panvk/csf: Fix BY_REGION dependencies

When only the tile buffers are touched, it's okay to take care of the
dependency at the draw level, with DCD_FLAGS_2, but as soon as one side
of the dep has side effects that could impact the other side, we need to
split the render pass and insert a real barrier, with a proper flush on
read-only L1 caches.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Aksel Hjerpbakk <aksel.hjerpbakk@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38950>

src/panfrost/ci/panfrost-g610-fails.txt

@@ -4,16 +4,8 @@ asan-dEQP-VK.glsl.loops.special.do_while_dynamic_iterations.dowhile_trap_vertex,
 # New failures with VKCTS 1.4.4.0
 dEQP-VK.binding_model.unused_invalid_descriptor.write.invalid.combined_image_sampler,Crash
 dEQP-VK.binding_model.unused_invalid_descriptor.write.invalid.sampled_image,Crash
-dEQP-VK.image.general_layout.memory_barrier.fragment.read_write.shader_read_write,Fail
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.sampled_read_storage_write,Fail
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.shader_read_write,Fail
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.storage_read_storage_write,Fail
 asan-dEQP-VK.binding_model.unused_invalid_descriptor.write.invalid.combined_image_sampler,Crash
 asan-dEQP-VK.binding_model.unused_invalid_descriptor.write.invalid.sampled_image,Crash
-asan-dEQP-VK.image.general_layout.memory_barrier.fragment.read_write.shader_read_write,Fail
-asan-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.sampled_read_storage_write,Fail
-asan-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.shader_read_write,Fail
-asan-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.storage_read_storage_write,Fail
 
 
 afbcp-spec@arb_shader_texture_lod@execution@arb_shader_texture_lod-texgradcube,Fail

src/panfrost/ci/panfrost-g610-flakes.txt

@@ -198,9 +198,6 @@ dEQP-VK.glsl.swizzles.vector_swizzles.mediump_vec4_x_fragment
 dEQP-VK.glsl.arrays.constructor.int4_fragment
 dEQP-VK.glsl.arrays.declaration.implicit_size_int_ivec3_fragment
 
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.shader_read_write
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.sampled_read_storage_write
-dEQP-VK.image.general_layout.memory_barrier.fragment.write_read.storage_read_storage_write
 dEQP-VK.image.store.with_format.2d_array.r8_snorm_single_layer
 dEQP-VK.image.subresource_layout.2d.4_levels.r8g8b8a8_uint_offset
 dEQP-VK.image.texel_view_compatible.graphic.extended.1d_image.texture_read.astc_5x4_unorm_block.r32g32b32a32_sint

src/panfrost/vulkan/csf/panvk_vX_cmd_buffer.c

@@ -386,6 +386,35 @@ add_memory_dependency(struct panvk_cache_flush_info *cache_flush,
    }
 }
 
+static bool
+frag_subqueue_needs_sidefx_barrier(VkAccessFlags2 src_access,
+                                   VkAccessFlags2 dst_access)
+{
+   bool src_reads_mem = src_access & (VK_ACCESS_2_SHADER_SAMPLED_READ_BIT |
+                                      VK_ACCESS_2_SHADER_STORAGE_READ_BIT |
+                                      VK_ACCESS_2_MEMORY_READ_BIT);
+   bool dst_reads_mem = dst_access & (VK_ACCESS_2_SHADER_SAMPLED_READ_BIT |
+                                      VK_ACCESS_2_SHADER_STORAGE_READ_BIT |
+                                      VK_ACCESS_2_MEMORY_READ_BIT);
+   bool src_writes_mem = src_access & (VK_ACCESS_2_MEMORY_WRITE_BIT |
+                                       VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT);
+   bool dst_writes_mem = dst_access & (VK_ACCESS_2_MEMORY_WRITE_BIT |
+                                       VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT);
+
+   /* If there's no read -> write, write -> write or write -> read
+    * memory dependency, we can skip, otherwise we have to split the
+    * render pass. We could possibly add the dependency at the draw level,
+    * using extra bits in the DCD2 flags to encode storage reads/writes and
+    * adding extra WAIT/WAIT_RESOURCE shader side, but we can't flush the
+    * texture cache, so it wouldn't work for SAMPLED_READ. Let's keep things
+    * simple and consider any side effect as requiring a split, until this
+    * proves to be a real bottleneck.
+    */
+   return (src_reads_mem && dst_writes_mem) ||
+          (src_writes_mem && dst_writes_mem) ||
+          (src_writes_mem && dst_reads_mem);
+}
+
 static bool
 should_split_render_pass(const uint32_t wait_masks[static PANVK_SUBQUEUE_COUNT],
                          VkAccessFlags2 src_access, VkAccessFlags2 dst_access)
@@ -406,15 +435,19 @@ should_split_render_pass(const uint32_t wait_masks[static PANVK_SUBQUEUE_COUNT],
        BITFIELD_BIT(PANVK_SUBQUEUE_FRAGMENT))
       return true;
 
-   /* split if the fragment subqueue self-waits with a feedback loop, because
+   if (wait_masks[PANVK_SUBQUEUE_FRAGMENT] &
-    * we lower subpassLoad to texelFetch
+       BITFIELD_BIT(PANVK_SUBQUEUE_FRAGMENT)) {
-    */
+      /* split if the fragment subqueue self-waits with a feedback loop, because
-   if ((wait_masks[PANVK_SUBQUEUE_FRAGMENT] &
+       * we lower subpassLoad to texelFetch
-        BITFIELD_BIT(PANVK_SUBQUEUE_FRAGMENT)) &&
+       */
-       (src_access & (VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT |
+      if ((src_access & (VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT |
-                      VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT)) &&
+                         VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT)) &&
-       (dst_access & VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT))
+          (dst_access & VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT))
-      return true;
+         return true;
+
+      if (frag_subqueue_needs_sidefx_barrier(src_access, dst_access))
+         return true;
+   }
 
    return false;
 }
@@ -433,13 +466,29 @@ collect_cache_flush_info(enum panvk_subqueue_id subqueue,
    add_memory_dependency(cache_flush, src_access, dst_access);
 }
 
+static bool
+can_skip_barrier(struct panvk_cmd_buffer *cmdbuf, const VkDependencyInfo *info,
+                 struct panvk_sync_scope src, struct panvk_sync_scope dst)
+{
+   bool inside_rp = cmdbuf->state.gfx.render.tiler || inherits_render_ctx(cmdbuf);
+   bool by_region = info->dependencyFlags & VK_DEPENDENCY_BY_REGION_BIT;
+
+   if (inside_rp && by_region &&
+       !frag_subqueue_needs_sidefx_barrier(src.access, dst.access))
+      return true;
+
+   return false;
+}
+
 static void
-collect_cs_deps(struct panvk_cmd_buffer *cmdbuf,
+collect_cs_deps(struct panvk_cmd_buffer *cmdbuf, const VkDependencyInfo *info,
                 struct panvk_sync_scope src, struct panvk_sync_scope dst,
                 struct panvk_cs_deps *deps)
 {
-   struct panvk_device *dev = to_panvk_device(cmdbuf->vk.base.device);
+   if (can_skip_barrier(cmdbuf, info, src, dst))
+      return;
 
+   struct panvk_device *dev = to_panvk_device(cmdbuf->vk.base.device);
    uint32_t wait_masks[PANVK_SUBQUEUE_COUNT] = {0};
    add_execution_dependency(wait_masks, src.stages, dst.stages);
 
@@ -581,7 +630,7 @@ panvk_per_arch(add_cs_deps)(struct panvk_cmd_buffer *cmdbuf,
                            VK_QUEUE_FAMILY_IGNORED,
                            barrier_stage);
 
-      collect_cs_deps(cmdbuf, src, dst, out);
+      collect_cs_deps(cmdbuf, in, src, dst, out);
    }
 
    for (uint32_t i = 0; i < in->bufferMemoryBarrierCount; i++) {
@@ -593,7 +642,7 @@ panvk_per_arch(add_cs_deps)(struct panvk_cmd_buffer *cmdbuf,
                            barrier->dstQueueFamilyIndex,
                            barrier_stage);
 
-      collect_cs_deps(cmdbuf, src, dst, out);
+      collect_cs_deps(cmdbuf, in, src, dst, out);
    }
 
    for (uint32_t i = 0; i < in->imageMemoryBarrierCount; i++) {
@@ -608,7 +657,7 @@ panvk_per_arch(add_cs_deps)(struct panvk_cmd_buffer *cmdbuf,
                            barrier->dstQueueFamilyIndex,
                            barrier_stage);
 
-      collect_cs_deps(cmdbuf, src, dst, out);
+      collect_cs_deps(cmdbuf, in, src, dst, out);
 
       if (barrier_stage == PANVK_BARRIER_STAGE_FIRST && transition.stages)
          out->needs_layout_transitions = true;
@@ -736,13 +785,6 @@ panvk_per_arch(CmdPipelineBarrier2)(VkCommandBuffer commandBuffer,
                                     const VkDependencyInfo *pDependencyInfo)
 {
    VK_FROM_HANDLE(panvk_cmd_buffer, cmdbuf, commandBuffer);
-
-   /* Intra render pass barriers can be skipped iff we're inside a render
-    * pass. */
-   if ((cmdbuf->state.gfx.render.tiler || inherits_render_ctx(cmdbuf)) &&
-       (pDependencyInfo->dependencyFlags & VK_DEPENDENCY_BY_REGION_BIT))
-      return;
-
    struct panvk_cs_deps deps = {0};
 
    panvk_per_arch(add_cs_deps)(cmdbuf, PANVK_BARRIER_STAGE_FIRST, pDependencyInfo, &deps, false);