radv: flush more stages when semaphore are waiting.

This still doesn't give us complete pWaitDstStageMask support,
but it should provide enough to be correct if not as efficent as
possible.

If we have wait semaphores we must flush between submits and
flush the shaders as well.

This fixes the remaining fails in:
dEQP-VK.synchronization.op.single_queue.semaphore.*ssbo*

Reviewed-by: Bas Nieuwenhuizen <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
(cherry picked from commit a52470402515c46cd9f33a5d83dc8d2bc9f7bae9)

2 files changed, 23 insertions, 3 deletions

diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index 8654b1e8ab9..46490059719 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -1049,6 +1049,22 @@ VkResult radv_CreateDevice(
 			break;
 		}
 		device->ws->cs_finalize(device->flush_cs[family]);
+
+		device->flush_shader_cs[family] = device->ws->cs_create(device->ws, family);
+		switch (family) {
+		case RADV_QUEUE_GENERAL:
+		case RADV_QUEUE_COMPUTE:
+			si_cs_emit_cache_flush(device->flush_shader_cs[family],
+			                       device->physical_device->rad_info.chip_class,
+			                       family == RADV_QUEUE_COMPUTE && device->physical_device->rad_info.chip_class >= CIK,
+					       family == RADV_QUEUE_COMPUTE ? RADV_CMD_FLAG_CS_PARTIAL_FLUSH : (RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_PS_PARTIAL_FLUSH) |
+			                       RADV_CMD_FLAG_INV_ICACHE |
+			                       RADV_CMD_FLAG_INV_SMEM_L1 |
+			                       RADV_CMD_FLAG_INV_VMEM_L1 |
+			                       RADV_CMD_FLAG_INV_GLOBAL_L2);
+			break;
+		}
+		device->ws->cs_finalize(device->flush_shader_cs[family]);
 	}
 
 	if (getenv("RADV_TRACE_FILE")) {
@@ -1124,6 +1140,8 @@ void radv_DestroyDevice(
 			device->ws->cs_destroy(device->empty_cs[i]);
 		if (device->flush_cs[i])
 			device->ws->cs_destroy(device->flush_cs[i]);
+		if (device->flush_shader_cs[i])
+			device->ws->cs_destroy(device->flush_shader_cs[i]);
 	}
 	radv_device_finish_meta(device);
 
@@ -1825,7 +1843,7 @@ VkResult radv_QueueSubmit(
 
 	for (uint32_t i = 0; i < submitCount; i++) {
 		struct radeon_winsys_cs **cs_array;
-		bool do_flush = !i;
+		bool do_flush = !i || pSubmits[i].pWaitDstStageMask;
 		bool can_patch = !do_flush;
 		uint32_t advance;
 
@@ -1852,7 +1870,9 @@ VkResult radv_QueueSubmit(
 					        (pSubmits[i].commandBufferCount + do_flush));
 
 		if(do_flush)
-			cs_array[0] = queue->device->flush_cs[queue->queue_family_index];
+			cs_array[0] = pSubmits[i].waitSemaphoreCount ?
+				queue->device->flush_shader_cs[queue->queue_family_index] :
+				queue->device->flush_cs[queue->queue_family_index];
 
 		for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
 			RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index 4ace068742d..08f53a169f2 100644
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -495,7 +495,7 @@ struct radv_device {
 	int queue_count[RADV_MAX_QUEUE_FAMILIES];
 	struct radeon_winsys_cs *empty_cs[RADV_MAX_QUEUE_FAMILIES];
 	struct radeon_winsys_cs *flush_cs[RADV_MAX_QUEUE_FAMILIES];
-
+	struct radeon_winsys_cs *flush_shader_cs[RADV_MAX_QUEUE_FAMILIES];
 	uint64_t debug_flags;
 
 	bool llvm_supports_spill;