std::gpu::isa

std::gpu::isa is the typed Silk spelling of the initial dependency-light AMDGPU instruction subset. It exists for kernel bring-up while ordinary Silk expressions are connected to the general GPU selector. Applications import this module instead of declaring compiler intrinsics with ext.

These functions may be called only inside attr(device=gpu) bodies and every argument must be an unsigned integer literal. Arguments name physical scalar or vector registers and byte offsets. They are not ordinary host operations.

The module's internal __silk_amdgpu_* declarations are compiler-owned device intrinsics only in this module. The same raw declarations remain available for the documented standalone --nostd --kind object --target amdgcn-... compatibility path; applications should prefer these typed wrappers whenever the call is part of a mixed host/device Silk program.

Common helpers#

• s_add_i32(dst, src0, src1)
• s_mov_b64(dst_pair_start, src_pair_start)
• s_waitcnt_vmcnt0_expcnt0_lgkmcnt0()
• v_add_f32(dst, src0, src1)
• v_mul_f32(dst, src0, src1)
• v_fma_f32(dst, src0, src1, src2)
• ds_write_b32(address, data, byte_offset)
• ds_read_b32(dst, address, byte_offset)
• global_load_dword(dst, address_pair_start)
• global_store_dword(address_pair_start, data)

Target-specific helpers#

• s_waitcnt_vscnt0() is available on gfx1100 and gfx1151 and rejected on gfx942.
• mfma_f32_16x16x16_f16(acc_dst_start, src_a_pair_start, src_b_pair_start, acc_src_start) is available on gfx942 and rejected on the current GFX11 targets.

The compiler accounts for every referenced register in kernel metadata and enables LDS metadata when a DS helper is used. It inserts the processor-specific program terminator when the source body does not end with return;.

This module is deliberately explicit about its low-level nature. New general language behavior belongs in the Silk IR-to-AMDGPU selector, not in a growing collection of opaque pseudo-builtins.