From dbf94b7dd366d95ec35f844e344d3e89ae67763b Mon Sep 17 00:00:00 2001
From: David Belanger <david.belanger@amd.com>
Date: Tue, 24 Jan 2023 09:55:04 -0500
Subject: [PATCH] libhsakmt: Fixed VGPR memory size for GFX11.0 and GFX11.1.

Fixed VGPR memory size, size was too small for some GPU, causing a memory overflow.
Refactored macro code into a function.
Thanks to Jay Cornwall for locating the problem and proposing the fix.

Change-Id: Iffedea1c4f341967f02c56d810ff048225b02c16
Signed-off-by: David Belanger <david.belanger@amd.com>


[ROCm/ROCR-Runtime commit: a847a7b80edda2103c7094c17b26b547c04957c2]
---
 projects/rocr-runtime/src/libhsakmt.h |  6 +++---
 projects/rocr-runtime/src/queues.c    | 24 +++++++++++++++++++++++-
 projects/rocr-runtime/src/topology.c  |  2 +-
 3 files changed, 27 insertions(+), 5 deletions(-)

diff --git a/projects/rocr-runtime/src/libhsakmt.h b/projects/rocr-runtime/src/libhsakmt.h
index 7f62c512de..0f80e5eb3b 100644
--- a/projects/rocr-runtime/src/libhsakmt.h
+++ b/projects/rocr-runtime/src/libhsakmt.h
@@ -149,6 +149,8 @@ enum full_gfx_versions {
 	GFX_VERSION_VANGOGH	 	= 0x0A0303,
 	GFX_VERSION_BEIGE_GOBY	 	= 0x0A0304,
 	GFX_VERSION_YELLOW_CARP	 	= 0x0A0305,
+	GFX_VERSION_PLUM_BONITO		= 0x0B0000,
+	GFX_VERSION_WHEAT_NAS		= 0x0B0001,
 };
 
 struct hsa_gfxip_table {
@@ -222,8 +224,6 @@ uint32_t get_num_sysfs_nodes(void);
 bool is_forked_child(void);
 
 /* Calculate VGPR and SGPR register file size per CU */
-#define VGPR_SIZE_PER_CU(gfxv)			\
-	(((gfxv) == GFX_VERSION_ARCTURUS ||	\
-	  (gfxv) == GFX_VERSION_ALDEBARAN) ? 0x80000 : 0x40000)
+uint32_t get_vgpr_size_per_cu(uint32_t gfxv);
 #define SGPR_SIZE_PER_CU 0x4000
 #endif
diff --git a/projects/rocr-runtime/src/queues.c b/projects/rocr-runtime/src/queues.c
index 7e28ba751a..c53d219c75 100644
--- a/projects/rocr-runtime/src/queues.c
+++ b/projects/rocr-runtime/src/queues.c
@@ -45,7 +45,7 @@
 	(((gfxv) >= 0x80000) ? 4096 : 0))
 
 #define WG_CONTEXT_DATA_SIZE_PER_CU(gfxv) 		\
-	(VGPR_SIZE_PER_CU(gfxv) + SGPR_SIZE_PER_CU +	\
+	(get_vgpr_size_per_cu(gfxv) + SGPR_SIZE_PER_CU +	\
 	 LDS_SIZE_PER_CU + HWREG_SIZE_PER_CU)
 
 #define CNTL_STACK_BYTES_PER_WAVE(gfxv)	\
@@ -87,6 +87,28 @@ struct process_doorbells {
 static unsigned int num_doorbells;
 static struct process_doorbells *doorbells;
 
+uint32_t get_vgpr_size_per_cu(uint32_t gfxv)
+{
+	uint32_t vgpr_size = 0;
+
+	switch (gfxv)
+	{
+		case GFX_VERSION_ARCTURUS:
+		case GFX_VERSION_ALDEBARAN:
+			vgpr_size = 0x80000;
+			break;
+		case GFX_VERSION_PLUM_BONITO:
+		case GFX_VERSION_WHEAT_NAS:
+			vgpr_size = 0x60000;
+			break;
+		default:
+			vgpr_size = 0x40000;
+			break;
+	}
+
+	return vgpr_size;
+}
+
 HSAKMT_STATUS init_process_doorbells(unsigned int NumNodes)
 {
 	unsigned int i;
diff --git a/projects/rocr-runtime/src/topology.c b/projects/rocr-runtime/src/topology.c
index 2887c661bc..7e99ec10d5 100644
--- a/projects/rocr-runtime/src/topology.c
+++ b/projects/rocr-runtime/src/topology.c
@@ -1239,7 +1239,7 @@ static HSAKMT_STATUS topology_sysfs_get_node_props(uint32_t node_id,
 
 		/* Get VGPR/SGPR size in byte per CU */
 		props->SGPRSizePerCU = SGPR_SIZE_PER_CU;
-		props->VGPRSizePerCU = VGPR_SIZE_PER_CU(HSA_GET_GFX_VERSION_FULL(props->EngineId.ui32));
+		props->VGPRSizePerCU = get_vgpr_size_per_cu(HSA_GET_GFX_VERSION_FULL(props->EngineId.ui32));
 
 	} else if (props->DeviceId)
 		/* still return success */