* Add option to set the SYCL architecture for all targets
* Convert GGML_SYCL_HIP_TARGET to the more generic GGML_SYCL_ARCH option
* Document that setting GGML_SYCL_ARCH can improve the performance
* vulkan: Optimize soft_max
Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.
Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.
Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.
* vulkan: Further soft_max optimizations
Restore the workgroup size of 512 case, use it for >1024.
Use unrollable loops for more iteration counts.
* metal : add kernel arg structs (wip)
* metal : fattn args
ggml-ci
* metal : cont + avoid potential int overflow [no ci]
* metal : mul mat struct (wip)
* cont : mul mat vec
* cont : pass by reference
* cont : args is first argument
* cont : use char ptr
* cont : shmem style
* cont : thread counters style
* cont : mul mm id
ggml-ci
* cont : int safety + register optimizations
ggml-ci
* metal : GGML_OP_CONCAT
ggml-ci
* metal : GGML_OP_ADD, GGML_OP_SUB, GGML_OP_MUL, GGML_OP_DIV
* metal : GGML_OP_REPEAT
* metal : GGML_OP_CPY
* metal : GGML_OP_RMS_NORM
* metal : GGML_OP_NORM
* metal : add TODOs for rest of ops
* ggml : add ggml-metal-impl.h
ggml-ci
Compute two result elements per workgroup (for Q{4,5}_{0,1}). This reuses
the B loads across the rows and also reuses some addressing calculations.
This required manually partially unrolling the loop, since the compiler
is less willing to unroll outer loops.
Add bounds-checking on the last iteration of the loop. I think this was at
least partly broken before.
Optimize the Q4_K shader to vectorize most loads and reduce the number of
bit twiddling instructions.
* use 128 bit loads (i've tried 256->128 to death and its slower)
* double accumulator
* avx bf16 vec dot
* +3% q4_0 inference
* +7% tg +5% pp compared to master
* slower f16c version, kep for reference
* 256b version, also slow. i tried :)
* revert f16
* faster with madd
* split to functions
* Q8_0 and IQ4_NL, 5-7% faster
* fix potential overflow (performance reduced)
* 16 bit add for q4_0 only
* merge
* sycl: Use syclcompat::dp4a
* Using the syclcompat version allow the compiler to optimize the
operation with native function
* Update news section
* Update CI Windows oneAPI version to 2025.0
* Reword doc
* Call syclcompat::dp4a inside dpct::dp4a
This reverts commit 90cb61d692.
Reuse the index calculations across all of src0/src1/dst. Add a shader
variant for when src0/src1 are the same dimensions and additional modulus
for src1 aren't needed. Div/mod are slow, so add "fast" div/mod that
have a fast path when the calculation isn't needed or can be done more
cheaply.
* Fixes broken build for the SYCL CUDA backend caused by non-explicit gemm call in outprod (merged in with RWKV6 in
Optimize RWKV6 Operator Naming and Implement Multi-core CPU/ SYCL Acceleration #10133)
* Marks permuted MUL_MAT as unsupported to be able to run test-backend-ops
* Fixes asserts in norm to fix debug builds.
* tests: Fix memory bandwidth calculation for perf tests
Add a flops calculation for flash attention.
Add one GGML_OP_CPY perf test.
* vulkan: Optimize contiguous copies
Add a variant of the copy shader for when the tensors are contiguous. Avoid
the complex addressing calculations, and do four elements per invocation
to hide some other overhead.
Apply similar changes to the scale shader, since scale is always contiguous.
Add a "progress bar" for shader compiles.
Fixes#9582
Spawning too many concurrent copies of glslc leads to "Failed to create pipes"
errors on Linux. This change applies the same throttling we use for
multithreaded pipeline creation.
This change upstreams llamafile's cpu matrix
multiplication kernels for ppc64le using MMA
builtins for FP32 datatype.
This change results in a consistent 90%
improvement in input processing time, and 20%
to 80% improvement in output processing time,
across various batch sizes.
The patch is tested with Meta-Lllama-3-8B,
Mistral-7B, Llama-2-7B-chat-hf models on a
IBM POWER10 machine.
Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>
* metal : opt-in compile flag for BF16
ggml-ci
* ci : use BF16
ggml-ci
* swift : switch back to v12
* metal : has_float -> use_float
ggml-ci
* metal : fix BF16 check in MSL
ggml-ci
* ggml : add ggml_flash_attn_ext_get_prec
* metal : use F16 precision in FA kernels
ggml-ci
* metal : minor clean-up
* metal : compile-guard bf16 FA kernels
ggml-ci
* build : remove obsolete compile flag [no ci]
* metal : prevent int overflows [no ci]
* cuda : disable BF16 FA
ggml-ci
* metal : fix BF16 requirement for FA kernels
ggml-ci
* make : clean-up [no ci]
* rwkv6: rename to wkv6
* rwkv6: support avx2 avx512 armv8 armv9
* rwkv6: update cuda file name
* rwkv6: rename params
* wkv on sycl
* sycl: add some ops
* sycl: Enhance OP support judgment
* wkv6: drop armv9 and tranfer to GGML style
ggml-ci
* sync : ggml
* update the function to use appropriate types
* fix define error
* Update ggml/src/ggml-cpu.c
* add appropriate asserts
* move element-wise functions outside
* put the declaration outside the loop
* rewrite to be more inline with the common pattern for distributing threads
* use recommended way GGML_TENSOR_LOCALS
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Diego Devesa <slarengh@gmail.com>
Co-authored-by: Plamen Minev <pacominev@gmail.com>
Co-authored-by: Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
Co-authored-by: Meng, Hengyu <airdldl@163.com>
* ggml : add initial BF16 support
ggml-ci
* metal : add mul_mat_id BF16 support
ggml-ci
* metal : check for bfloat support on the Metal device
ggml-ci
* metal : better var names [no ci]
* metal : do not build bfloat kernels when not supported
ggml-ci
* metal : try to fix BF16 support check
ggml-ci
* metal : this should correctly check bfloat support
* metal : add quantized FA (vec) support
ggml-ci
* metal : add quantized FA (non-vec) support
* metal : fix support check
ggml-ci
* metal : clean-up
* metal : clean-up (cont)
* metal : fix shared memory calc + reduce smem + comments
* metal : float-correctness
* metal : minor [no ci]
* q6_k instruction reordering attempt
* better subtract method
* should be theoretically faster
small improvement with shuffle lut, likely because all loads are already done at that stage
* optimize bit fiddling
* handle -32 offset separately. bsums exists for a reason!
* use shift
* Update ggml-quants.c
* have to update ci macos version to 13 as 12 doesnt work now. 13 is still x86
* llama : fix buffer checks for mamba and rwk
* llama : fix missing worst case flag during reserve
* cuda : fix supports_op for norm
* disable sched SET_CAUSE
* ggml : fix gguf string leak when reading kv pairs fails
* ggml : avoid crashing with GGML_ABORT when the KV has an invalid type
* ggml : avoid crashing on failed memory allocations when loading a gguf file
* ggml: Add POOL2D OP for GPU ACC to the Vulkan.
- The MobileVLM model now supports inference acceleration through GPU by utilizing the Vulkan backend.
- A GGML_OP_POOL_2D shader has been added. (Pooling)
- The encoding performance of the CLIP model improved from 2.8s on the CPU to 0.7s on the GPU.
Signed-off-by: Changyeon Kim <cyzero.kim@samsung.com>
* [fix] Correct the incorrect order of the parameters.
fix casting to int.
Signed-off-by: Changyeon Kim <cyzero.kim@samsung.com>
---------
Signed-off-by: Changyeon Kim <cyzero.kim@samsung.com>
* metal : support permuted matrix multiplicaions
ggml-ci
* cont : use nb01 directly for row steps
ggml-ci
* cont : add comments [no ci]
* metal : minor refactor
* metal : minor
This commit removes the setting of the `used` field of the contexts in
the global state (g_state) in `ggml_init`.
The motivation for this change is that I believe that this additional
initialization might not be required after the changes in Commit
45fc4fed0b9fb5b1af4a8525cbebb95e11208732 ("sync : latest changes from
whisper.cpp"), which changed the initialization of the contexts field
from `{ 0 }` to `{ { 0 } }`:
```console
g_state = (struct ggml_state) {
- /*.contexts =*/ { 0 },
+ /*.contexts =*/ { { 0 } },
};
```
My understanding is that the `{0}` initialization might not have
zero-initialized all the nested fields in every array element because of
compiler differences, and might have been the reason for having the
explicit setting of the `used` fields to false.
* [CANN] Adapt to dynamically loadable backends mechanism
* Fix the Bug: inference running result is garbled in debug running model for LM models who's type is Q4_0 class
* Handle the review comments of this pull request
add intel amx isa detection
add vnni kernel for gemv cases
add vnni and amx kernel support for block_q8_0
code cleanup
fix packing B issue
enable openmp
fine tune amx kernel
switch to aten parallel pattern
add error message for nested parallelism
code cleanup
add f16 support in ggml-amx
add amx kernels for QK_K quant formats: Q4_K, Q5_K, Q6_K and IQ4_XS
update CMakeList
update README
fix some compilation warning
fix compiler warning when amx is not enabled
minor change
ggml-ci
move ggml_amx_init from ggml.c to ggml-amx/mmq.cpp
ggml-ci
update CMakeLists with -mamx-tile, -mamx-int8 and -mamx-bf16
ggml-ci
add amx as an ggml-backend
update header file, the old path for immintrin.h has changed to ggml-cpu-impl.h
minor change
update CMakeLists.txt
minor change
apply weight prepacking in set_tensor method in ggml-backend
fix compile error
ggml-ci
minor change
ggml-ci
update CMakeLists.txt
ggml-ci
add march dependency
minor change
ggml-ci
change ggml_backend_buffer_is_host to return false for amx backend
ggml-ci
fix supports_op
use device reg for AMX backend
ggml-ci
minor change
ggml-ci
minor change
fix rebase
set .buffer_from_host_ptr to be false for AMX backend
* fix: use `vm_allocate` to allocate CPU backend buffer on macOS
* fix: switch to `posix_memalign` to keep existing `free()` usages work
* feat: move `GGML_ALIGNED_MALLOC` to `ggml-backend-impl.h`, add support for `vm_allocate` on macOS
* style: formatting
* fix: move const outside of `#ifndef`
* style: formatting
* fix: unused var
* fix: transform `GGML_ALIGNED_MALLOC` and `GGML_ALIGNED_FREE` into functions and add them to `ggml-impl.h`
* fix: unused var
* fix: page align to `GGUF_DEFAULT_ALIGNMENT`
* fix: page align to `TENSOR_ALIGNMENT`
* fix: convert `TENSOR_ALIGNMENT` to a macro
* fix: increase page size to `32` on iOS
* fix: iOS page size
* fix: `hbw_posix_memalign` alignment
This commit removes the buffer_id field from the leaf_alloc struct.
The motivation for is that this field is only written to and never
read/used as far as I can tell. Each tensor_alloc has a buffer_id field
and this is what caused me to look into this more closely, to
understand what the buffer_id in leaf_alloc was used for.
* Vectorize load instructions in dmmv f16 CUDA kernel
Replaces scalar with vector load instructions, which substantially
improves performance on NVIDIA HBM GPUs, e.g. gives a 1.27X overall
speedup for Meta-Llama-3-8B-Instruct-F16 BS1 inference evaluation on
H100 SXM 80GB HBM3. On GDDR GPUs, there is a slight (1.01X) speedup.
* addressed comment
* Update ggml/src/ggml-cuda/dmmv.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* ggml : do not use BLAS with types without to_float
* ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies
* ggml : rename ggml_internal_get_type_traits -> ggml_get_type_traits
it's not really internal if everybody uses it
* docs : clarify building Android on Termux
* docs : update building Android on Termux
* docs : add cross-compiling for Android
* cmake : link dl explicitly for Android
* ggml : add metal backend registry / device
ggml-ci
* metal : fix names [no ci]
* metal : global registry and device instances
ggml-ci
* cont : alternative initialization of global objects
ggml-ci
* llama : adapt to backend changes
ggml-ci
* fixes
* metal : fix indent
* metal : fix build when MTLGPUFamilyApple3 is not available
ggml-ci
* fix merge
* metal : avoid unnecessary singleton accesses
ggml-ci
* metal : minor fix [no ci]
* metal : g_state -> g_ggml_ctx_dev_main [no ci]
* metal : avoid reference of device context in the backend context
ggml-ci
* metal : minor [no ci]
* metal : fix maxTransferRate check
* metal : remove transfer rate stuff
---------
Co-authored-by: slaren <slarengh@gmail.com>
* Single allocation of encode_async block with non-ARC capture in ggml-metal.m
* Moving Block_release to the deallocation code
* Release encode block when re-setting encoding buffer count if needed
* Update ggml/src/ggml-metal.m
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Add scaffolding for ggml logging macros
* Metal backend now uses GGML logging
* Cuda backend now uses GGML logging
* Cann backend now uses GGML logging
* Add enum tag to parameters
* Use C memory allocation funcs
* Fix compile error
* Use GGML_LOG instead of GGML_PRINT
* Rename llama_state to llama_logger_state
* Prevent null format string
* Fix whitespace
* Remove log callbacks from ggml backends
* Remove cuda log statement
* vulkan : do not use tensor->extra
This patch allows using the Vulkan backend with the RPC backend as
tensor->extra is no longer used.
Ref: #8536
* Adapt GGML_VULKAN_CHECK_RESULTS to extra removal (#2)
---------
Co-authored-by: 0cc4m <picard12@live.de>
When the device's warp size is less than 16,
it is possible for loadstride_a (mul_mm.comp:114)
and loadstride_b (mul_mm.comp:115) to be set to 0.
Because they are calculated as: the workgroup size,
multiplied by LOAD_VEC_* (which can be 1) and divided by 16.
And the workgroup size is set to be the same as the
warp/subgroup size.
The loadstride_* variables are used as increments in the
loops that populate the buffers used for the multiplication.
When they are 0 they cause an infinite loop.
But infinite loops without side-effects are UB and the
values of loadstride_* are known at compile time.
So, the compiler quietly optimizes all the loops away.
As a consequence, the buffers are not populated and
the multiplication result is just a matrix with all elements
set to 0.
We prevent the UB by making sure that the workgroup size
will never be less than 16, even if our device has a
smaller warp size (e.g. 8).
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
a return before a barrier (that happens only in some threads in
a workgroup) leads to UB.
While the old code actually works on some devices,
it fails on some others (i.e. "smaller" GPUs).
BTW, I think it would be better to set specialization constants
when the graph is built, in that way the local workgroup
could be sized appropriately.
But it would take a lot of work.
Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com>
* ggml: Added run-time detection of neon, i8mm and sve
Adds run-time detection of the Arm instructions set features
neon, i8mm and sve for Linux and Apple build targets.
* ggml: Extend feature detection to include non aarch64 Arm arch
* ggml: Move definition of ggml_arm_arch_features to the global data section
* ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels
* added fallback mechanism when the offline re-quantized model is not
optimized for the underlying target.
* fix for build errors
* remove prints from the low-level code
* Rebase to the latest upstream
Make sure n_barrier and n_barrier_passed do not share the cache line to avoid cache line bouncing.
This optimization shows performance improvements even for n_threads <= 8 cases.
Resurect TSAN (Thread Sanitizer) check so that we can avoid doing expensive read-modify-write
in the normal case and just use thread-fence as originally intended.
---
Here is the original description and suggestions from Willy Tarreau :
There's currently some false sharing between n_barrier and
n_barrier_passed that is amplified in ggml_barrier() by the fact that
all threads need to increment n_barrier when entering, while all
previous threads continue to read n_barrier_passed, waiting for the last
one to release them all. The side effect is that all these readers are
slowing down all new threads by making the cache line bounce back and
forth between readers and writers.
Just placing them in two distinct cache lines is sufficient to boost
the performance by 21% on a 80-core ARM server compared to the
no-openmp version, and by 3% compared to the openmp version.
Note that the variables could have been spread apart in the structure
as well, but it doesn't seem that the size of this threadpool struct is
critical so here we're simply aligning them.
Finally, the same issue was present when leaving the barrier since all
threads had to update the n_barrier_passed counter, though only one
would add a non-zero value. This alone is responsible for half of the
cost due to undesired serialization.
It might be possible that using a small array of n_barrier counters
could make things even faster on many-core systems, but it would likely
complicate the logic needed to detect the last thread.
Co-authored-by: Willy Tarreau <w@1wt.eu>
* AVX512 version of ggml_gemm_q4_0_8x8_q8_0
* Remove zero vector parameter passing
* Rename functions and rearrange order of macros
* Edit commments
* style : minor adjustments
* Update x to start from 0
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
