콘텐츠로 이동
The Waffuru waffle markWaffuru Docs

HIP + Vulkan backends — Phase 1 (RX 9070 XT / gfx1201)

이 콘텐츠는 아직 번역되지 않았습니다.

Companion to CUDA_BACKEND_SCOPE.md. Same backend seam, same #[kernel] macro, same IR — additive Target variants + per-backend runtime.

Status

Section titled “Status”
Backend Codegen Runtime Smoke green Notes
HIP / ROCm wh-iron-codegen::hip::HipGenerator (CudaGenerator + textual transform) wh-iron-runtime::HipDevice (hand-rolled amdhip64 + hiprtc FFI) 4 / 4 on RX 9070 XT (gfx1201) wave32 RDNA 4; wave64 CDNA gated by TargetProfile::hip_wave64, not yet exercised
Vulkan / SPIR-V wh-iron-codegen::spirv::GlslGenerator (fresh elementwise walker → GLSL 460 compute) wh-iron-runtime::VulkanDevice (hand-rolled vulkan-1 + shaderc_combined FFI) 4 / 4 on RX 9070 XT Phase-1 Elementwise only; Reduction / Grid3D / Coop are Phase 2+

Build: cargo test -p wh-iron-runtime --features hip,vulkan. Bit-accuracy: vector_add is bit-exact on both backends (max |Δ| = 0); scale_add_exp (exp + constexpr) passes at max_rel ≈ 1.2e-7 on both — the same precision band as the CUDA --fmad=false path. The HIP path additionally runs row_reduce_sum green (max_rel ≈ 3.4e-7), exercising the warp-shuffle + shared-memory tree.

What was added (no Metal / CUDA behavior change)

Section titled “What was added (no Metal / CUDA behavior change)”
  1. Target::Hip, Target::Spirv enum variants + matching TargetProfile constructors (hip(), hip_wave64(), vulkan()).
  2. Extended MmaStrategy with AmdWmma, AmdMfma, Software, VkCooperativeMatrix (all data-only — no behavior wired yet).
  3. wh-iron-codegen/src/hip/mod.rs — HipGenerator composes CudaGenerator and post-processes the emitted source with three surgical rewrites (header includes, bf16 type name, shuffle-mask width).
  4. wh-iron-codegen/src/spirv/mod.rs — GlslGenerator (fresh walker for the elementwise op subset) + GlslBindingPlan side-table used by the runtime to bind storage buffers without reparsing the shader.
  5. wh-iron-runtime/src/device/hip/{mod,ffi}.rs — HipDevice.
  6. wh-iron-runtime/src/device/vulkan/{mod,ffi}.rs — VulkanDevice.
  7. Cargo features hip and vulkan; build.rs extended with Windows-aware HIP_PATH / VULKAN_SDK linker discovery.
  8. Smoke tests tests/hip_smoke.rs (4) and tests/vulkan_smoke.rs (4).

The CUDA emitter, the CUDA runtime, and the Metal path are unchanged.

Gotchas — first-bring-up notes

Section titled “Gotchas — first-bring-up notes”

HIP

Section titled “HIP”
  1. hipRTC does not bundle hip/hip_fp16.h or hip/hip_bf16.h. The CUDA preamble’s #include <cuda_fp16.h> becomes #include <hip/hip_fp16.h> after the text transform, but hipRTC then fails to find the header. Fix: pass -I<HIP_PATH>/include as a compile option (device/hip/mod.rs::compile). Without this the build fails for any kernel, not just bf16/fp16 ones, because both headers are always emitted in the preamble.
  2. HIP 7.1 __shfl_*_sync enforces a 64-bit mask type. A static_assert(sizeof(MaskT) == 8) in the HIP runtime header rejects 0xffffffffu (the CUDA emitter’s wave32 mask). The text transform rewrites it to 0xffffffffull. Wave64 (CDNA) needs 0xffffffffffffffffull — a wave64-specific transform pass for Phase 2.
  3. hipDeviceGetAttribute(HIP_DEVICE_ATTRIBUTE_WARP_SIZE) returns nonsense on ROCm 7.1 / Windows. The first run returned 65536. The attribute enum index is unstable across ROCm releases. Phase 1 derives wave size from the gfx family instead: gfx9* → 64, otherwise 32.
  4. gfx target detection. Windows hipDeviceGetName returns the marketing name (“AMD Radeon RX 9070 XT”), not the gfx code. Phase 1 reads IRON_HIP_GFX if set; default gfx1201 matches the user’s RX 9070 XT. Override for RDNA 3 (gfx1100), MI300 (gfx942), MI350 (gfx950).
  5. Wave32 + the 32-lane shuffle table = free port. Beyond the mask rewrite, zero further HIP-specific work was needed to run row_reduce_sum. The CUDA emitter’s __shfl_down_sync + per-warp shared-mem tree ported verbatim because RDNA 4 wave32 ≙ NVIDIA warp32.

Vulkan / SPIR-V

Section titled “Vulkan / SPIR-V”
  1. shaderc_compute_shader is enum value 2, not 4. Mismatched shader_kind doesn’t error — shaderc silently treats the source as a different stage, then GLSL fails with 'local_size_x': no such layout identifier for this stage and gl_GlobalInvocationID undeclared. Always verify the enum mapping against shaderc/env.h.
  2. shaderc on Windows: link shaderc_combined.lib, not shaderc_shared.lib. The combined static lib bundles libshaderc + glslang + SPIRV-Tools — one link, no DLL dep chain on PATH. Linux keeps shaderc_shared.
  3. Zero-byte Vulkan allocations are rejected. vkCreateBuffer(size=0) returns VK_ERROR_INVALID_*. Phase-1 alloc clamps small buffers to 4 B.
  4. GLSL.std.450 is missing erf / erfinv / expm1 / log10. The preamble synthesises them via iron_erf/iron_erfinv/iron_expm1/ iron_log10. iron_erfinv is a coarse Winitzki approximation — sufficient for Phase-1 smoke but should be tuned to the CPU oracle’s tolerance in Phase 2.
  5. SPIR-V is u32 words, but shaderc returns bytes. Verify spv.len() % 4 == 0 and check the magic number 0x07230203 (LE bytes: 03 02 23 07) before feeding it to vkCreateShaderModule.
  6. No gfx-equivalent identifier yet. VulkanDevice::name() returns the placeholder "vulkan-device" — Phase 2 wires vkGetPhysicalDeviceProperties for the actual device name + the subgroup-size-control query (VK_EXT_subgroup_size_control) needed for the subgroup fast-path reductions per spec §4.1.

Reproduce

Section titled “Reproduce”
# HIP (RDNA 4 / gfx1201). Add ROCm to PATH so amdhip64_7.dll loads:
$env:Path += ";C:\Program Files\AMD\ROCm\7.1\bin"
cargo test -p wh-iron-runtime --features hip --test hip_smoke -- --nocapture


# Vulkan. shaderc_combined is statically linked, so no extra PATH munging:
cargo test -p wh-iron-runtime --features vulkan --test vulkan_smoke -- --nocapture

Override gfx (e.g. for RDNA 3 / MI300): $env:IRON_HIP_GFX = "gfx1100".

What’s next (Phase 2 outline)

Section titled “What’s next (Phase 2 outline)”
  • HIP: the corpus tests that worked for CUDA should largely work for HIP since the kernel surface is identical post-transform. Wire the same test harness used in cuda_kernel_corpus against HipDevice to measure what fraction passes. Wave64 / CDNA is a separate axis (mask width + reductions sized to 64-lane warps).
  • Vulkan: add Reduction-mode lowering using the portable workgroup- shared barrier-tree (subgroup-width agnostic; the §4.1 hazard mitigation). This unlocks row_reduce_sum, rms_norm, qgemv. The subgroup-op fast path lives behind a feature query.
  • Both: wire --target hip / --target vulkan into iron build and iron bench. Add device-spec rows for the roofline view.