[require]
VK_KHR_spirv_1_4
VK_KHR_variable_pointers
VariablePointerFeatures.variablePointersStorageBuffer
VariablePointerFeatures.variablePointers
VK_KHR_storage_buffer_storage_class
VK_KHR_get_physical_device_properties2

[compute shader spirv]
; Test operates on a Workgroup array (%wg_var) containing an array of 16 arrays of four ints.

; Each invocation writes out the result of OpPtrDiff for elements of the
; outer array to index 16 of the corresponding output variable. Pointers are
; based the gid'th element vs the zero'th element.

; Each invocation loops 4 times and writes out 4 results each for OpPtrDiff.
; The pointers are the lid'th element of gid'th element of the outer array vs
; the iteration count mod 4'th element of the gid'th element of the outer
; array. Results are ordered so each thread writes continguous results.

OpCapability Shader
OpCapability VariablePointers
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %main "main" %wg_var %diff_var %lid_var %gid_var
OpExecutionMode %main LocalSize 16 4 1

OpDecorate %struct_17 Block
OpMemberDecorate %struct_17 0 Offset 0
OpDecorate %runtime_17 ArrayStride 68
OpDecorate %array_17 ArrayStride 4

OpDecorate %diff_var DescriptorSet 0
OpDecorate %diff_var Binding 0

OpDecorate %lid_var BuiltIn LocalInvocationId
OpDecorate %gid_var BuiltIn GlobalInvocationId

%void = OpTypeVoid
%bool = OpTypeBool
%int = OpTypeInt 32 1
%int_0 = OpConstant %int 0
%int_1 = OpConstant %int 1
%int_4 = OpConstant %int 4
%int_16 = OpConstant %int 16
%int_17 = OpConstant %int 17
%int_64 = OpConstant %int 64
%int3 = OpTypeVector %int 3

%ptr_input_int3 = OpTypePointer Input %int3

%array_4 = OpTypeArray %int %int_4
%array_array_4 = OpTypeArray %array_4 %int_16
%ptr_array_array_4 = OpTypePointer Workgroup %array_array_4
%ptr_array_4 = OpTypePointer Workgroup %array_4
%ptr_wg_int = OpTypePointer Workgroup %int

%array_17 = OpTypeArray %int %int_17
%runtime_17 = OpTypeRuntimeArray %array_17
%struct_17 = OpTypeStruct %runtime_17
%ptr_struct_17 = OpTypePointer StorageBuffer %struct_17
%ptr_array_17 = OpTypePointer StorageBuffer %array_17
%ptr_int = OpTypePointer StorageBuffer %int

%wg_var = OpVariable %ptr_array_array_4 Workgroup
%diff_var = OpVariable %ptr_struct_17 StorageBuffer
%lid_var = OpVariable %ptr_input_int3 Input
%gid_var = OpVariable %ptr_input_int3 Input

%void_fn = OpTypeFunction %void
%main = OpFunction %void None %void_fn
%entry = OpLabel
%gid = OpLoad %int3 %gid_var
%gid_x = OpCompositeExtract %int %gid 0
%gid_y = OpCompositeExtract %int %gid 1
%lid = OpLoad %int3 %lid_var
%lid_x = OpCompositeExtract %int %lid 0
%lid_y = OpCompositeExtract %int %lid 1
%array_gep_0 = OpAccessChain %ptr_array_4 %wg_var %int_0
%array_gep = OpAccessChain %ptr_array_4 %wg_var %lid_x
%lid_y_is_1 = OpIEqual %bool %lid_y %int_1
OpSelectionMerge %loop None
OpBranchConditional %lid_y_is_1 %then %loop

%then = OpLabel
; Compute results for outer array
%large_diff = OpPtrDiff %int %array_gep %array_gep_0
%large_diff_gep = OpAccessChain %ptr_int %diff_var %int_0 %gid_x %int_16
OpStore %large_diff_gep %large_diff
;
OpBranch %loop

%loop = OpLabel
%i = OpPhi %int %int_0 %entry %int_0 %then %inc_i %loop
%inc_i = OpIAdd %int %i %int_1
%i_cmp = OpIEqual %bool %inc_i %int_4
%lid_offset = OpIMul %int %lid_y %int_4
%out_gep_index = OpIAdd %int %i %lid_offset
; Compute results for inner array
%ref_gep = OpAccessChain %ptr_wg_int %array_gep %lid_y
%cmp_gep = OpAccessChain %ptr_wg_int %array_gep %i
%diff = OpPtrDiff %int %ref_gep %cmp_gep
%diff_gep = OpAccessChain %ptr_int %diff_var %int_0 %gid_x %out_gep_index
OpStore %diff_gep %diff
;
OpLoopMerge %exit %loop None
OpBranchConditional %i_cmp %exit %loop

%exit = OpLabel
OpReturn
OpFunctionEnd


[test]
# 16 x 4 x 2 = 128 work items
# 4 x 4 = 16 bytes per invocation
# 2048 bytes from looping
# 16 x 2 = 32 coarse results (1 per lid)
# 128 bytes
# 2176 total bytes
ssbo 0:0 2176

compute entrypoint main
compute 2 1 1

# diff var
probe ssbo int 0:0 0    == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 0
probe ssbo int 0:0 68   == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 1
probe ssbo int 0:0 136  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 2
probe ssbo int 0:0 204  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 3
probe ssbo int 0:0 272  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 4
probe ssbo int 0:0 340  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 5
probe ssbo int 0:0 408  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 6
probe ssbo int 0:0 476  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 7
probe ssbo int 0:0 544  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 8
probe ssbo int 0:0 612  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 9
probe ssbo int 0:0 680  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 10
probe ssbo int 0:0 748  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 11
probe ssbo int 0:0 816  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 12
probe ssbo int 0:0 884  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 13
probe ssbo int 0:0 952  == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 14
probe ssbo int 0:0 1020 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 15
probe ssbo int 0:0 1088 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 0
probe ssbo int 0:0 1156 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 1
probe ssbo int 0:0 1224 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 2
probe ssbo int 0:0 1292 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 3
probe ssbo int 0:0 1360 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 4
probe ssbo int 0:0 1428 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 5
probe ssbo int 0:0 1496 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 6
probe ssbo int 0:0 1564 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 7
probe ssbo int 0:0 1632 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 8
probe ssbo int 0:0 1700 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 9
probe ssbo int 0:0 1768 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 10
probe ssbo int 0:0 1836 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 11
probe ssbo int 0:0 1904 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 12
probe ssbo int 0:0 1972 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 13
probe ssbo int 0:0 2040 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 14
probe ssbo int 0:0 2108 == 0 -1 -2 -3 1 0 -1 -2 2 1 0 -1 3 2 1 0 15