Mesa (staging/20.0): ac/llvm: fix 16-bit fmed3 on GFX8 and older gens

[GitLab Mirror]

Module: Mesa
Branch: staging/20.0
Commit: 710388f0067d978b932bf647a398221c242cf6ad
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=710388f0067d978b932bf647a398221c242cf6ad

Author: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Date:   Wed Feb 26 15:09:40 2020 +0100

ac/llvm: fix 16-bit fmed3 on GFX8 and older gens

16-bit med3 is only supported on GFX9+.

Fixes dEQP-VK.spirv_assembly.instruction.amd_trinary_minmax.mid3.f16.*.

Fixes: d6a07732c9c ("ac: use llvm.amdgcn.fmed3 intrinsic for nir_op_fmed3")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3962>
(cherry picked from commit 30ac733680c3dfbfd1300c5498dd1b0c0a680905)

---

 .pick_status.json            | 2 +-
 src/amd/llvm/ac_llvm_build.c | 6 ++++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/.pick_status.json b/.pick_status.json
index 43f623965f6..09b0c3448ba 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -1057,7 +1057,7 @@
         "description": "ac/llvm: fix 16-bit fmed3 on GFX8 and older gens",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "master_sha": null,
         "because_sha": "d6a07732c9c155c73f7d2cddc10faa7eab768df9"
     },
diff --git a/src/amd/llvm/ac_llvm_build.c b/src/amd/llvm/ac_llvm_build.c
index db7964d6aa9..116abf942c2 100644
--- a/src/amd/llvm/ac_llvm_build.c
+++ b/src/amd/llvm/ac_llvm_build.c
@@ -2725,8 +2725,10 @@ LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
 {
 	LLVMValueRef result;
 
-	if (bitsize == 64) {
-		/* Lower 64-bit fmed because LLVM doesn't expose an intrinsic. */
+	if (bitsize == 64 || (bitsize == 16 && ctx->chip_class <= GFX8)) {
+		/* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
+		 * or lower 16-bit fmed because it's only supported on GFX9+.
+		 */
 		LLVMValueRef min1, min2, max1;
 
 		min1 = ac_build_fmin(ctx, src0, src1);