| 1 |
From 79c865f2310df6b843702bc47d3393d125801701 Mon Sep 17 00:00:00 2001 |
| 2 |
From: Jason Ekstrand <jason.ekstrand@intel.com> |
| 3 |
Date: Wed, 17 Oct 2018 11:34:32 -0500 |
| 4 |
Subject: [PATCH 75/78] intel/peephole_ffma: Fix swizzle propagation |
| 5 |
|
| 6 |
The num_components value passed into get_mul_for_src is used to only |
| 7 |
compose the parts of the swizzle that we know will be used so we don't |
| 8 |
compose invalid swizzle components. However, we had a bug where we |
| 9 |
passed the number of components of the add all the way through. For the |
| 10 |
given source, we need the number of components read from that source. |
| 11 |
In the case where we have a narrow add, say 2 components, that is |
| 12 |
sourced from a chain of wider instructions, we may not compose all the |
| 13 |
swizzles. All we really need to do is pass through the right number of |
| 14 |
components at each level. |
| 15 |
|
| 16 |
Fixes: 2231cf0ba3a "nir: Fix output swizzle in get_mul_for_src" |
| 17 |
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> |
| 18 |
(cherry picked from commit 1ede463b6e66eb0a6df5250261810b6985c35eb9) |
| 19 |
--- |
| 20 |
src/intel/compiler/brw_nir_opt_peephole_ffma.c | 11 +++++++---- |
| 21 |
1 file changed, 7 insertions(+), 4 deletions(-) |
| 22 |
|
| 23 |
diff --git a/src/intel/compiler/brw_nir_opt_peephole_ffma.c b/src/intel/compiler/brw_nir_opt_peephole_ffma.c |
| 24 |
index cc225e1847..7271bdbca4 100644 |
| 25 |
--- a/src/intel/compiler/brw_nir_opt_peephole_ffma.c |
| 26 |
+++ b/src/intel/compiler/brw_nir_opt_peephole_ffma.c |
| 27 |
@@ -68,7 +68,7 @@ are_all_uses_fadd(nir_ssa_def *def) |
| 28 |
} |
| 29 |
|
| 30 |
static nir_alu_instr * |
| 31 |
-get_mul_for_src(nir_alu_src *src, int num_components, |
| 32 |
+get_mul_for_src(nir_alu_src *src, unsigned num_components, |
| 33 |
uint8_t swizzle[4], bool *negate, bool *abs) |
| 34 |
{ |
| 35 |
uint8_t swizzle_tmp[4]; |
| 36 |
@@ -93,16 +93,19 @@ get_mul_for_src(nir_alu_src *src, int num_components, |
| 37 |
switch (alu->op) { |
| 38 |
case nir_op_imov: |
| 39 |
case nir_op_fmov: |
| 40 |
- alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs); |
| 41 |
+ alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components, |
| 42 |
+ swizzle, negate, abs); |
| 43 |
break; |
| 44 |
|
| 45 |
case nir_op_fneg: |
| 46 |
- alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs); |
| 47 |
+ alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components, |
| 48 |
+ swizzle, negate, abs); |
| 49 |
*negate = !*negate; |
| 50 |
break; |
| 51 |
|
| 52 |
case nir_op_fabs: |
| 53 |
- alu = get_mul_for_src(&alu->src[0], num_components, swizzle, negate, abs); |
| 54 |
+ alu = get_mul_for_src(&alu->src[0], alu->dest.dest.ssa.num_components, |
| 55 |
+ swizzle, negate, abs); |
| 56 |
*negate = false; |
| 57 |
*abs = true; |
| 58 |
break; |
| 59 |
-- |
| 60 |
2.20.1 |
| 61 |
|
Parent Directory
| 
Revision Log