kc3-lang/angle

2018-06-28T08:32:54

ES31: Support atomic functions on D3D11 - Part I This patch is the first one of the implementation of atomic functions in D3D11. There are mainly two differences in the usage of GLSL and HLSL atomic functions: 1. All GLSL atomic functions have return values, which all represent the original value of the shared or ssbo variable; while all HLSL atomic functions don't, and the original value can be stored in the last parameter of the function call. 2. For HLSL atomic functions, the last parameter that stores the original value is optional except for InterlockedExchange and InterlockedCompareExchange. Missing original_value in the call of InterlockedExchange and InterlockedCompareExchange results in a compile error from HLSL compiler. To handle these differences, we plan to implement the translation in two steps: 1. Support direct translations from GLSL atomic functions to HLSL ones. Direct translation can only handle the following two situations: (1) The sentence is a GLSL atomic function call without requesting a return value and it is not atomicExchange or atomicCompSwap: e.g. GLSL: atomicAdd(mem, value); -> HLSL: InterlockedAdd(mem, value); (2) The sentence is a simple assignment expression: its right is a GLSL atomic function call and its left is a declared variable. e.g. GLSL: oldValue = atomicAdd(mem, value); -> HLSL: InterlockedAdd(mem, value, oldValue); 2. Support atomic functions in the situations that don't support direct translations. We will modify the intermediate tree to make direct translation work on all these situations. e.g. atomicExchange(mem, value); -> int oldValue; oldValue = atomicExchange(mem, value); int oldValue = atomicAdd(mem, value); -> int oldValue; oldValue = atomicAdd(mem, value); return atomicAdd(mem, value); -> int temp; temp = atomicAdd(mem, value); return temp; for (i = 0; i < atomicAdd(mem, value); ++i) -> int temp; temp = atomicAdd(mem, value); for (i = 0; i < temp; ++i) { ... temp = atomicAdd(mem, value); } int result = isTrue ? atomicAdd(mem, value) : 0; -> int result; if (isTrue) { result = atomicAdd(mem, value); } else { result = 0; } This patch completes Step 1 which mainly focus on the translation from GLSL atomic functions to HLSL ones. BUG=angleproject:2682 TEST=angle_end2end_tests Change-Id: I3b655b6e286dad4fd97f255f7fe87521c94db30c Reviewed-on: https://chromium-review.googlesource.com/1121835 Commit-Queue: Jiawei Shao <jiawei.shao@intel.com> Reviewed-by: Olli Etuaho <oetuaho@nvidia.com>