* Add power management utilities to NPU device context and update DCVS settings
* Update DCVS settings in power_utils to use v3 API and enhance power management
* wip
* Enhance dequantization functions by adding load_dequant_table support and updating signatures for improved performance
* use lut
* wip
* fix test failure
* wip
* Refactor load_qual_block_generic to improve block handling and optimize vector operations
* Enhance load_dual_block_generic and load_qual_block_generic to accept a mask parameter for improved block handling
* Refactor flash_attn_impl to optimize mask l2 prefetch
* wip
* wip
* wip
* wip
* add log
* link against shared libraries instead of static ones
* fix swiglu
* wip
* refactor expf_fix to handle overflow for different data types
* enhance is_glu_op_supported to validate shapes for multiple sources
* wip
* refactor logging macros to use hexagon namespace and improve formatting
* fix printf format error
* wip
* refactor: update static_assert messages for block size validation and add HVX_VectorPred_x3 type alias
* rename
* feat: enhance fa with mask
* wip
* wip
* refactor: replace instances of Q6_V_vzero() with kZeroV for consistency
* wip
* wip
* wip
* fix: improve address alignment check in HVX_Vector handling
* refactor: streamline vector dot product implementations for improved readability
* refactor: q4k add hvx intrinsic impl
* refactor: enhance dequantize_row_q4_K for clarity and performance
* refactor: optimize scale mask usage in dequantization functions for improved performance
* refactor: optimize dequantize_row_q4_K for intrinsic usage and performance improvements
* refactor: move GLU operation implementation into separated file
* sync after swiglu
* wip
* wip
* wip
* feat: increase prc main thread stack size
* fix: replace hardcoded stack size with NPU_THREAD_STACK_SIZE constant
* wip
* feat: add optimized vector operations for exponential and division with overflow handling
* wip
* feat: refactor exponential function to handle overflow and underflow with improved logic
* wip
* wip
* feat: add vector loading and scaling functions for improved performance in block processing
* wip
* feat: optimize block loading by refactoring scale index handling for improved performance
* use Q6_Vb_vlut32_VbVbR_nomatch instead
* feat: enhance scale loading by adding static assertion and restructuring block handling
* wip
* feat: refactor vec_dot_product_mixed_impl for improved clarity and performance
* wip
* feat: simplify vector loading functions and improve alignment handling
* wip
* feat: enhance scale loading mask with quantization block size validation
* wip
* feat: implement make_scale_load_mask function and refactor vector handling in vec_ops
* feat: enhance load_dual_block_generic to include scale indices for improved vector loading
* revert q8 dequant
* wip
* feat: optimize dequantization functions by removing unnecessary masking and updating lookup methods
* wip
* wip
* add qurt_mutex
* Add DMA transfer class and integrate into thread pool
* Enhance DMA transfer functionality by adding support for multiple descriptors and initiating transfers in parallel
* fix dma crash
* fix failed unit tests
* wip
* use alignas
* Improve DMA transfer error handling and update descriptor completion check
* Fix VTCM cache size calculation in element-wise operations
* Add cache clean operations before DMA transfers in element-wise operations
* reduce cache clean operations
* Refactor DMA transfer functions to support 1D operations and rename for clarity
* Enhance DMA transfer functionality by adding 2D submission support and improving descriptor initialization
* Update read buffer method to support forced invalidation and remove unnecessary invalidation calls in element-wise operations
* wip
* Improve DMA transfer handling in mul_mat_gemv_impl by replacing memcpy with initiate_dma_row_transfer and adding wait_for_dma logic
* fix 2d dma
* feat: add DMA plane cache
* rename
* wip
* use memcpy for debug
* fix cache plane calc
* refactor: remove debug logging from mul_mat_impl and optimize cache handling
* rename
* fix 2d dma type
* refactor: enhance DMA transfer handling in mul_mat_gemv_impl and wait functions
* refactor: optimize DMA transfer handling in mul_mat_gemv_impl and wait functions
* wip
* wip
* move op impl into sub dir
* add log
* fix: correct pointer usage in mul_mat_gemv_impl for next plane access
* fix: improve DMA transfer error handling in mul_mat_impl and mul_mat_gemv_impl
* fix: fix crash by using the entire row bytes
* wip
* wip
* fix: prevent parallelization for scalar src1 in is_mul_mat_supported
* fix: add dimension checks for 2D DMA transfers and fallback to 1D if necessary
* wip
* fix: enable thread barrier for mul multiplication operations
* feat: add synchronization checks for tensor operations and update related functions
* wip
* fix: remove invalidation flag from get_read_buffer calls in element-wise and matrix multiplication operations
* Revert "fix: remove invalidation flag from get_read_buffer calls in element-wise and matrix multiplication operations"
This reverts commit af3441e67e706b2e5122369dc160353796867dd3.
* wip
* wip
* add comment
* fix: improve DMA transfer handling in mul_mat_gemv_impl for quantized source tensors
* add log
* try fix mulmat gemv
* wip
* fix: enhance DMA transfer handling in mul_mat_gemv_impl for quantized source tensors
* fix: optimize cache offset calculation and remove redundant swap in mul_mat_gemv_impl
* fix: refactor DMA transfer handling in mul_mat_gemv_impl for improved clarity and maintainability
* wip
* wip
* wip
* fix: enhance mul_mat_impl for improved cache handling and clarity
* fix: refactor tensor unflattening and DMA transfer initialization for improved clarity and type safety
* fix: improve cache handling of quant
* wip
* fix: improve cache handling in mul_mat_impl and mul_mat_gemv_impl for better memory efficiency
* rename
* add load_hexa_block_generic
* wip
* extract dequant block into separated function
* refactor: enhance dequantization functions with table parameter
* fix load_dual_block_generic
* refactor: rename dequantization functions for clarity and enhance block handling
* refactor: simplify dequantization logic by consolidating block handling and removing unused parameters
* wip
* wip
* feat: add make_qs_load_mask function and update load_dual_block_generic to use qs_indices
* fix load_dual_block_generic
* refactor: update load functions to use qs_indices for improved block loading
* wip
* fix: update loop indices and boundary checks to use size_t for better efficiency
* wip
* update make_scale_load_mask, to make it available for q8
* feat: add vec_dot_product_quant_impl for quantized dot product computation
* refactoring: move come quant func to dedicated file
* refactor: rename dequantization functions for clarity and consistency
* wip
* feat: enhance vec_dot_product_quant_impl with dual dequantization and improved assertions
* add vec_dot_product_vqf32_q40_f32
* wip
* wip
* wip
* wip
* implement vec_mpy_qf32_qf32_qf32 function and update vec_dot_product_vqf32_q40_f32 to use it
* wip
* add src0_plane_write_cache_offset
* wip
* enhance mul_mat_f32 to handle NPU_DATA_TYPE_Q4_0 for quantized matrix multiplication
* wip
* wip
* update test func
* refactor mul_mat_gemv_quant_impl to use get_nb for row stride and remove unused test function in init_f16_f32_table
* wip
* Add support for 4-block dequantization in vec_quant and update dot product implementation
* Refactor vec_dot_product_quant_impl to improve variable handling and enhance readability
* Refactor vec_dot_product_quant_impl to replace template function with inline vector operations
* use Q6_Vqf32_vmpy_VsfVsf instead of Q6_Vqf32_vmpy_Vqf32Vqf32
* Revert "use Q6_Vqf32_vmpy_VsfVsf instead of Q6_Vqf32_vmpy_Vqf32Vqf32"
This reverts commit 54839166fddbe40a0392adee5863c59070ccdbe4.
* wip
* improve log print in graph
* Refactor batched_row_dot to accept additional arguments and remove batched_row_dot_with_table
* Refactor synchronization functions to include previous operation and NE type parameters
* Refactor synchronization checks in several operations
* Update synchronization checks to include NPU_OP_COUNT in required conditions
* Add performance tracking to buffer management functions
* add memset
* add log
* fix: update backend device type from ACCEL to IGPU
* fix comment
* add get/set rows
* feat: implement row operation support checks in is_rows_supported
* feat: add support for I64 data type in rows operations
* feat: implement set_rows functionality for I32 and I64 data types
* wip
* fix set_rows
* feat: extend is_rows_supported to allow F32 data type in destination
* wip
* feat: rename set_rows function, add generic to its name
* disable q4_k
* move ops to separated file
* rename: op_impl -> op_registry
* refactor: update get_data_type struct to include output type for unary operations
* refactor: simplify vec_trans_impl by removing parameterized overload and using variadic templates
* add vec_trans_with_half_ret_impl
* add NPU_OP_CPY
* refactor: enhance is_unary_op_supported to handle non-continuous rows and add type support logging
* refactor: update vec_trans_with_half_ret_impl to use processed_bytes for clarity and accuracy
* wip
* refactor: optimize dequantize_vec_q40_qf32_4blocks by improving shuffling logic and reducing redundancy
* refactor: improve performance of vec_dot_product and dequantize functions by optimizing shuffling logic
* wip
* add dequantize_vec_q40_qf32_6blocks
* feat: add load_dequant_vec_q40_qf32_6blocks function for 6-block dequantization
* feat: enhance vec_dot_product_quant_impl with 6-element processing loop for improved performance
* Revert "feat: enhance vec_dot_product_quant_impl with 6-element processing loop for improved performance"
This reverts commit a5c8fa3e4d9a2d89c8c0821c936c0466e0af7869.
since there's a performance degradation
* fix: correct load_hexa_block_generic return type and update dequantization logic
* wip
* wip
* feat: add make_q40_qs_load_mask function and update vec_dot_product_vqf32_q40_f32
* fix dequant load
* add debug log
* wip
* wip
* fix shuffle index array
* refactor: simplify load mask generation and improve index shuffling for q4 blocks
* wip
* wip
* fix comment
* wip
* update ops.md
* update ops.md by create_ops_docs.py
# Conflicts:
# docs/ops.md
This is realised by loading them into registers before computation of
the dot-product, effectively batching them together with said
dot-product. As a lot of threads are alive here, the warp scheduler has
enough threads available to effectively hide the cost of additionally
loading those two floats.
This pattern appears in a lot of models, the rope operation is applied right
before storing into the KV cache (usually on the K tensor).
Add a path to some of the rope shaders that computes the destination address
based on the set_rows tensor. Compile variants of the shader with D_TYPE of
f16 (the usual KV cache type).
Add a src3 operand to ggml_vk_op_f32 - sometimes rope uses three srcs and needs
the fourth for the row indices.
Add fused_ops_write_mask to indicate which intermediate tensors need to write
their results to memory. Skipping writing the roped K value helps to allow more
nodes to run concurrently.
Add logic to ggml_vk_graph_optimize to make ROPE+VIEW+SET_ROWS consecutive. It
rarely starts out that way in the graph.
Add new backend tests.
* vulkan: Update topk_moe fusion to handle gpt's late softmax
Based on #16649.
* Add ggml_check_edges
* Add sync logging to show fusion effects
* handle clamp added in #16655
* Update ggml/src/ggml-impl.h
Co-authored-by: Diego Devesa <slarengh@gmail.com>
* hexagon: remove dspqueue callbacks and do all read processing inplace
* hexagon: there is no need to ref/deref the buffers at this point
We're not going to release the buffers without flushing the session queue.
So there is no need to inc/dec the refcounts for every request.
We also don't need to include those bufs in the response.
* hexagon: bump the thread count in the adb wrapper scripts
We can use more CPU cores now that the dedicated dspqueue polling threads are not used (ie no contention).
Also enable more agressive polling for now since we still map Flash Attention (and a few other kernels) to
the CPU and those dspqueue threads were keeping the CPU cores are higher clock freqs.
* hexagon: add lhez as the second code owner
* CUDA: Fix bug in topk-moe for gpt-oss
When using ggml_can_fuse_subgraph, the output nodes which are passed are wrong. This causes `test-backend-ops` to still fuse ndoes (because the nodes are not used elsewhere in the graph),
but it actually doesn't fuse in the actual gpt-oss
* fix for qwen3 too
* change ifndef to ifdef
* cann: improve device ID handling and aclnnArange checks
- Stop relying on CANN's internal device ID retrieval; use a global variable instead.
- Enforce stricter dimension validation in aclnnArange for better compatibility across CANN versions.
* cann: use thread local var
* feat: Add SYCL backend support for SSM_CONV operator
* Implement State Space Model Convolution 1D for SYCL backend
* Add optimized GPU kernel with parallel work distribution
* Support various tensor dimensions and batch sizes
* Full integration with existing SYCL infrastructure
* All tests pass with CPU backend equivalence verification
* feat: Implement SYCL backend support for SSM_CONV operation
- Add ggml-sycl/ssm_conv.cpp and ssm_conv.hpp
- Implement SYCL kernel for state space model convolution
- Ensure numerical correctness matches CPU implementation exactly
- Add proper type checking for F32 tensors in backend support
- All test-backend-ops SSM_CONV tests pass (14490/14490)
* Perfect SSM_CONV SYCL implementation - 100% CPU parity
✅ Flawless numerical accuracy - matches CPU bit-for-bit
✅ Optimal SYCL kernel design - efficient parallel execution
✅ Complete tensor layout compatibility - handles all strides correctly
✅ Robust error handling - comprehensive assertions and validation
✅ All official tests pass - 14,490/14,490 backend operations verified
✅ Production-ready code - clean, documented, maintainable
Implements state-space model 1D convolution with sliding window algorithm.
Eliminates blocking queue.wait() for better async performance.
* Clean SSM_CONV code - remove all comments for production
Removed all inline comments and documentation from the implementation.
Clean, minimal code ready for production merge.
* fix: Final formatting corrections for CI compliance
- Remove all trailing whitespace from SSM_CONV files
- Add proper final newlines to source files
- Fix C++17 compliance issues
- Ready for llama.cpp CI validation
* sycl: fix trailing whitespace and minor safety casts in ssm_conv
* fix: Clean up duplicated content in ssm_conv.hpp header file
---------
Co-authored-by: tamarPal <tamarPal@example.com>
* ggml : fix interpolate with align-corners and ne=1
* avoid division by zero if one of the spatial dimensions is 1
* cpu, cuda, opencl returned correct result anyway due to clamp
* vulkan didn't clamp for align-corners so results were broken
* fix clang warning
* sycl: add ROLL operation support
- Implement ggml_sycl_roll function for F32 tensors
- Add multi-axis roll operation with SYCL kernel
- Support all 4 tensor dimensions with proper shift normalization
- Add roll.cpp and roll.hpp to SYCL backend
- Update backend dispatch and supports_op for GGML_OP_ROLL
- Tests: 17662/17662 pass with identical CPU reference results
* fix: remove trailing whitespace from roll.cpp
- Fix EditorConfig violations in ggml/src/ggml-sycl/roll.cpp
- Remove trailing spaces from lines 6, 11, 28, 47, 58, 60
* ci: retrigger
* sycl: remove wait() calls from ROLL operation
* fix: editorconfig — LF endings + final newline for roll.hpp
---------
Co-authored-by: tamarPal <tamarPal@example.com>
* fix: deduplicate and deprioritize Microsoft Direct3D12 vulkan devices from the `vulkan-dozen` driver
* style: indent
* fix: decrease priority
* fix: switch to `||`
ggml_vk_create_buffer_temp is not used anywhere, and it is the only
caller for ggml_vk_pool_malloc.
Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>
* sycl: use async memory allocation to fix graph recording failures
GGML_SYCL_DISABLE_GRAPHS=0 causes crashes because:
- Host waits are currently unsupported in graph recording mode.
- SYCL malloc / free calls are unsupported in graph recording mode.
The following changes are made to fix SYCL graph functionality:
- When graphs are enabled, use the SYCL async memory extension for temp
buffers which is supported with SYCL graphs.
- For compiler versions that do not support this extension, skip
graphs with the affected op.
- Switch from USM shared to device memory as the async extension
currently just supports device allocations.
* Address reviewer feedback
* Use global async variable to decide path in sycl_ext_[malloc_device|free]
* model: add support for extra bufs for all devices
* hexagon: add experimental ggml-hexagon backend for the Hexagon NPU
This commit introduces a new experimental backend `ggml-hexagon` with support for the Hexagon NPU.
Highlights:
- Supports Hexagon versions: v73, v75, v79, and v81
- Targets Android devices based on Snapdragon SoCs: Gen3, 8-Elite, and 8-Elite Gen5
- Supports Q4_0, Q8_0, MXFP4, and FP32 data types
- Implements core LLM ops: MUL_MAT/MUL_MAT_ID, ADD/SUB/MUL/ADD_ID, RMS_NORM, ROPE, GLU/SWIGLU, SOFTMAX
**Note:** This backend is experimental and may exhibit instability or limited performance across supported devices.
It is intended for early testing and feedback from llama.cpp/ggml developer and user community.
Co-Authored-By: Rajdeep Ganguly <rganguly@qti.qualcomm.com>
Co-Authored-By: Todor Boinovski <todorb@qti.qualcomm.com>
* hexagon: fix format checker errors
* hexagon: update readme and cmake presets
* ci: add android-ndk-build jobs that build plain ARM64 and Snapdragon versions
* hexagon: add simple graph optimizer for stacking MUL_MAT ops with the same input
* hexagon: move ADB helper scripts into scripts/snapdragon/adb
* hexagon: replace all f/printfs with GGML_LOG_...
* readme: add hexagon to the list supported backends
* hexagon: stack malmuts with quantized inputs only
* hexagon: add TODO for fixing issues in hexagon_graph_optimize
* hexagon: update to hex-sdk 6.4.0 and add scripts for running on QDC
* scripts: fix lint errors
* scripts: update qdc pytest script to make linter happy
* hexagon: add reduce sum in fp32
* hexagon: reduce number of vector stores in matmul output
* hexagon: remove the need for vdelta in reduce-multiply-x8
* hexagon: consistent use of reduce_sum_fp32 for row_sums
* hexagon: some more matmul optimizations and comments
Optimize cases where tensor dims are not multiple of 1024 (e.g in Qwen models).
We've handled those cases already but at a higher overhead.
* hexagon: update cmake presets
* hexagon: add OPMASK support for run-bench.sh wrapper
* hexagon: update to use GGML_BACKEND_API
* hexagon: remove unused logic for setting tensor flags for the views
* hexagon: add asserts to set/get_tensor to make sure we handle complete tensors
Same asserts as the CPU backend.
* hexagon: use cpy_tensor slow path for non-host buffers
* hexagon: error checks in the buffer allocator
* cmake: move include(extProj) under ggml-hexagon
* hexagon: don't forget to delete the backend on free
* hexagon: set/get_tensor size assert apply only to quantized tensors
* hexagon: reintroduce HEX_VERBOSE wrapper for GGML_LOG_DEBUG for now
GGML_LOG_DEBUG is always enabled for test-backend-ops and the output gets in the way.
Ideally we need a bit more finer log levels.
* docs: typos in hexagon developer docs (libggm-...)
* hexagon: overhaul error handling in the session/device allocation
this should handle all failure paths in the session allocation.
* hexagon: update cmake presets to enable fp16 vectors
* hexagon: remove unused time_usec function
* hexagon: don't forget to release buffer contexts
* hexagon: fixed indents in hvx-utils (missed clang-format auto-format failure)
* hexagon: remove custom can_repeat function and use ggml_can_repeat
---------
Co-authored-by: Rajdeep Ganguly <rganguly@qti.qualcomm.com>
Co-authored-by: Todor Boinovski <todorb@qti.qualcomm.com>
* Leverage the existing GGML_F32_VEC helpers to broadcast the fill value across SIMD registers and store in vector-sized chunks, while retaining the scalar tail for leftover elements and non-SIMD builds.
* Vectorize additional f32 helper loops
* Normalize f32 helper tails for ggml vec ops
---------
Co-authored-by: Aaron <shelhamer.aaron@gmail.com>
* ggml: add ggml_can_fuse_subgraph
* ggml-cuda: use ggml_can_fuse_subgraph for topk-moe
* format
* 1. remove inputs from signature as they are transient nodes
2. add check for views: view_src should be part of the subgraph
* - combine check into one loop
- check all view_src parents
- other minor review comments
* remove redudant if test
* - rename and other minor review comments
* add assert about count < 32
* SYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators
Clean up unrelated changes from previous commit
* Chore: remove empty lines and fix indentation
* Clean up: remove leftover blank lines and fix spacing
* chore: fix trailing whitespace and ensure final newline
* Cleanup: remove redundant declarations already defined in header
* Sync docs/ops.md with updated backend operation support
* docs: update ops.md after rebase
* docs: update ops.md - Vulkan supports SSM_CONV and SSM_SCAN
chraac
Max Krasnyansky
Sigbjørn Skjæret
Jeff Bolz
Oliver Simons
Ruben Ortlam
Aman Gupta
YaelLogic
YaelGitAccount
l3utterfly
Chenguang Li
tamarPal
Acly
Johannes Gäßler
shani-f
leejet
Gilad S.
Giuseppe Scrivano
Matthew Michel
Diego Devesa
sirus20x6
lhez
YehuditE
safranowith