From: Iago Toral Quiroga <itoral@igalia.com>
Date: Mon, 21 Jan 2019 11:11:44 +0000 (+0100)
Subject: intel/compiler: workaround for SIMD8 half-float MAD in gen8
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=0986199b31ab2a6086131887e474bc8f79fbc28d;p=mesa.git

intel/compiler: workaround for SIMD8 half-float MAD in gen8

Empirical testing shows that gen8 has a bug where MAD instructions with
a half-float source starting at a non-zero offset fail to execute
properly.

This scenario usually happened in SIMD8 executions, where we used to
pack vector components Y and W in the second half of SIMD registers
(therefore, with a 16B offset). It looks like we are not currently doing
this any more but this would handle the situation properly if we ever
happen to produce code like this again.

v2 (Jason):
 - Move this workaround to the lower_regioning pass as an additional case
   to has_invalid_src_region()
 - Do not apply the workaround if the stride of the source operand is 0,
   testing suggests the problem doesn't exist in that case.

v3 (Jason):
 - We want offset % REG_SIZE > 0, not just offset > 0
 - Use a helper to compute the offset

Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com> (v1)
---

diff --git a/src/intel/compiler/brw_fs_lower_regioning.cpp b/src/intel/compiler/brw_fs_lower_regioning.cpp
index c60d4700419..a76fd262a10 100644
--- a/src/intel/compiler/brw_fs_lower_regioning.cpp
+++ b/src/intel/compiler/brw_fs_lower_regioning.cpp
@@ -127,20 +127,37 @@ namespace {
    has_invalid_src_region(const gen_device_info *devinfo, const fs_inst *inst,
                           unsigned i)
    {
-      if (is_unordered(inst) || inst->is_control_source(i)) {
+      if (is_unordered(inst) || inst->is_control_source(i))
          return false;
-      } else {
-         const unsigned dst_byte_stride = inst->dst.stride * type_sz(inst->dst.type);
-         const unsigned src_byte_stride = inst->src[i].stride *
-            type_sz(inst->src[i].type);
-         const unsigned dst_byte_offset = reg_offset(inst->dst) % REG_SIZE;
-         const unsigned src_byte_offset = reg_offset(inst->src[i]) % REG_SIZE;
 
-         return has_dst_aligned_region_restriction(devinfo, inst) &&
-                !is_uniform(inst->src[i]) &&
-                (src_byte_stride != dst_byte_stride ||
-                 src_byte_offset != dst_byte_offset);
+      /* Empirical testing shows that Broadwell has a bug affecting half-float
+       * MAD instructions when any of its sources has a non-zero offset, such
+       * as:
+       *
+       * mad(8) g18<1>HF -g17<4,4,1>HF g14.8<4,4,1>HF g11<4,4,1>HF { align16 1Q };
+       *
+       * We used to generate code like this for SIMD8 executions where we
+       * used to pack components Y and W of a vector at offset 16B of a SIMD
+       * register. The problem doesn't occur if the stride of the source is 0.
+       */
+      if (devinfo->gen == 8 &&
+          inst->opcode == BRW_OPCODE_MAD &&
+          inst->src[i].type == BRW_REGISTER_TYPE_HF &&
+          reg_offset(inst->src[i]) % REG_SIZE > 0 &&
+          inst->src[i].stride != 0) {
+         return true;
       }
+
+      const unsigned dst_byte_stride = inst->dst.stride * type_sz(inst->dst.type);
+      const unsigned src_byte_stride = inst->src[i].stride *
+         type_sz(inst->src[i].type);
+      const unsigned dst_byte_offset = reg_offset(inst->dst) % REG_SIZE;
+      const unsigned src_byte_offset = reg_offset(inst->src[i]) % REG_SIZE;
+
+      return has_dst_aligned_region_restriction(devinfo, inst) &&
+             !is_uniform(inst->src[i]) &&
+             (src_byte_stride != dst_byte_stride ||
+              src_byte_offset != dst_byte_offset);
    }
 
    /*