From bca53a9c76e1edbbbdff74ec4e299951303a2a7c Mon Sep 17 00:00:00 2001
From: Hansung Kim <hansung_kim@berkeley.edu>
Date: Fri, 19 Jul 2024 16:37:51 -0700
Subject: [PATCH] sgemm_tcore: Skip load at last k-iter; do DMA by default

---
 tests/regression/sgemm_tcore/Makefile   | 2 +-
 tests/regression/sgemm_tcore/kernel.cpp | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/tests/regression/sgemm_tcore/Makefile b/tests/regression/sgemm_tcore/Makefile
index c9e2e02c..deb2c1ca 100644
--- a/tests/regression/sgemm_tcore/Makefile
+++ b/tests/regression/sgemm_tcore/Makefile
@@ -2,7 +2,7 @@ PROJECT = sgemm_tcore
 
 SRCS = main.cpp common.h
 
-VX_SRCS = kernel.activation.cpp
+VX_SRCS = kernel.cpp
 
 OPTS ?= -n16
 
diff --git a/tests/regression/sgemm_tcore/kernel.cpp b/tests/regression/sgemm_tcore/kernel.cpp
index 6859b7c9..6e7a275a 100644
--- a/tests/regression/sgemm_tcore/kernel.cpp
+++ b/tests/regression/sgemm_tcore/kernel.cpp
@@ -7,7 +7,7 @@
 #include "include/gemmini.h"
 #include "gemmini_mmio.h"
 
-#define GEMMINI_DMA 0
+#define GEMMINI_DMA 1
 #if SMEM_SIZE == 0x4000
 #define SMEM_ADDR_Q0 ((float * const) 0xff000000)
 #define SMEM_ADDR_Q1 ((float * const) 0xff001000)
@@ -368,7 +368,7 @@ inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
         // this is either done using DMA or SIMT cores depending on GEMMINI_DMA
 
 #if (GEMMINI_DMA == 1)
-        if (tid_in_threadblock == 0) {
+        if ((tid_in_threadblock == 0) && ((block_k * BK) != (dim_k - BK))) {
           // configure dma gmem address to load from
           // FIXME: block_k is wrong
           ROCC_INSTRUCTION_RS1_RS2(