the src op is:
%0 = bufferization.to_memref %cst : memref<40xf16, #spirv.storage_class<StorageBuffer>>
and when convert to spirv, it generates the unrealized_conversion_cast op
%6 = "builtin.unrealized_conversion_cast"(%5) : (!spirv.array<40 x f16>) -> !spirv.ptr<!spirv.struct<(!spirv.array<40 x f16, stride=2> [0])>, StorageBuffer>
is there some existed pass should i add or i need to add a pass convert from memref to spirv.ptr<spirv.struct<…>>. could anyone help, many thanks.
The memref type is already converted to !spirv.ptr there. I don’t have the full context and flow of what you are trying to do; just from the source op I think you should use a proper way to store the constant to the memref, like with memref.store or vector.transfer_write or whatever.
@antiagainst Great thanks for reply. my context is as follows, after comprehensive-bufferize pass: the ir is
// -----// IR Dump After IREEComprehensiveBufferize (iree-codegen-iree-comprehensive-bufferize) //----- //
module {
func.func @forward_dispatch_0_matmul_transpose_b_16x16x16_f16() {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f16
%cst_0 = arith.constant dense<[8.227530e-02, 1.696780e-01, -8.789060e-03, -2.351070e-01, -1.804200e-01, -1.469730e-01, -3.710940e-02, 8.691400e-02, -1.838380e-01, 8.837890e-02, -2.265630e-01, -2.456050e-01, -2.158200e-01, 1.945800e-01, 3.955080e-02, 1.032710e-01]> : tensor<16xf16>
%0 = bufferization.to_memref %cst_0 : memref<16xf16>
%1 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
memref.assume_alignment %1, 64 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
%2 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
memref.assume_alignment %2, 64 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
%3 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
memref.assume_alignment %3, 64 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
%workgroup_id_x = hal.interface.workgroup.id[0] : index
%4 = affine.apply affine_map<()[s0] -> (s0 * 16)>()[%workgroup_id_x]
linalg.fill ins(%cst : f16) outs(%3 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>)
linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"]} ins(%1, %2 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>, memref<16x16xf16, #hal.descriptor_type<storage_buffer>>) outs(%3 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>) attrs = {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[16, 16], [16, 16], [0, 0, 16], [16, 16, 16]]>} {
^bb0(%in: f16, %in_1: f16, %out: f16):
%5 = arith.mulf %in, %in_1 : f16
%6 = arith.addf %out, %5 : f16
linalg.yield %6 : f16
}
%subview = memref.subview %0[%4] [16] [1] : memref<16xf16> to memref<16xf16, strided<[1], offset: ?>>
linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d1)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%subview : memref<16xf16, strided<[1], offset: ?>>) outs(%3 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>>) {
^bb0(%in: f16, %out: f16):
%5 = arith.addf %out, %in : f16
linalg.yield %5 : f16
}
memref.copy %3, %3 : memref<16x16xf16, #hal.descriptor_type<storage_buffer>> to memref<16x16xf16, #hal.descriptor_type<storage_buffer>>
return
}
}
Then after many passes till to convert-to-spirv, the %cst_0 convert to a spirv.Constant which is a !spirv.array and bufferization.to_memref to unrealized_conversion_cast op.
%5 = "spirv.Constant"() <{value = dense<[8.227530e-02, 1.696780e-01, -8.789060e-03, -2.351070e-01, -1.804200e-01, -1.469730e-01, -3.710940e-02, 8.691400e-02, -1.838380e-01, 8.837890e-02, -2.265630e-01, -2.456050e-01, -2.158200e-01, 1.945800e-01, 3.955080e-02, 1.032710e-01]> : tensor<16xf16>}> : () -> !spirv.array<16 x f16>
%6 = "builtin.unrealized_conversion_cast"(%5) : (!spirv.array<16 x f16>) -> !spirv.ptr<!spirv.struct<(!spirv.array<16 x f16, stride=2> [0])>, StorageBuffer>
and this array will be accessed by index
%51 = "spirv.AccessChain"(%6, %46, %50) : (!spirv.ptr<!spirv.struct<(!spirv.array<16 x f16, stride=2> [0])>, StorageBuffer>, i32, i32) -> !spirv.ptr<f16, StorageBuffer>
Should I convert the bufferization.to_memref op to a memref.store or vector.transfer_write op in some pass? currently this op is unchanged until convert-to-spirv pass.
hi @antiagainst, currently I get rid of the *to_memref op caused by the one-shot bufferization by adding convert dialect bufferization pass, and get the memref.global op and memref.get_global op, but currently memref to spirv pass not support these op conversion. So could you give some advice for this conversion. Is it good that convert memref.global to spirv.GlobalVariable and memref.get_global to spirv.mlir.addressof ? Thanks.