radv: Add on-demand compilation of built-in shaders.

In environments where we cannot cache, e.g. Android (no homedir),
ChromeOS (readonly rootfs) or sandboxes (cannot open cache), the
startup cost of creating a device in radv is rather high, due
to compiling all possible built-in pipelines up front. This meant
depending on the CPU a 1-4 sec cost of creating a Device.

For CTS this cost is unacceptable, and likely for starting random
apps too.

So if there is no cache, with this patch radv will compile shaders
on demand. Once there is a cache from the first run, even if
incomplete, the driver knows that it can likely write the cache
and precompiles everything.

Note that I did not switch the buffer and itob/btoi compute pipelines
to on-demand, since you cannot really do anything in Vulkan without
them and there are only a few.

This reduces the CTS runtime for the no caches scenario on my
threadripper from 32 minutes to 8 minutes.

Reviewed-by: Dave Airlie <[email protected]>

1 files changed, 28 insertions, 6 deletions

diff --git a/src/amd/vulkan/radv_meta_resolve_cs.c b/src/amd/vulkan/radv_meta_resolve_cs.c
index 2d79cb09fec..fca49a01bb0 100644
--- a/src/amd/vulkan/radv_meta_resolve_cs.c
+++ b/src/amd/vulkan/radv_meta_resolve_cs.c
@@ -212,6 +212,12 @@ create_resolve_pipeline(struct radv_device *device,
 	VkResult result;
 	struct radv_shader_module cs = { .nir = NULL };
 
+	mtx_lock(&device->meta_state.mtx);
+	if (*pipeline) {
+		mtx_unlock(&device->meta_state.mtx);
+		return VK_SUCCESS;
+	}
+
 	cs.nir = build_resolve_compute_shader(device, is_integer, is_srgb, samples);
 
 	/* compute shader */
@@ -239,14 +245,16 @@ create_resolve_pipeline(struct radv_device *device,
 		goto fail;
 
 	ralloc_free(cs.nir);
+	mtx_unlock(&device->meta_state.mtx);
 	return VK_SUCCESS;
 fail:
 	ralloc_free(cs.nir);
+	mtx_unlock(&device->meta_state.mtx);
 	return result;
 }
 
 VkResult
-radv_device_init_meta_resolve_compute_state(struct radv_device *device)
+radv_device_init_meta_resolve_compute_state(struct radv_device *device, bool on_demand)
 {
 	struct radv_meta_state *state = &device->meta_state;
 	VkResult res;
@@ -255,6 +263,9 @@ radv_device_init_meta_resolve_compute_state(struct radv_device *device)
 	if (res != VK_SUCCESS)
 		goto fail;
 
+	if (on_demand)
+		return VK_SUCCESS;
+
 	for (uint32_t i = 0; i < MAX_SAMPLES_LOG2; ++i) {
 		uint32_t samples = 1 << i;
 
@@ -353,16 +364,27 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer,
 			      }
 				      });
 
-	VkPipeline pipeline;
+	VkPipeline *pipeline;
 	if (vk_format_is_int(src_iview->image->vk_format))
-		pipeline = device->meta_state.resolve_compute.rc[samples_log2].i_pipeline;
+		pipeline = &device->meta_state.resolve_compute.rc[samples_log2].i_pipeline;
 	else if (vk_format_is_srgb(src_iview->image->vk_format))
-		pipeline = device->meta_state.resolve_compute.rc[samples_log2].srgb_pipeline;
+		pipeline = &device->meta_state.resolve_compute.rc[samples_log2].srgb_pipeline;
 	else
-		pipeline = device->meta_state.resolve_compute.rc[samples_log2].pipeline;
+		pipeline = &device->meta_state.resolve_compute.rc[samples_log2].pipeline;
+
+	if (!*pipeline) {
+		VkResult ret = create_resolve_pipeline(device, samples,
+		                                       vk_format_is_int(src_iview->image->vk_format),
+		                                       vk_format_is_srgb(src_iview->image->vk_format),
+		                                       pipeline);
+		if (ret != VK_SUCCESS) {
+			cmd_buffer->record_result = ret;
+			return;
+		}
+	}
 
 	radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
-			     VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
+			     VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
 
 	unsigned push_constants[4] = {
 		src_offset->x,