Improved Scan #855
Improved Scan #855
Conversation
include/nbl/builtin/hlsl/subgroup2/arithmetic_portability_impl.hlsl
Outdated
Show resolved
Hide resolved
| using op_t = subgroup::impl::inclusive_scan<binop_t, native>; | ||
| // assert T == scalar type, binop::type == T | ||
|
|
||
| T operator()(NBL_CONST_REF_ARG(T) value) | ||
| { | ||
| op_t op; | ||
| return op(value); | ||
| } | ||
| }; | ||
|
|
||
| template<class Binop, typename T, bool native> | ||
| struct exclusive_scan<Binop, T, 1, native> | ||
| { | ||
| using binop_t = Binop; | ||
| using op_t = subgroup::impl::exclusive_scan<binop_t, native>; | ||
|
|
||
| T operator()(NBL_CONST_REF_ARG(T) value) | ||
| { | ||
| op_t op; | ||
| return op(value); | ||
| } | ||
| }; | ||
|
|
||
| template<class Binop, typename T, bool native> | ||
| struct reduction<Binop, T, 1, native> | ||
| { | ||
| using binop_t = Binop; | ||
| using op_t = subgroup::impl::reduction<binop_t, native>; |
There was a problem hiding this comment.
benchmark is invalid if you do stuff in terms of subgroup functions, because you are are supposed to use the Params::Configuration::SizeLog2 to make sure your loops unroll, as opposed to the subgroup v1 loops which can't unroll because the loop invariant depends on gl_SubgroupSize which is a uniform and not a compile time constant (you can only hope that the IHV compiler is not dump and actually uses the subgroup size you provide in pipeline creation parameters when lowering SPIR-V to ISA)
TL;DR there can be no dependency between subgroup2 and subgroup namespace, copy the code over
include/nbl/builtin/hlsl/subgroup2/arithmetic_portability_impl.hlsl
Outdated
Show resolved
Hide resolved
| for (uint32_t i = 1; i < ItemsPerInvocation; i++) | ||
| retval[i] = binop(retval[i-1], value[i]); | ||
|
|
||
| exclusive_scan_op_t op; | ||
| scalar_t exclusive = op(retval[ItemsPerInvocation-1]); | ||
|
|
||
| //[unroll(ItemsPerInvocation)] | ||
| for (uint32_t i = 0; i < ItemsPerInvocation; i++) | ||
| retval[i] = binop(retval[i], exclusive); |
There was a problem hiding this comment.
this only works if the subgroup invocations are not coalesced
| inclusive_scan_op_t op; | ||
| value = op(value); | ||
|
|
||
| type_t left = glsl::subgroupShuffleUp<type_t>(value,1); |
There was a problem hiding this comment.
yeah, if each invocation holds consecutive input and output elements, this shift becomes a mess (see that loop you have at the end)
also there was never a need to shuffle the entire vector, because you only ever used the last component
There was a problem hiding this comment.
if you do coalesced, then a plain subgroup shuffle on the vector and then conditional set of first element (literal vectorized version of old code) will achieve what you want
const uint32_t invocationID = glsl::gl_SubgroupInvocationID();
// cyclic/modulo shuffle instead of relative needed
const type_t left = ItemsPerInvocation ? glsl::subgroupShuffle<type_t>(value,(invocationID-1)&SubgroupMask):glsl::subgroupShuffleUp<type_t>(value,1);
type_t newFirst; newFirst[0] = binop_t::identity;
[unroll]
for (uint32_t i=1; i<ItemsPerInvocation; i++)
newFirst[i] = left[i-1];
return mix(newFirst,left,bool(glsl::gl_SubgroupInvocationID()));P.S. also use mix(T,T,bool) instead of ? bevcause of HLSL short circuiting and turning ternaries into branches.
There was a problem hiding this comment.
btw the subgroupShuffle with a modulo SubgroupSize can be replaced with new intrinsic from SPV_KHR_subgroup_rotate if you extend the device_limits.json and so on (so that device_capability_traits gets it)
include/nbl/builtin/hlsl/subgroup2/arithmetic_portability_impl.hlsl
Outdated
Show resolved
Hide resolved
Description
Testing
TODO list: