An alternative to the Indirect Scan (#374)

Console devs can suck it...

Author: devshgraphicsprogramming

include/nbl/builtin/glsl/scanning_append/scanning_append.glsl

@@ -0,0 +1,74 @@
+// No header guards!!! In case you want to have multiple append buffers,
+// you might want to `#include` this header multiple times with different suffices
+#ifndef NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX
+#define NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX
+#endif
+
+// LSB
+#ifdef NBL_GLSL_EXT_shader_atomic_int64
+#define nbl_glsl_scanning_append_counter_t uint64_t
+#else
+#define nbl_glsl_scanning_append_counter_t uvec2
+#endif
+
+struct nbl_glsl_scanning_append_result_t
+{
+   uint outIndex;
+   uint exclusivePrefix;
+};
+
+// Elements with value 0 do not get appended
+// Note: If NBL_GLSL_EXT_shader_atomic_int64 is not present, then the call to this function needs to be subgroup uniform
+nbl_glsl_scanning_append_result_t NBL_GLSL_CONCATENATE2(nbl_glsl_scanning_append,NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX)(in uint value)
+{
+#ifndef NBL_GLSL_SCANNING_APPEND_COUNTER_NAME
+#error "Need to define NBL_GLSL_SCANNING_APPEND_COUNTER_NAME for the `nbl_glsl_scanning_append` function, cause GLSL is dumb and `buffer` cannot be passed around."
+#endif
+
+  const bool willAppend = bool(value);
+#ifdef NBL_GLSL_EXT_shader_atomic_int64
+  uint64_t add = value;
+  if (willAppend)
+     add |= 0x100000000ull;
+  const uint64_t count_reduction = atomicAdd(NBL_GLSL_SCANNING_APPEND_COUNTER_NAME,add);
+  return nbl_glsl_scanning_append_result_t{uint(count_reduction>>32),uint(count_reduction)};
+#else
+  #error "Untested Path, won't touch this until we actually need to ship something on Vulkan mobile or GL!"
+  uint localIndex = nbl_glsl_subgroupBallotExclusiveBitCount(nbl_glsl_subgroupBallot(willAppend));
+  uint partialPrefix = nbl_glsl_subgroupExclusiveAdd(value);
+
+  const uint lastSubgroupInvocationID = nbl_glsl_SubgroupSize-1u;
+  uint subgroupIndex,subgroupPrefix;
+  if (nbl_glsl_SubgroupInvocationID==lastSubgroupInvocationID)
+  {
+    // crude mutex, reuse MSB bit
+    const uint lockBit = 0x80000000u;
+    // first subgroup to set the bit to 1 (old value 0) proceeds with the lock
+    while (bool(atomicOr(NBL_GLSL_SCANNING_APPEND_COUNTER_NAME[1],lockBit))) {}
+    // now MSB is always 1
+    subgroupPrefix = atomicAdd(NBL_GLSL_SCANNING_APPEND_COUNTER_NAME[0],partialPrefix+value);
+    // set the MSB to 0 while adding by making sure MSB overflows
+    subgroupIndex = atomicAdd(NBL_GLSL_SCANNING_APPEND_COUNTER_NAME[1],localIndex+(willAppend ? (lockBit+1):lockBit));
+  }
+  return nbl_glsl_scanning_append_result_t{
+    nbl_glsl_subgroupBroadcast(subgroupIndex,lastSubgroupInvocationID)+localIndex,
+    nbl_glsl_subgroupBroadcast(subgroupPrefix,lastSubgroupInvocationID)+partialPrefix
+  };
+#endif
+}
+
+// optimized version which tries to omit the atomicAdd and locks if it can, in return it may return garbage/invalid value when invocation's `value==0`
+nbl_glsl_scanning_append_result_t NBL_GLSL_CONCATENATE2(nbl_glsl_scanning_append,NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX)(in uint value)
+{
+  const bool willAppend = bool(value);
+#ifdef NBL_GLSL_EXT_shader_atomic_int64
+  if (willAppend)
+#else
+  if (subgroupAny(willAppend))
+#endif
+    return NBL_GLSL_CONCATENATE2(nbl_glsl_scanning_append,NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX)(value);
+
+  return nbl_glsl_scanning_append_result_t{~0u,~0u};
+}
+
+#undef NBL_GLSL_SCANNING_APPEND_FUNCNAME_SUFFFIX

src/nbl/builtin/CMakeLists.txt

@@ -143,6 +143,8 @@ set(nbl_resources_to_embed
   "nbl/builtin/glsl/scan/indirect.comp"
   "nbl/builtin/glsl/scan/parameters_struct.glsl"
   "nbl/builtin/glsl/scan/virtual_workgroup.glsl"
+  # faster and easier scan
+  "nbl/builtin/glsl/scanning_append/scanning_append.glsl"
   # scene
   "nbl/builtin/glsl/scene/animation.glsl"
   "nbl/builtin/glsl/scene/keyframe.glsl"