From e53f1dbb46f2cca40abcd96905e5bbbe7083c1c3 Mon Sep 17 00:00:00 2001
From: Rostyslav Geyyer <46627076+geyyer@users.noreply.github.com>
Date: Wed, 5 Jul 2023 18:29:56 -0500
Subject: [PATCH] Add the missing archs (#785)

[ROCm/composable_kernel commit: 61dc9aa93277561bde816098ba8487d985b1b439]
---
 include/ck/utility/amd_xdlops.hpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/ck/utility/amd_xdlops.hpp b/include/ck/utility/amd_xdlops.hpp
index ca38077cf5..ea7755036f 100644
--- a/include/ck/utility/amd_xdlops.hpp
+++ b/include/ck/utility/amd_xdlops.hpp
@@ -364,7 +364,7 @@ struct intrin_mfma_f32_32x32x16f8f8<32, 32>
     template <class FloatC>
     __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
     {
-#if defined(__gfx940__)
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
         reg_c.template AsType<float16_t>()(Number<0>{}) =
             __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
                 bit_cast<long>(reg_a),
@@ -396,7 +396,7 @@ struct intrin_mfma_f32_16x16x32f8f8<16, 16>
     template <class FloatC>
     __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
     {
-#if defined(__gfx940__)
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
         reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
             bit_cast<long>(reg_a),
             bit_cast<long>(reg_b),