || | ||
| (lanes == 1 && rewrite(h_and(broadcast(x, arg_lanes) < ramp(y, z, arg_lanes), lanes), | ||
| x <= y + min(z * (arg_lanes - 1), 0))) || | ||
| false) { | ||
| return mutate(rewrite.result, info); | ||
| } | ||
|
|
@@ -191,7 +199,8 @@ Expr Simplify::visit(const VectorReduce *op, ExprInfo *info) { | |
| rewrite(h_or(x && broadcast(y, arg_lanes), lanes), h_or(x, lanes) && broadcast(y, lanes)) || | ||
| rewrite(h_or(broadcast(x, arg_lanes) && y, lanes), h_or(y, lanes) && broadcast(x, lanes)) || | ||
| rewrite(h_or(broadcast(x, arg_lanes), lanes), broadcast(x, lanes)) || | ||
| rewrite(h_or(broadcast(x, c0), lanes), h_or(x, lanes), factor % c0 == 0) || | ||
| rewrite(h_or(broadcast(x, c0), lanes), broadcast(h_or(x, lanes / c0), c0), lanes % c0 == 0) || | ||
| rewrite(h_or(broadcast(x, c0), lanes), broadcast(h_or(x, 1), lanes), c0 >= lanes) || | ||
| // type of arg_lanes is somewhat indeterminate | ||
| rewrite(h_or(ramp(x, y, arg_lanes) < broadcast(z, arg_lanes), lanes), | ||
| x + min(y * (arg_lanes - 1), 0) < z) || | ||
|
|
||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It looks like the analog to the third h_add rule is missing: