Modified from https://github.com/ROCm/llvm-project/commit/794ebeffcafbf6f4d86cb1bfd7a5a0d1d30f1fc7 From 794ebeffcafbf6f4d86cb1bfd7a5a0d1d30f1fc7 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Thu, 4 Jan 2024 20:40:30 +0700 Subject: [PATCH] device-libs: Fix native_rsqrt test failures for f16-as-f32 case At some point the code quality regressed for these. Change-Id: Ib5c4c1d093a6056dd6213ed6eaf157ad806101fb --- amd/device-libs/test/compile/native_rsqrt.cl | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/test/compile/native_rsqrt.cl b/test/compile/native_rsqrt.cl index 94ef375e4d522f..291cc6f32c3351 100644 --- a/test/compile/native_rsqrt.cl +++ b/test/compile/native_rsqrt.cl @@ -6,16 +6,21 @@ half __ocml_native_rsqrt_f16(half); -// GCN: {{^}}test_native_rsqrt_f16: -// GFX600: v_rsq_f32 -// GFX700: v_rsq_f32 +// FIXME: Promoted case using full expansion +// GCN-LABEL: {{^}}test_native_rsqrt_f16: +// GFX600: v_sqrt_f32 +// GFX600: v_rcp_f32 + +// GFX700: v_sqrt_f32 +// GFX700: v_rcp_f32 + // GFX803: v_rsq_f16 kernel void test_native_rsqrt_f16(global half* restrict out, global half* restrict in) { int id = get_local_id(0); out[id] = __ocml_native_rsqrt_f16(in[id]); } -// GCN: {{^}}test_native_rsqrt_f32: +// GCN-LABEL: {{^}}test_native_rsqrt_f32: // GCN: v_rsq_f32 kernel void test_native_rsqrt_f32(global float* restrict out, global float* restrict in) { int id = get_local_id(0);