drm/amdkfd: fix cu mask for asics with wgps

GFX10 and up have work group processors (WGP) and WGP mode is the native
compile mode.

KFD and ROCr have no visibility into whether a dispatch is operating
in CU or WGP mode.

Enforce CU masking to be pairwise continguous in enablement and
round robin distribute CUs across the SEs in a pairwise manner to
assume WGP mode at all times.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c

@@ -100,7 +100,9 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm,
 {
 	struct kfd_cu_info cu_info;
 	uint32_t cu_per_sh[KFD_MAX_NUM_SE][KFD_MAX_NUM_SH_PER_SE] = {0};
-	int i, se, sh, cu, cu_bitmap_sh_mul;
+	bool wgp_mode_req = KFD_GC_VERSION(mm->dev) >= IP_VERSION(10, 0, 0);
+	uint32_t en_mask = wgp_mode_req ? 0x3 : 0x1;
+	int i, se, sh, cu, cu_bitmap_sh_mul, inc = wgp_mode_req ? 2 : 1;
 
 	amdgpu_amdkfd_get_cu_info(mm->dev->adev, &cu_info);
 
@@ -167,13 +169,13 @@ void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm,
 		se_mask[i] = 0;
 
 	i = 0;
-	for (cu = 0; cu < 16; cu++) {
+	for (cu = 0; cu < 16; cu += inc) {
 		for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) {
 			for (se = 0; se < cu_info.num_shader_engines; se++) {
 				if (cu_per_sh[se][sh] > cu) {
-					if (cu_mask[i / 32] & (1 << (i % 32)))
-						se_mask[se] |= 1 << (cu + sh * 16);
-					i++;
+					if (cu_mask[i / 32] & (en_mask << (i % 32)))
+						se_mask[se] |= en_mask << (cu + sh * 16);
+					i += inc;
 					if (i == cu_mask_count)
 						return;
 				}

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c

@@ -498,6 +498,21 @@ int pqm_update_mqd(struct process_queue_manager *pqm,
 		return -EFAULT;
 	}
 
+	/* ASICs that have WGPs must enforce pairwise enabled mask checks. */
+	if (minfo && minfo->update_flag == UPDATE_FLAG_CU_MASK && minfo->cu_mask.ptr &&
+			KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) {
+		int i;
+
+		for (i = 0; i < minfo->cu_mask.count; i += 2) {
+			uint32_t cu_pair = (minfo->cu_mask.ptr[i / 32] >> (i % 32)) & 0x3;
+
+			if (cu_pair && cu_pair != 0x3) {
+				pr_debug("CUs must be adjacent pairwise enabled.\n");
+				return -EINVAL;
+			}
+		}
+	}
+
 	retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
 							pqn->q, minfo);
 	if (retval != 0)